Mechine Learning CH

ᒿ 䀰
ᵪಘᆖҐ䘉䰘ᆖ、ᡰ‫Ⲵ⌘ޣ‬䰞仈ᱟ˖䇑㇇ᵪ〻ᒿྲօ䲿⵰㓿傼〟㍟㠚ࣘᨀ儈ᙗ㜭˛䘁ᒤ
ᶕˈᵪಘᆖҐ㻛ᡀ࣏ൠᓄ⭘Ҿᖸཊ亶ฏˈӾỰ⍻ؑ⭘঑Ӕ᱃Ⅺ䇸Ⲵᮠᦞ᥆ᧈ〻ᒿˈࡠ㧧ਆ⭘
ᡧ䰵䈫‫ޤ‬䏓Ⲵؑ᚟䗷└㌫㔏ˈ޽ࡠ㜭൘儈䙏‫ޜ‬䐟к㠚ࣘ㹼傦Ⲵ⊭䖖Ǆ਼ᰦˈ䘉њᆖ、Ⲵส⹰
⨶䇪઼㇇⌅ҏᴹҶ䟽བྷⲴ䘋ኅǄ
䘉ᵜᮉᶀⲴⴞḷᱟኅ⧠ᵪಘᆖҐѝṨᗳⲴ㇇⌅઼⨶䇪ǄᵪಘᆖҐӾᖸཊᆖ、੨᭦Ҷᡀ᷌
઼ᾲᘥˈवᤜ㔏䇑ᆖǃӪᐕᲪ㜭ǃଢᆖǃؑ᚟䇪ǃ⭏⢙ᆖǃ䇔⸕、ᆖǃ䇑㇇༽ᵲᙗ઼᧗ࡦ䇪
ㅹǄᡁ⴨ؑˈ⹄ウᵪಘᆖҐⲴᴰ֣䙄ᖴᱟӾ䘉Ӌᆖ、Ⲵ㿲⛩ⴻᖵᵪಘᆖҐˈᒦфԕ↔ᶕ⨶䀓
䰞仈Ⲵ㛼Ჟǃ㇇⌅ԕ৺ަѝ䳀ਜ਼Ⲵ‫ٷ‬ᇊǄ䘉Ӌ൘ԕᖰᖸ䳮‫ˈࡠڊ‬ഐѪ൘䘉а亶ฏ㕪ቁवᇩᒯ
⌋Ⲵ৏࿻䍴ᯉǄ䘉ᵜҖⲴѫ㾱ⴞⲴቡᱟᨀ‫׋‬䘉ṧⲴаԭ䍴ᯉǄ
⭡Ҿ㍐ᶀⲴཊᆖ、ᙗˈ䘉ᵜҖн㾱≲䈫㘵ާᴹ⴨ᓄⲴ⸕䇶㛼Ჟˈ㘼ᱟ൘ᗵ㾱ᰦӻ㓽ަԆ
аӋᆖ、Ⲵสᵜᾲᘥˈྲ㔏䇑ᆖǃӪᐕᲪ㜭ǃؑ᚟䇪ㅹǄӻ㓽Ⲵ䟽⛩ᱟоᵪಘᆖҐ‫ޣ‬㌫ᴰᇶ
࠷Ⲵ䛓ӋᾲᘥǄᵜҖਟԕ֌Ѫ䇑㇇ᵪ、ᆖоᐕ〻ǃ㔏䇑ᆖ઼⽮Պ、ᆖㅹуъⲴབྷᆖ⭏ᡆ⹄ウ
⭏Ⲵᮉᶀˈҏਟ֌Ѫ䖟Ԧ⹄ウӪઈᡆӾъӪઈⲴ৲㘳Ǆᤷሬ䘉ᵜҖ߉֌Ⲵєᶑ৏ࡉѪ˖1.
ᆳᱟ൘ṑབྷᆖ⭏ਟԕ⨶䀓Ⲵ˗2.ᆳᓄ䈕वਜ਼ঊ༛⭏൘ᔰ࿻⹄ウᵪಘᆖҐࡽ㾱ᦼᨑⲴ޵ᇩǄ
ᤷሬ䘉ᵜҖ߉֌Ⲵㅜйᶑ৏ࡉᱟ˖ᆳᓄ䈕փ⧠⨶䇪઼ᇎ䐥є㘵Ⲵᒣ㺑ǄᵪಘᆖҐ⨶䇪㠤
࣋Ҿഎㆄ䘉ṧⲴ䰞仈ĀᆖҐᙗ㜭ᱟᘾṧ䲿⵰㔉ᇊⲴ䇝㓳ṧֻⲴᮠ䟿ਈॆⲴ˛ā઼ĀሩҾн਼
㊫රⲴᆖҐԫ࣑ˈଚњᆖҐ㇇⌅ᴰ䘲ਸ˛ā࡙⭘ᶕ㠚㔏䇑ᆖǃ䇑㇇༽ᵲᙗ઼䍍ਦᯟ࠶᷀Ⲵ⨶
䇪ᡀ᷌ˈ䘉ᵜҖ䇘䇪Ҷ䘉а㊫⨶䇪䰞仈Ǆ਼ᰦᵜҖҏ㾶ⴆҶᖸཊᇎ䐥ᯩ䶒Ⲵ޵ᇩ˖ӻ㓽Ҷ䘉
а亶ฏⲴѫ㾱㇇⌅ˈᒦ䱀᰾Ҷ㇇⌅Ⲵ䘀㹼䗷〻ǄаӋ㇇⌅Ⲵᇎ⧠઼ᮠᦞਟԕ൘ӂ㚄㖁к䙊䗷
㖁൰ http://www.cs.cmu.edu/~tom/mlbook.html ᗇࡠǄަѝवᤜ⭘ҾӪ㝨䇶࡛Ⲵ⾎㓿㖁㔌ǃ⭘
Ҿؑ䍧࠶᷀ⲴߣㆆṁᆖҐǃ৺࠶᷀᮷ᵜ᮷ẓⲴ䍍ਦᯟ࠶㊫ಘ਴㠚ⲴⓀԓ⸱઼ᡰ䴰ᮠᦞǄᡁᖸ
ᝏ䉒䛓Ӌᑞࣙᡁࡋᔪ䘉Ӌ൘㓯䍴ⓀⲴ਼һˈवᤜ Jason RennieǃPaul HsiungǃJeff Shufeltǃ
Matt GlickmanǃScott DaviesǃJoseph O’SullivanǃKen LangǃAndrew McCallum ઼ Thorsten
JoachimsǄ
㠤䉒
൘߉֌䘉ᵜҖⲴ䗷〻ѝˈᡁᒨ䘀ൠᗇࡠҶᵪಘᆖҐ亶ฏᖸཊᆖ、࠶᭟ⲴᢰᵟуᇦⲴᑞ
ࣙǄ⋑ᴹԆԜⲴᑞࣙ䘉ᵜҖᱟнਟ㜭ᆼᡀⲴǄᡁ␡␡ൠᝏ◰л䶒Ⲵ、ᆖᇦԜˈԆԜ㣡ᰦ䰤Ự
䰵ᵜҖⲴ㥹はˈᒦфԕԆԜ਴㠚亶ฏⲴу䮯ሩᡁ䘋㹼ҶᤷሬǄ
˄ĂĂ˅
ᡁҏᖸᝏ䉒਴ᡰབྷᆖⲴᖸཊ䇢ᐸ઼ᆖ⭏ˈԆԜᇎൠ⍻䈅ҶᵜҖⲴᖸཊ㥹はᒦᨀࠪҶԆԜ
Ⲵᔪ䇞Ǆቭ㇑⋑ᴹ䏣ཏⲴ⡸䶒ᶕᝏ䉒кⲮ਽Ⲵᆖ⭏ǃ䇢ᐸ઼ަԆ⍻䈅Ҷ㥹はⲴӪˈᡁ㾱ᝏ䉒
л䶒਴սˈᝏ䉒ԆԜ⢩࡛ᴹᑞࣙⲴᔪ䇞઼䇘䇪Ǆ
˄ĂĂ˅
ᡁᝏ䉒 Joan Mitchell ᔪ・ҶᵜҖⲴ㍒ᕅǄᡁҏᝏ䉒 Jean Harpley ᑞࣙ㕆䗁ҶᖸཊᨂമǄ
ETP Harrison Ⲵ Jane Loftus ᑞࣙᮤ⨶ҶᵜҖⲴ᡻はǄᡁⲴ㕆䗁ˈMcGraw Hill ࠪ⡸⽮Ⲵ Eric
Munson ൘亩ⴞⲴᮤњ䗷〻ѝᨀ‫׋‬Ҷ啃઼࣡᜿㿱Ǆ
䙊ᑨˈањӪᴰ䈕ᝏ䉒ⲴᱟԆⲴ਼һǃᴻ৻઼ᇦᓝǄሩҾᡁˈቔަ㾱㺘䗮㠚ᐡⲴᝏ◰Ǆ
ᡁᖸ䳮ᜣ䊑ᴹӪ∄ᡁ൘ Carnegie Mellon ᤕᴹᴤྭⲴᲪ㘵Ӂ䳶Ⲵ⧟ຳ઼ᴤཊⲴ唾࣋⴨ࣙⲴᴻ
৻Ǆ൘䘉Ӌᖸཊᑞࣙ䗷ᡁⲴӪᖃѝˈᡁ⢩࡛ᝏ䉒 Sebastian Thrunˈ൘䘉њ亩ⴞⲴ㠚࿻㠣㓸ˈ
Ԇаⴤሩᡁ䘋㹼⵰㋮⾎啃࣡ǃᢰᵟᤷሬㅹ਴⿽᭟ᤱǄᡁⲴ⡦⇽ˈоԕᖰаṧⲴ啃࣡ᡁᒦ൘ᴰ
ᚠᖃⲴᰦ‫ى‬䰞Āᐢ㓿ᆼᡀҶੇ˛āᴰਾˈᡁаᇊ㾱ᝏ䉒ᡁⲴᇦӪ˖MeghanˈShannon ઼ JoanǄ
ԆԜ൘н⸕н㿹ѝԕ਴⿽ᯩᔿሩ↔Җ֌ࠪҶ䍑⥞Ǆ䘉ᵜҖᱟ⥞㔉ԆԜⲴǄ
Tom M. Mitchell
ㅢ1ㄖ㔠䇰
㠚Ӿ䇑㇇ᵪ㻛ਁ᰾ԕᶕˈӪԜቡᜣ⸕䚃ᆳԜ㜭н㜭ᆖҐǄྲ᷌ᡁԜ⨶䀓Ҷ䇑㇇ᵪᆖҐⲴ
޵൘ᵪࡦˈণᘾṧ֯ᆳԜṩᦞ㓿傼ᶕ㠚ࣘᨀ儈ˈ䛓Ѹᖡ૽ሶᱟオࡽⲴǄᜣ䊑алˈ൘ᵚᶕˈ
䇑㇇ᵪ㜭Ӿ५⯇䇠ᖅѝᆖҐˈ㧧ਆ⋫⯇ᯠ⯮⯵Ⲵᴰᴹ᭸ᯩ⌅˗տᆵ㇑⨶㌫㔏࠶᷀տᡧⲴ⭘⭥
⁑ᔿˈԕ䱽վ㜭Ⓚ⎸㙇˗њӪ䖟Ԧࣙ⨶䐏䑚⭘ᡧⲴ‫ޤ‬䏓ˈᒦѪަ䘹ᤙᴰᝏ‫ޤ‬䏓Ⲵ൘㓯ᯠ
䰫ĂĂǄሩ䇑㇇ᵪᆖҐⲴᡀ࣏⨶䀓ሶᔰ䗏ࠪ‫ޘ‬ᯠⲴᓄ⭘亶ฏˈᒦ֯ަ䇑㇇㜭઼࣋ਟᇊࡦᙗк
ॷࡠᯠⲴቲ⅑Ǆ਼ᰦˈ䘿ᖫൠ⨶䀓ᵪಘᆖҐⲴؑ᚟༴⨶㇇⌅ˈҏՊᴹࣙҾᴤྭൠ⨶䀓Ӫ㊫Ⲵ
ᆖҐ㜭࣋Ǆ
ⴞࡽˈᡁԜ䘈н⸕䚃ᘾṧ֯䇑㇇ᵪⲴᆖҐ㜭઼࣋Ӫ㊫⴨ტ㖾Ǆ❦㘼аӋ䪸ሩ⢩ᇊᆖҐԫ
࣑Ⲵ㇇⌅ᐢ㓿ӗ⭏Ǆ‫ޣ‬ҾᆖҐⲴ⨶䇪䇔䇶ᐢᔰ࿻䙀↕ᖒᡀǄӪԜᔰਁࠪҶᖸཊᇎ䐥ᙗⲴ䇑㇇
ᵪ〻ᒿᶕᇎ⧠н਼㊫රⲴᆖҐˈаӋ୶ъॆⲴᓄ⭘ҏᐢ㓿ࠪ⧠ǄֻྲሩҾ䈝丣䇶࡛䘉ṧⲴ䈮
仈ˈ㠣ӺѪ→ˈสҾᵪಘᆖҐⲴ㇇⌅᰾ᱮ㜌䗷ަԆⲴᯩ⌅Ǆ൘ᮠᦞ᥆ᧈ亶ฏˈᵪಘᆖҐ㇇⌅
⨶ᡰᖃ❦ൠᗇࡠᓄ⭘ˈӾवਜ਼䇮༷㔤ᣔ䇠ᖅǃُ䍧⭣䈧ǃ䠁㶽Ӕ᱃ǃ५⯇䇠ᖅㅹ㊫լؑ᚟Ⲵ
བྷරᮠᦞᓃѝਁ⧠ᴹԧ٬Ⲵؑ᚟Ǆ䲿⵰ሩ䇑㇇ᵪⲴ⨶䀓Ⲵᰕ⳺ᡀ⟏ˈᵪಘᆖҐᗵሶ൘䇑㇇ᵪ
、ᆖ઼ᢰᵟѝᢞ╄䎺ᶕ䎺䟽㾱Ⲵ䀂㢢ʽ
䙊䗷аӋ⢩ᇊⲴᡀቡᡁԜਟԕⴻࡠ䘉䰘ᢰᵟⲴ⧠⣦˖䇑㇇ᵪᐢ㓿㜭ཏᡀ࣏ൠ䇶࡛Ӫ㊫Ⲵ
䇢䈍˄Waibel 1989˗Lee 1989˅˗亴⍻㛪⚾ᛓ㘵Ⲵᓧ༽⦷˄Cooper et al. 1997˅˗Ự⍻ؑ⭘঑
Ⅺ䇸˗൘儈䙏‫ޜ‬䐟к傮傦˄Pomerleau 1989˅˗ԕ᧕䘁Ӫ㊫ц⭼ߐߋⲴ≤ᒣሩᔸ㾯⌻ৼ䱶ỻķ䘉
ṧⲴ⑨ᠿ˄Tesauro 1992, 1995˅ǄᐢᴹҶᖸཊ⨶䇪ᡀ᷌㜭ཏሩ䇝㓳ṧֻᮠ䟿ǃ‫ٷ‬䇮オ䰤བྷሿǃ
઼ᆖᗇ‫ٷ‬䇮䭉䈟⦷䘉й㘵䰤Ⲵสᵜ‫ޣ‬㌫䘋㹼࡫⭫ǄᡁԜ↓൘ᔰ࿻㧧ਆӪ㊫઼ࣘ⢙ᆖҐⲴ৏࿻
⁑රˈ⭘ԕ⨶䀓ᆳԜ઼䇑㇇ᵪⲴᆖҐ㇇⌅䰤Ⲵ‫ޣ‬㌫˄ֻྲˈLaird et al. 1986˗Anderson 1991˗
Qin et al. 1992˗Chi & Bassock 1989˗Ahn & Brewer 1993˅Ǆ൘䗷৫Ⲵॱᒤѝᰐ䇪ᱟᓄ⭘ǃ㇇
⌅ǃ⨶䇪ˈ䘈ᱟ⭏⢙㌫㔏Ⲵ⹄ウ䜭ਆᗇҶ٬ᗇ⌘ⴞⲴ䘋↕ǄᵪಘᆖҐᴰ䘁Ⲵࠐ⿽ᓄ⭘㻛ᖂ㓣
൘㺘 1-1 ѝǄLangley & Simon˄1995˅ԕ৺Rumelhart et al.˄1994˅䈳ḕҶᵪಘᆖҐⲴаӋ
ަԆᓄ⭘Ǆ
㺘 1-1 ᵪಘᆖҐⲴаӋᡀ࣏ᓄ⭘
x ᆖҐ䇶࡛Ӫ㊫Ⲵ䇢䈍
ᡰᴹᴰᡀ࣏Ⲵ䈝丣䇶࡛㌫㔏䜭֯⭘ҶḀ⿽ᖒᔿⲴᵪಘᆖҐᢰᵟǄֻྲˈSphinx ㌫㔏˄৲㿱
Lee 1989˅ਟᆖҐ⢩ᇊ䇢䈍㘵Ⲵ䈝丣䇶࡛ㆆ⮕ˈӾỰ⍻ࡠⲴ䈝丣ؑਧѝ䇶࡛ࠪสᵜⲴ丣㍐
˄phoneme˅઼অ䇽Ǆ⾎㓿㖁㔌ᆖҐᯩ⌅ ˄ֻྲ Waibel et al. 1989˅઼䳀ᔿ傜ቄਟཛ⁑ර˄hidden
Markov model˅ⲴᆖҐᯩ⌅˄ֻྲ Lee 1989˅൘䈝丣䇶࡛㌫㔏ѝҏ䶎ᑨᴹ᭸ˈᆳԜਟԕ䇙㌫
㔏㠚ࣘ䘲ᓄн਼Ⲵ䇢䈍㘵ǃ䇽≷ǃ哖‫ݻ‬仾⢩ᙗ઼㛼Ჟಚ丣ㅹㅹǄ㊫լⲴᢰᵟ൘ᖸཊؑਧ䀓
䟺䈮仈ѝᴹᓄ⭘▌࣋Ǆ
x ᆖҐ傮傦䖖䖶
ķ
䈁⌘˖а⿽㊫լ伎㹼ỻⲴ⑨ᠿˈৼᯩ਴ᤱॱӄᆀˈ䙊䗷᧧僠ᆀᶕߣᇊỻᆀ〫ࣘⲴ↕ᮠǄ
ᵪಘᆖҐᯩ⌅ᐢ㻛⭘Ҿ䇝㓳䇑㇇ᵪ᧗ࡦⲴ䖖䖶ˈ֯ަ൘਴⿽㊫රⲴ䚃䐟к↓⺞㹼傦Ǆֻྲ
ALVINN ㌫㔏˄Pomerleau 1989˅ᐢ㓿࡙⭘ᆳᆖՊⲴㆆ⮕⤜㠚൘儈䙏‫ޜ‬䐟ⲴަԆ䖖䖶ѻ䰤྄
傠ˈԕ 70 㤡䟼Ⲵᰦ䙏‫ޡ‬㹼傦Ҷ 90 㤡䟼Ǆ㊫լⲴᢰᵟਟ㜭൘ᖸཊสҾՐᝏಘⲴ᧗ࡦ䰞仈ѝ
ᗇࡠᓄ⭘Ǆ
x ᆖҐ࠶㊫ᯠⲴཙ᮷㔃ᶴ
ᵪಘᆖҐᯩ⌅ᐢ㓿㻛⭘ҾӾ਴⿽བྷ㿴⁑Ⲵᮠᦞᓃѝਁ⧠䳀㯿Ⲵа㡜㿴ᖻǄֻྲˈߣㆆṁᆖ
Ґ㇇⌅ᐢ㓿㻛㖾ഭഭᇦ㡚オ઼㡚ཙተ˄NASA˅⭘ᶕ࠶㊫ཙփˈᮠᦞᶕ㠚ㅜҼᑅ⍋傜ཙ᮷ਠ
ཚオ䈳ḕ˄Fayyad et al. 1995˅Ǆ䘉а㌫㔏⧠൘㻛⭘Ҿ㠚ࣘ࠶㊫ཚオ䈳ḕѝⲴᡰᴹཙփˈަ
ѝवਜ਼Ҷ 3T ᆇ㢲Ⲵമ‫ۿ‬ᮠᦞǄ
x ᆖҐԕц⭼㓗Ⲵ≤ᒣሩᔸ㾯⌻ৼ䱶ỻ
ᴰᡀ࣏Ⲵঊᔸ㊫˄ྲ㾯⌻ৼ䱶ỻ˅䇑㇇ᵪ〻ᒿᱟสҾᵪಘᆖҐ㇇⌅ⲴǄֻྲˈц⭼ᴰྭⲴ
㾯⌻ৼ䱶ỻ〻ᒿ TD-Gammon˄Tesauro 1992, 1995˅ᱟ䙊䗷аⲮз⅑ԕкⲴ઼㠚ᐡሩᔸᶕᆖ
Ґަㆆ⮕ⲴǄ⧠൘ᆳⲴ≤ᒣ㜭оӪ㊫Ⲵц⭼ߐߋ⴨ᖃǄ㊫լⲴᢰᵟ㻛ᓄ⭘Ҿ䇨ཊᇎ䱵䰞仈ˈ
ަѝ䴰㾱儈᭸ൠᩌ㍒ᓎབྷⲴᩌ㍒オ䰤Ǆ
ᵜҖ䪸ሩᵪಘᆖҐ䘉њ亶ฏˈ᧿䘠Ҷཊ⿽ᆖҐ㤳රǃ㇇⌅ǃ⨶䇪ԕ৺ᓄ⭘ǄᵪಘᆖҐӾ
ᵜ䍘кᱟањཊᆖ、Ⲵ亶ฏǄᆳ੨ਆҶӪᐕᲪ㜭ǃᾲ⦷㔏䇑ǃ䇑㇇༽ᵲᙗ⨶䇪ǃ᧗ࡦ䇪ǃؑ
᚟䇪ǃଢᆖǃ⭏⨶ᆖǃ⾎㓿⭏⢙ᆖㅹᆖ、Ⲵᡀ᷌Ǆ㺘 1-2 ᖂ㓣Ҷ䘉Ӌᆖ、ѝᖡ૽ᵪಘᆖҐⲴ
‫ޣ‬䭞ᙍᜣǄᵜҖⲴ㍐ᶀสҾн਼ᆖ、Ⲵᡀ᷌ˈ❦㘼䈫㘵нᗵ㋮䙊⇿ањᆖ、Ǆᶕ㠚䘉Ӌᆖ、
Ⲵ‫ޣ‬䭞⨶䇪ሶ֯⭘䶎уъⲴ䇽≷䇢䀓ˈަѝн⟏ᚹⲴᵟ䈝઼ᾲᘥՊ൘䴰㾱ᰦ࣐ԕӻ㓽Ǆ
㺘 1-2 аӋᆖ、઼ᆳԜሩᵪಘᆖҐⲴᖡ૽
x ӪᐕᲪ㜭
ᆖҐᾲᘥⲴㅖਧ㺘⽪Ǆ֌Ѫᩌ㍒䰞仈ⲴᵪಘᆖҐǄ֌Ѫᨀ儈䰞仈≲䀓㜭࣋䙄ᖴⲴᆖҐǄ֯
⭘‫ݸ‬傼Ⲵ⸕䇶઼䇝㓳ᮠᦞа䎧ᕅሬᆖҐǄ
x 䍍ਦᯟᯩ⌅
֌Ѫ䇑㇇‫ٷ‬䇮ᾲ⦷Ⲵส⹰Ⲵ䍍ਦᯟ⌅ࡉǄᵤ㍐䍍ਦᯟ࠶㊫ಘǄՠ䇑ᵚ㿲⍻ࡠਈ䟿Ⲵ٬Ⲵ㇇
⌅Ǆ
x 䇑㇇༽ᵲᙗ⨶䇪
н਼ᆖҐԫ࣑ѝപᴹⲴ༽ᵲᙗⲴ⨶䇪䗩⭼ˈԕ䇑㇇䟿ǃ䇝㓳ṧֻᮠ䟿ǃࠪ䭉ᮠ䟿ㅹ㺑䟿Ǆ
x ᧗ࡦ䇪
ѪҶՈॆ亴ᇊⴞḷˈᆖҐሩ਴⿽༴⨶䗷〻䘋㹼᧗ࡦˈᆖҐ亴⍻㻛᧗ࡦⲴ䗷〻Ⲵлањ⣦ᘱǄ
x ؑ᚟䇪
⟥઼ؑ᚟޵ᇩⲴᓖ䟿ǄᆖҐⲴᴰሿ᧿䘠䮯ᓖᯩ⌅Ǆ㕆⸱‫ٷ‬䇮ᰦˈᆳⲴᴰ֣㕆⸱઼оᴰ֣䇝
㓳ᒿࡇⲴ‫ޣ‬㌫Ǆ
x ଢᆖ
Āྕൾ࿶Ⲵࡳ࠰ā˄Occam’s razor˅ķ˖ᴰㆰঅⲴ‫ٷ‬䇮ᱟᴰྭⲴǄӾ㿲ሏࡠⲴᮠᦞ⌋ॆⲴ⨶
⭡࠶᷀Ǆ
x ᗳ⨶ᆖ઼⾎㓿⭏⢙ᆖ
ᇎ䐥Ⲵᑲᇊᖻ˄power law of practice˅ˈ䈕ᇊᖻᤷࠪሩҾᖸབྷ㤳ത޵ⲴᆖҐ䰞仈ˈӪԜⲴ৽
ᓄ䙏ᓖ䲿⵰ᇎ䐥⅑ᮠⲴᑲ㓗ᨀ儈Ǆ◰ਁӪᐕ⾎㓿㖁㔌ⲴᆖҐ⁑ᔿⲴ⾎㓿⭏⢙ᆖ⹄ウǄ
x 㔏䇑ᆖ
ķ
䈁⌘˖ҏ〠Ā੍ஜᖻ˄Law of Parsimony’āᡆĀ㢲㓖ᖻ˄Law of Economy˅āˈѫ㾱ᙍᜣѪㆰঅⲴ⨶䇪
˄ᡆ‫ٷ‬䇮˅ՈҾ༽ᵲⲴˈഐ㤡ഭଢᆖᇦྕൾ࿶˄1285~1349˅仁㑱֯⭘䘉а৏ࡉˈ᭵〠ѪĀྕൾ࿶ࡳ࠰āǄ
൘ՠ䇑ᴹ䲀ᮠᦞṧᵜкⲴ‫ٷ‬䇮㋮ᓖᰦࠪ⧠Ⲵ䈟ᐞ˄ֻྲ‫ٿ‬ᐞ઼ᯩᐞ˅Ⲵ࡫⭫Ǆ㖞ؑ४䰤ˈ
㔏䇑Ự傼Ǆ
1.1 ᆜҖ䰤从Ⲻḽ߼ᨅ䘦
䇙ᡁԜӾࠐњᇎ䱵ⲴᆖҐԫ࣑ᔰ࿻⹄ウᵪಘᆖҐǄṩᦞᵜҖⲴⴞⲴˈᡁԜ㔉ᆖҐањᇭ
ᒯⲴᇊѹˈԕ֯ަवᤜԫօ䇑㇇ᵪ〻ᒿ䙊䗷㓿傼ᶕᨀ儈Ḁԫ࣑༴⨶ᙗ㜭Ⲵ㹼ѪǄᴤ߶⺞ൠ䇢ˈ
//࡙⭘㓿傼᭩ழ㌫㔏㠚䓛Ⲵᙗ㜭//
ᇐѿφ 对于某类任务 T 和性能度量 P，如果一个计算机程序在 T(任务)上以 P（性

能标准）衡量的性能随着经验 E 而自我完善，那么我们称这个计算机程序在从经验 E
学习。
ֻྲˈሩҾᆖҐл㾯⌻䐣ỻķⲴ䇑㇇ᵪ〻ᒿˈᆳਟԕ䙊䗷઼㠚ᐡлỻ㧧ਆ㓿傼ˈᆳᣵ䍏
Ⲵԫ࣑ᱟ৲о㾯⌻䐣ỻሩᔸˈᆳⲴᙗ㜭⭘ᆳ䎒ỻⲴ㜭࣋ᶕ㺑䟿Ǆ䙊ᑨˈѪҶᖸྭൠᇊѹањ
ᆖҐ䰞仈ˈᡁԜᗵ享᰾⺞䘉ṧйњ⢩ᖱ˖ԫ࣑Ⲵ⿽㊫˗㺑䟿ԫ࣑ᨀ儈Ⲵḷ߶˗㓿傼ⲴᶕⓀǄ
㾯⌻䐣ỻᆖҐ䰞仈˖
x ԫ࣑ T˖л㾯⌻䐣ỻ
x ᙗ㜭ḷ߶ P˖∄䎋ѝࠫ䍕ሩ᡻ⲴⲮ࠶∄
x 䇝㓳㓿傼 E˖઼㠚ᐡ䘋㹼ሩᔸ
ᡁԜਟԕ⭘ԕкᯩ⌅ᇊѹᖸཊᆖҐ䰞仈ˈֻྲᆖҐ᡻߉䇶࡛ǃᆖҐ㠚ࣘ傮傦ᵪಘӪ⊭䖖Ǆ
᡻߉䇶࡛ᆖҐ䰞仈˖
x ԫ࣑ T˖䇶઼࡛࠶㊫മ‫ۿ‬ѝⲴ᡻߉᮷ᆇ
x ᙗ㜭ḷ߶ P˖࠶㊫Ⲵ↓⺞⦷
x 䇝㓳㓿傼 E˖ᐢ⸕࠶㊫Ⲵ᡻߉᮷ᆇᮠᦞᓃ
ᵪಘӪ傮傦ᆖҐ䰞仈˖
x ԫ࣑ T˖䙊䗷㿶㿹Րᝏಘ൘ഋ䖖䚃儈䙏‫ޜ‬䐟к傮傦
ķ
䈁⌘˖ѪҶᴤྭ⨶䀓ᵜֻˈл䶒ㆰ㾱ӻ㓽ал䘉⿽䐣ỻǄỻⴈѪ 8h8 ᯩṬˈ␡㢢ỻṬнਟ⵰ᆀǄਟঅ
↕㹼䎠ˈӖਟ⇿↕䐘ሩᯩаᆀঅ䐣ᡆ䘎䐣ˈ㻛䐘䎺Ⲵᆀ㻛ᵰࠪተǄࡠ䗮ሩᯩᓅ㓯ⲴᆀᡀѪ⦻ˈਟഎੁ㹼䎠
˄ᡀѪ⦻ࡽਚਟࡽ㹼˅ˈ৸ਟ䳄オṬ伎㹼ǄлമѪ㾯⌻䐣ỻỻⴈ⽪ֻ˄䎧࿻⣦ᘱ˅Ǆ
x ᙗ㜭ḷ߶ P˖ᒣ൷ᰐᐞ䭉㹼傦䟼〻˄ᐞ䭉⭡Ӫ㊫Ⲵⴁⶓ㻱ᇊ˅
x 䇝㓳㓿傼 E˖⌘㿶Ӫ㊫傮傦ᰦᖅࡦⲴа㌫ࡇമ‫઼ۿ‬傮傦ᤷԔ
䘉䟼ሩᆖҐⲴᇊѹᖸᇭᒯˈ䏣ԕवᤜབྷཊᮠᜟҾ㻛〠ѪĀᆖҐāⲴԫ࣑ˈቡ‫ۿ‬ᡁԜᰕᑨ
֯⭘Ⲵ䘉њ䇽аṧǄ਼ᰦˈᆳҏवᤜҶԕ䶎ᑨㆰ᰾Ⲵᯩᔿ䙊䗷㓿傼㠚ᡁᨀ儈Ⲵ䇑㇇ᵪ〻ᒿǄ
ֻྲˈањ‫ݱ‬䇨⭘ᡧᴤᯠᮠᦞᶑⴞⲴᮠᦞᓃ㌫㔏ˈҏㅖਸᡁԜሩᆖҐ㌫㔏Ⲵᇊѹ˖ᆳṩᦞӾ
ᮠᦞᓃᴤᯠᗇࡠⲴ㓿傼ᨀ儈ᆳഎㆄᮠᦞḕ䈒Ⲵ㜭࣋Ǆоަᣵᗳ䘉⿽㹼ѪоĀᆖҐā䘉њ䇽ᰕ
ᑨ䈸䇪Ⲵ䶎↓ᔿਜ਼ѹ⴨␧⏶ˈᡁԜ㍒ᙗㆰঅൠ䟷⭘ᡁԜⲴ、ᢰරᇊѹüüа㊫䇑㇇ᵪ〻ᒿ䙊
䗷㓿傼ᨀ儈Ⲵ䗷〻Ǆ൘䘉њ㤳⮤޵ˈᡁԜՊਁ⧠ᖸཊ䰞仈ᡆཊᡆቁ䴰㾱䖳༽ᵲⲴ䀓ߣ࣎⌅Ǆ
䘉䟼ᡁԜᒦ䶎㾱࠶᷀ĀᆖҐā䘉њঅ䇽Ⲵᰕᑨਜ਼ѹǄ㘼ᱟ㾱㋮⺞ൠᇊѹа㊫೺ᤜᡁԜᝏ‫ޤ‬䏓
ⲴᆖҐᖒᔿⲴ䰞仈ˈ᧒㍒䀓ߣ䘉㊫䰞仈Ⲵᯩ⌅ˈᒦ⨶䀓ᆖҐ䰞仈Ⲵส⹰㔃ᶴ઼䗷〻Ǆ
1.2 䇴䇗жѠᆜҖ㌱㔕
ѪҶ╄⽪аӋᵪಘᆖҐⲴสᵜ䇮䇑ᯩ⌅઼䙄ᖴˈ㘳㲁䇮䇑ањᆖҐл㾯⌻䐣ỻⲴ〻ᒿǄ
ᡁԜⲴⴞḷᱟ䇙ᆳ䘋‫ޕ‬㾯⌻䐣ỻц⭼䭖ḷ䎋ǄᡁԜ䟷⭘ᴰᱮ㘼᱃㿱Ⲵḷ߶㺑䟿ᆳⲴᙗ㜭˖൘
ц⭼䭖ḷ䎋кᢃ䎒Ⲵ∄䎋ঐᙫ৲䎋⅑ᮠⲴⲮ࠶∄Ǆ
1.2.1 䘿᤟䇣㓹ᯯᕅ
ᡁԜ䶒ѤⲴㅜањ䇮䇑䰞仈ᱟ䘹ਆ䇝㓳㓿傼Ⲵ㊫රˈ֯㌫㔏Ӿѝ䘋㹼ᆖҐǄ㔉ᆖҐಘᨀ
‫Ⲵ׋‬䇝㓳㓿傼ሩᆳⲴᡀ䍕ᴹ䟽བྷⲴᖡ૽Ǆањ‫ޣ‬䭞኎ᙗᱟ䇝㓳㓿傼㜭੖Ѫ㌫㔏Ⲵߣㆆᨀ‫ⴤ׋‬
᧕ᡆ䰤᧕Ⲵ৽侸ǄֻྲˈሩҾᆖҐл㾯⌻䐣ỻˈ㌫㔏ਟԕӾⴤ᧕Ⲵ˄direct˅䇝㓳ṧֻˈণ
਴⿽ỻⴈ⣦ᘱ઼⴨ᓄⲴ↓⺞䎠ᆀѝᆖҐǄਖа⿽ᛵߥˈᆳਟ㜭ӵᴹ䰤᧕˄indirect˅Ⲵؑ᚟ˈ
वਜ਼ᖸཊ䗷৫Ⲵሩᔸᒿࡇ઼ᴰ㓸㔃ተǄሩҾਾа⿽ᛵߥˈ‫ޣ‬Ҿঊᔸѝ䖳ᰙ䎠ᆀⲴ↓⺞ᙗᗵ享
Ӿሩᔸᴰ㓸Ⲵ䗃䎒ᶕ᧘ᯝǄ䘉ᰦᆖҐಘ৸仍ཆ䶒Ѥањؑ⭘࠶䝽˄credit assignment˅䰞仈ˈ
ҏቡᱟ㘳㲁⇿а⅑䎠ᆀሩᴰ㓸㔃᷌Ⲵ䍑⥞〻ᓖǄؑ⭘࠶䝽ਟ㜭ᱟањ䶎ᑨ䳮ԕ䀓ߣⲴ䰞仈ˈ
ഐѪྲ᷌ਾ䶒лᗇᖸᐞˈ䛓Ѹণ֯䎧ࡍⲴ䎠ᆀᱟᴰ֣Ⲵˈ䘉ⴈỻҏՊ䗃ᦹǄᡰԕ䙊ᑨӾⴤ᧕
Ⲵ䇝㓳৽侸ᶕᆖҐ∄䰤᧕ⲴㆰঅǄ
䇝㓳㓿傼ⲴㅜҼњ䟽㾱኎ᙗᱟᆖҐಘਟԕ൘ཊབྷ〻ᓖк᧗ࡦ䇝㓳ṧֻᒿࡇǄֻྲˈᆖҐ
ಘਟ㜭‫׍‬䎆ᯭᮉ㘵䘹ਆỻⴈ⣦ᘱˈ઼ᨀ‫⇿׋‬а⅑Ⲵ↓⺞〫ࣘǄᡆ㘵ˈᆖҐಘਟ㜭㠚ᐡᨀࠪᆳ
䇔Ѫ⢩࡛ഠᜁⲴỻተᒦੁᯭᮉ㘵䈒䰞↓⺞Ⲵ䎠ᆀǄᡆ㘵ˈᆖҐಘਟԕᆼ‫ࡦ᧗ޘ‬ỻተ઼˄䰤᧕
Ⲵ˅䇝㓳࠶㊫ˈቡ‫⋑ۿ‬ᴹᯭᮉ㘵ᰦᆳ઼㠚ᐡሩᔸ䘋㹼ᆖҐаṧǄ⌘᜿ሩҾᴰਾа⿽ᛵߥᆖҐ
ಘਟ㜭䘹ᤙԕлє⿽ᛵߥѝⲴа⿽˖ㅜаˈ䈅傼ᆳ䘈ᵚ㘳㲁䗷Ⲵ‫ޘ‬ᯠỻተ˗ㅜҼˈ൘ᆳⴞࡽ
ਁ⧠Ⲵᴰཿ᭸Ⲵ䐟㓯Ⲵᗞሿਈॆкሩᔸˈԕ⼘⹪ᆳⲴᢰ㜭Ǆਾ㔝Ⲵㄐ㢲㘳㲁аӋᆖҐṶᷦˈ
वᤜҶԕлࠐ⿽ᛵߥ˖䇝㓳㓿傼ᱟԕ䎵ѾᆖҐಘ᧗ࡦⲴ䲿ᵪ䗷〻ᨀ‫˗Ⲵ׋‬ᆖҐಘਟੁᯭᮉ㘵
ᨀࠪн਼㊫රⲴḕ䈒˗ԕ৺ᆖҐಘ䙊䗷㠚ࣘ᧒㍒⧟ຳᶕᩌ䳶䇝㓳ṧֻǄ
䇝㓳㓿傼Ⲵㅜйњ䟽㾱኎ᙗᱟˈ䇝㓳ṧֻⲴ࠶ᐳ㜭ཊྭൠ㺘⽪ᇎֻ࠶ᐳˈ㘼ᴰ㓸㌫㔏Ⲵ
ᙗ㜭 P ᱟ䙊䗷ਾ㘵ᶕ㺑䟿ⲴǄа㡜㘼䀰ˈᖃ䇝㓳ṧֻⲴ࠶ᐳ઼ሶᶕⲴ⍻䈅ṧֻⲴ࠶ᐳ⴨լ
ᰦˈᆖҐާᴹᴰབྷⲴਟؑᓖǄሩҾᡁԜⲴ㾯⌻䐣ỻᆖҐˈᙗ㜭ᤷḷ P ᱟ䈕㌫㔏൘ц⭼䭖ḷ
䎋к䎒ỻⲴⲮ࠶∄Ǆྲ᷌ᆳⲴ䇝㓳㓿傼 E ӵ⭡઼ᆳ㠚ᐡሩᔸⲴ䇝㓳㓴ᡀˈ‫ׯ‬ᆈ൘ањ᰾ᱮ
Ⲵড䲙˖䘉њ䇝㓳ਟ㜭н㜭‫࠶ݵ‬ൠ㺘⽪䈕㌫㔏ԕਾ㻛⍻䈅ᰦⲴᛵᖒǄֻྲˈᆖҐಘਟ㜭൘䇝
㓳ѝӾᶕᵚ䙷ࡠ䗷ḀӋ䶎ᑨ‫ޣ‬䭞ᙗⲴỻተˈ㘼ᆳԜ৸䶎ᑨਟ㜭㻛Ӫ㊫ц⭼ߐߋ䟷⭘Ǆᇎ䱵кˈ
ᆖҐⲴṧֻ䙊ᑨоᴰ㓸㌫㔏㻛䇴ՠᰦⲴṧֻᴹаᇊᐞᔲˈᆖҐಘᗵ享㜭Ӿѝ䘋㹼ᆖҐ˄Ѯֻ
ᶕ䈤ˈц⭼㓗Ⲵ㾯⌻䐣ỻߐߋਟ㜭нՊᴹ‫ޤ‬䏓ᮉањ〻ᒿлỻ˅Ǆ䘉Ⲵ⺞ᱟањ䰞仈ˈഐѪ
ᦼᨑҶṧֻⲴа⿽࠶ᐳˈнаᇊՊሬ㠤ሩަԆⲴ࠶ᐳҏᴹྭⲴᙗ㜭Ǆਟԕⴻࡠˈⴞࡽཊᮠᵪ
ಘᆖҐ⨶䇪䜭ᱟสҾ䇝㓳ṧֻо⍻䈅ṧֻ࠶ᐳа㠤䘉аࡽᨀǄቭ㇑ᡁԜ䴰㾱䘉ṧⲴࡽᨀԕ‫ׯ‬
ᗇࡠ⨶䇪Ⲵ㔃᷌ˈն਼ṧᗵ享䇠տ൘ᇎ䐥ѝ䘉њ‫ٷ‬䇮㓿ᑨᱟнѕṬᡀ・ⲴǄ
л䶒㔗㔝䘋㹼㇇⌅䇮䇑ˈᡁԜߣᇊ㌫㔏ሶ䙊䗷઼㠚ᐡሩᔸᶕ䇝㓳Ǆ䘉ṧⲴྭ༴ᱟн䴰㾱
ཆ⭼Ⲵ䇝㓳㘵ˈᡰԕਟԕ䇙㌫㔏ӗ⭏ᰐ䲀ཊⲴ䇝㓳ᮠᦞˈਚ㾱ᰦ䰤‫ݱ‬䇨Ǆ⧠൘ᴹҶањᆼᮤ
ⲴᆖҐԫ࣑Ǆ
㾯⌻䐣ỻᆖҐ䰞仈˖
x ᙗ㜭ḷ߶ P˖ц⭼䭖ḷ䎋кࠫ䍕ሩ᡻ⲴⲮ࠶∄
x 䇝㓳㓿傼 E˖઼㠚ᐡ䘋㹼ሩᔸ
ѪҶᆼᡀ䘉њᆖҐ㌫㔏Ⲵ䇮䇑ˈ⧠൘䴰㾱䘹ᤙ˖
1. 㾱ᆖҐⲴ⸕䇶Ⲵ⺞࠷㊫ර
2. ሩҾ䘉њⴞḷ⸕䇶Ⲵ㺘⽪
3. а⿽ᆖҐᵪࡦ
1.2.2 䘿᤟ⴤḽ࠳ᮦ
лањ䇮䇑䘹ᤙᱟߣᇊ㾱ᆖҐⲴ⸕䇶Ⲵ⺞࠷㊫රˈԕ৺ᢗ㹼〻ᒿᘾṧ֯⭘䘉Ӌ⸕䇶Ǆᡁ
ԜӾањሩҾԫօỻተ䜭㜭ӗ⭏ਸ⌅˄legal˅䎠ᆀⲴ㾯⌻䐣ỻঊᔸ〻ᒿᔰ࿻Ǆ䛓Ѹˈᴰ㓸Ⲵ
〻ᒿӵ享ᆖՊӾ䘉Ӌਸ⌅Ⲵ䎠ᆀѝ䘹ᤙᴰ֣ⲴǄ䘉њᆖҐԫ࣑ԓ㺘Ҷаབྷ㊫ԫ࣑˖ਸ⌅䎠ᆀ
ᇊѹҶḀњ‫ݸ‬傼ᐢ⸕Ⲵᐘབྷᩌ㍒オ䰤ˈնᴰ֣Ⲵᩌ㍒ㆆ⮕ᵚ⸕ǄᖸཊᴰՈॆ䰞仈䜭ਟᖂҾ↔
㊫ˈֻྲሩҾ⭏ӗ䗷〻Ⲵ䈳ᓖ઼᧗ࡦ䰞仈ˈ⭏ӗѝⲴ⇿а↕䜭ᖸ␵ᾊˈն䈳ᓖ䘉Ӌ↕僔Ⲵᴰ
֣ㆆ⮕ᵚ⸕Ǆ
ѪҶᆖҐӾਸ⌅䎠ᆀѝ֌ࠪ䘹ᤙˈᖸ᰾ᱮˈ㾱ᆖҐⲴؑ᚟㊫රቡᱟањ〻ᒿᡆ࠭ᮠˈᆳ
ሩ ԫ օ 㔉 ᇊ Ⲵ ỻ ተ 㜭䘹 ࠪ ᴰ ྭ Ⲵ 䎠 ⌅ Ǆ ਟ 〠 ↔ ࠭ ᮠ Ѫ ChooseMove ˈ ᒦ ⭘ 䇠 ⌅
ChooseMove:BM ᶕ㺘⽪䘉њ࠭ᮠԕਸ⌅ỻተ䳶ਸѝⲴỻⴈ⣦ᘱ֌Ѫ䗃‫ˈޕ‬ᒦӾਸ⌅䎠ᆀ
䳶ਸѝӗ⭏Ḁњ䎠ᆀ֌Ѫ䗃ࠪǄ൘‫ޣ‬ҾᵪಘᆖҐⲴᡰᴹ䇘䇪ѝˈᡁԜਁ⧠ਟԕᢺሩԫ࣑ T
ᨀ儈ᙗ㜭 P Ⲵ䰞仈ㆰॆѪᆖҐ䊑 ChooseMove 䘉ṧḀњ⢩ᇊⲴⴞ ⴞḷ࠭ᮠ˄target function˅
Ⲵ䰞仈Ǆᡰԕⴞḷ࠭ᮠⲴ䘹ᤙᱟањ‫ޣ‬䭞Ⲵ䇮䇑䰞仈Ǆ
ቭ㇑൘ֻᆀѝᖸ᰾ᱮᓄᢺ ChooseMove ֌Ѫⴞḷ࠭ᮠˈնᡁԜՊਁ⧠ᆖҐ䘉њⴞḷ࠭ᮠ

ᱟ䶎ᑨഠ䳮Ⲵˈ৏ഐᱟᨀ‫׋‬㔉㌫㔏Ⲵᱟ䰤᧕Ⲵ䇝㓳㓿傼Ǆਖཆањਟ‫׋‬䘹ᤙⲴⴞḷ࠭ᮠᱟа
њ䇴ՠ࠭ᮠˈᆳѪԫօ㔉ᇊỻተ䍻ҸањᮠᆇⲴ䇴࠶Ǆਟԕਁ⧠ˈሩҾᵜֻˈᆖҐ䘉њⴞḷ
࠭ᮠᴤㆰঅǄԔ䘉њⴞḷ࠭ᮠѪ Vˈᒦ⭘ V˖B ᶕ㺘⽪ V ᢺԫօਸ⌅Ⲵỻተ᱐ሴࡠḀа
њᇎᮠ٬˄⭘ᶕ㺘⽪ᇎᮠ䳶ਸ˅ǄᡁԜᢃ㇇䇙䘉њⴞḷ࠭ᮠ V 㔉ྭⲴỻተ䍻Ҹ䖳儈Ⲵ䇴࠶Ǆ
ྲ᷌㌫㔏㜭ཏᡀ࣏ൠᆖՊ䘉њⴞḷ࠭ᮠ Vˈ䛓Ѹᆳ‫ׯ‬㜭֯⭘↔࠭ᮠ䖫ᶮൠ᢮ࡠᖃࡽỻተⲴᴰ
֣䎠⌅Ǆᇎ⧠Ⲵᯩ⌅ᱟˈ‫ݸ‬ӗ⭏⇿ањਸ⌅䎠ᆀሩᓄⲴᡰᴹਾ㔝ỻተˈ❦ਾ֯⭘ V ᶕ䘹ਆ
ަѝᴰ֣Ⲵਾ㔗ỻተˈӾ㘼䘹ᤙᴰྭⲴ䎠ᆀǄ
ሩҾԫ᜿ỻተˈⴞḷ࠭ᮠ V Ⲵ߶⺞٬ᓄ䈕ᱟཊቁ઒˛ᖃ❦ԫօሩ䖳ྭⲴỻተ䍻Ҹ䖳儈
Ⲵ࠶ᮠⲴ䇴ՠ࠭ᮠ䜭䘲⭘Ǆ❦㘼ˈᴰྭ൘䛓Ӌӗ⭏ᴰ֣ሩᔸⲴՇཊᯩ⌅ѝᇊѹањ⢩ᇊⲴⴞ
ḷ࠭ᮠ VǄਟԕⴻࡠˈ䘉ሶ֯ᗇ䇮䇑ањ䇝㓳㇇⌅ਈᗇㆰঅǄഐ↔ˈሩҾ䳶ਸ B ѝⲴԫ᜿Ⲵ
ỻተ⣦ᘱ bˈᡁԜྲлᇊѹⴞḷ࠭ᮠ V(b)˖
1. ྲ᷌ b ᱟаᴰ㓸Ⲵ㜌ተˈ䛓Ѹ V(b)=100
2. ྲ᷌ b ᱟаᴰ㓸Ⲵ䍏ተˈ䛓Ѹ V(b)=-100
3. ྲ᷌ b ᱟаᴰ㓸Ⲵ઼ተˈ䛓Ѹ V(b)=0
4. ྲ᷌ b нᱟᴰ㓸ỻተˈ䛓Ѹ V(b)=V(b)ˈަѝ bᱟӾ b ᔰ࿻ৼᯩ䜭䟷ਆᴰՈሩ

ᔸਾਟ䗮ࡠⲴ㓸ተǄ
❦㘼ˈ⭡Ҿ䘉њᇊѹⲴ䙂ᖂᙗˈᆳⲴ䘀㇇᭸⦷н儈ˈᡰԕ䘉њᇊѹሩҾ㾯⌻䐣ỻ∄䎋㘵
нਟ⭘Ǆ䲔Ҷᰐ‫ޣ‬㍗㾱Ⲵࡽй⿽㓸ተⲴᛵߥˈሩҾḀањỻⴈ⣦ᘱ˄ᛵߥ 4˅b㾱ߣᇊᆳⲴ
٬V˄b˅䴰㾱ੁࡽᩌ㍒ࡠ䗮㓸ተⲴᡰᴹ䐟㓯ʽ⭡Ҿ䘉њᇊѹн㜭⭡㾯⌻䐣ỻ〻ᒿ儈᭸ൠ䘀
㇇ˈ䘉њᇊѹ㻛〠Ѫнਟ᫽֌ⲴᇊѹǄǄᖃࡽⲴⴞḷᱟਁ⧠ањਟ᫽֌ⲴᇊѹVˈᆳ㜭ཏ㻛㾯
⌻䐣ỻ〻ᒿ⭘ᶕ൘ਸ⨶Ⲵᰦ䰤޵䇴ՠỻተᒦ䘹ਆ䎠⌅Ǆ
䘉ṧˈ䘉⿽ᛵߥⲴᆖҐԫ࣑㻛ㆰॆᡀਁ⧠ањ⨶ᜣⴞḷ࠭ᮠ V Ⲵਟ᫽֌᧿䘠Ǆ䙊ᑨ㾱
ᆼ㖾ൠᆖҐ䘉ṧањ V Ⲵਟ᫽֌Ⲵᖒᔿᱟ䶎ᑨഠ䳮ⲴǄһᇎкˈᡁԜ㓿ᑨᐼᵋᆖҐ㇇⌅ӵ
ᗇࡠⴞḷ࠭ᮠⲴḀњ䘁
䘁լ˄approximation˅ˈ⭡Ҿ䘉њ৏ഐᆖҐⴞḷ࠭ᮠⲴ䗷〻ᑨ㻛〠Ѫ࠭
ᮠ䙬䘁˄function approximation˅Ǆ൘ᖃࡽⲴ䇘䇪ѝˈ⭘ Vˆ ᶕ㺘⽪〻ᒿѝᇎ䱵ᆖҐࡠⲴ࠭ᮠˈ
ԕ४࡛⨶ᜣⴞḷ࠭ᮠ VǄ
1.2.3 䘿᤟ⴤḽ࠳ᮦⲺ㺞⽰
㠣↔ˈᡁԜᐢ㓿⺞ᇊҶⴞḷ࠭ᮠ Vˈ᧕лᶕᗵ享䘹ᤙањ㺘⽪ˈ㻛ᆖҐ〻ᒿ⭘ᶕ᧿䘠㾱
ᆖҐⲴ࠭ᮠ Vˆ Ǆሩ↔ҏᴹᖸཊ䇮䇑䘹ᤙǄֻྲˈਟԕሶ Vˆ 㺘⽪Ѫаᕐབྷ㺘ˈሩҾ⇿њᜏа
Ⲵỻⴈ⣦ᘱ bˈ㺘ѝᴹᜏаⲴ㺘亩ᶕ⺞ᇊᆳⲴ⣦ᘱ٬ Vˆ (b)Ǆᡆ㘵ˈਟԕ䇙〻ᒿ⭘ањ㿴ࡉ䳶
ਸᶕ३䝽ỻተⲴ⢩ᖱԕ㺘⽪ Vˆ ˈᡆ䟷⭘ањо亴ᇊѹỻⴈ⢩ᖱᴹ‫Ⲵޣ‬Ҽ⅑ཊ亩ᔿ࠭ᮠˈᡆ
㘵⭘Ӫᐕ⾎㓿‫ݳ‬㖁㔌Ǆ䙊ᑨˈ䘹ᤙ䘉њ᧿䘠वਜ਼ањ䟽㾱Ⲵᵳ㺑䗷〻Ǆаᯩ䶒ˈᡁԜᙫᐼᵋ
䘹ਆањ䶎ᑨᴹ㺘ᖱ࣋Ⲵ᧿䘠ˈԕᴰབྷਟ㜭ൠ䙬䘁⨶ᜣⲴⴞḷ࠭ᮠ VǄਖаᯩ䶒ˈ䎺ᴹ㺘ᖱ
࣋Ⲵ᧿䘠䴰㾱䎺ཊⲴ䇝㓳ᮠᦞˈ֯〻ᒿ㜭Ӿᆳ㺘⽪Ⲵཊ⿽‫ٷ‬䇮ѝ‫ࠪڊ‬䘹ᤙǄѪҶㆰॆ䇘䇪ˈ
⧠൘䘹ᤙањㆰঅⲴ㺘⽪⌅˖ሩҾԫօ㔉ᇊⲴỻⴈ⣦ᘱˈ࠭ᮠ Vˆ ਟԕ䙊䗷ԕлỻⴈ৲ᮠⲴ
㓯ᙗ㓴ਸᶕ䇑㇇˖
z x1˖ỻⴈк唁ᆀⲴᮠ䟿
z x2˖ỻⴈк㓒ᆀⲴᮠ䟿
z x3˖ỻⴈк唁⦻Ⲵᮠ䟿
z x4˖ỻⴈк㓒⦻Ⲵᮠ䟿
z x5˖㻛㓒ᆀေ㛱Ⲵ唁ᆀᮠ䟿˄ণՊ൘ла⅑㻛㓒ਲ਼ᦹⲴᆀ˅
z x6˖㻛唁ᆀေ㛱Ⲵ㓒ᆀᮠ䟿
ҾᱟˈᆖҐ〻ᒿᢺ Vˆ (b)㺘⽪Ѫањ㓯ᙗ࠭ᮠ
Vˆ (b)=w0+w1x1+w2x2+w3x3+w4x4+w5x5+w6x6
ަѝw0ࡠw6Ѫᮠᆇ㌫ᮠˈᡆਛᵳˈ⭡ᆖҐ㇇⌅ᶕ䘹ᤙǄ൘ߣᇊḀањỻⴈ⣦ᘱⲴ࠶٬ᰦˈ
w1 ࡠ w6ߣᇊҶн਼Ⲵỻⴈ⢩ᖱⲴ⴨ሩ䟽㾱ᙗˈ㘼ᵳw0Ѫањ䱴࣐Ⲵᑨ䟿Ǆ
ᾲᤜалⴞࡽѪ→Ⲵ䇮䇑ǄᡁԜᐢ㓿䈖㓶䱀䘠Ҷ䘉њᆖҐ䰞仈Ⲵ৏රˈণѪᆳ䘹ᤙа⿽
㊫රⲴ䇝㓳㓿傼ǃањ㾱ᆖҐⲴⴞḷ࠭ᮠ઼䘉њⴞḷ࠭ᮠⲴа⿽㺘⽪⌅Ǆ⧠൘ⲴᆖҐԫ࣑ᱟ˖
㾯⌻䐣ỻ〻ᒿⲴ䜘࠶䇮䇑
x ᙗ㜭ḷ߶ P˖ц⭼䭖ḷ䎋кࠫ䍕ሩ᡻ⲴⲮ࠶∄
x 䇝㓳㓿傼 E˖઼㠚ᐡ䘋㹼䇝㓳ሩᔸ
x ⴞḷ࠭ᮠ˖V˖B
x ⴞḷ࠭ᮠⲴ㺘⽪˖ Vˆ (b)=w0+w1x1+w2x2+w3x3+w4x4+w5x5+w6x6
ࡽйᶑᱟሩᆖҐԫ࣑Ⲵ䈤᰾ˈਾєᶑࡦᇊҶѪᇎ⧠䘉њᆖҐ〻ᒿⲴ䇮䇑ᯩṸǄ⌘᜿䘉њ
䇮䇑Ⲵ‫ޣ‬䭞֌⭘ᱟᢺᆖҐ㾯⌻䐣ỻᡈ⮕Ⲵ䰞仈ㆰॆѪᆖҐⴞḷ࠭ᮠ᧿䘠ѝ㌫ᮠw0ࡠw6٬Ⲵ
䰞仈Ǆ
1.2.4 䘿᤟࠳ᮦ䙲䘇㇍⌋
ѪҶᆖҐⴞḷ࠭ᮠ Vˆ ˈ䴰㾱а㌫ࡇ䇝㓳ṧֻˈ⇿ањṧֻ᧿䘠Ҷ⢩ᇊⲴỻⴈ⣦ᘱb઼
ᆳⲴ䇝㓳٬Vtrain(b)Ǆᦒ䀰ѻˈ⇿ањ䇝㓳ṧֻᱟᖒᔿѪ<bˈVtrain(b)>Ⲵᒿ‫ڦ‬ǄѮֻᶕ䈤ˈл
䶒Ⲵ䇝㓳ᇎֻ᧿䘠Ҷањ唁ỻ㜌࡙˄⌘᜿x2=0 㺘⽪㓒ỻᐢ㓿⋑ᴹᆀҶ˅Ⲵỻⴈ⣦ᘱbˈᆳⲴ
ⴞḷ࠭ᮠ٬Vtrain(b)Ѫ 100Ǆ
<<x1=3ˈx2=0ˈx3=1ˈx4=0ˈx5=0ˈx6=0>ˈ+100>
л᮷᧿䘠Ҷањ䗷〻ˈᆳ‫ݸ‬ӾᆖҐಘਟᗇⲴ䰤᧕䇝㓳㓿傼ѝሬࠪк䶒Ⲵ䇝㓳ṧֻˈ❦ਾ
䈳ᮤᵳ٬wiԕᴰ֣ᤏਸ䘉Ӌ䇝㓳ṧֻǄ
1.2.4.1 զ䇗䇣㓹‫ٲ‬
ṩᦞԕкⲴᆖҐ⁑රˈᆖҐಘਟԕᗇࡠⲴ䇝㓳ؑ᚟ӵᱟሩᔸᴰਾⲴ㜌䍏Ǆ ਖаᯩ䶒ˈ
ᡁԜ䴰㾱䇝㓳ṧֻѪ⇿њỻⴈ⣦ᘱ䍻Ҹањ࠶٬Ǆ㔉ሩᔸ㔃ᶏᰦⲴỻⴈ⣦ᘱ䇴࠶ᱟᇩ᱃Ⲵˈ
㘼㾱㔉ሩᔸ㔃ᶏࡽⲴབྷ䟿ѝ䰤ỻተ䇴࠶ቡн䛓Ѹᇩ᱃ҶǄഐѪˈаⴈỻⲴᴰ㓸䗃䎒ᵚᗵ㜭䈤
᰾䘉ⴈỻᖃѝⲴ⇿ањỻⴈ⣦ᘱⲴྭᡆൿǄֻྲˈণ֯Ḁњ〻ᒿ䗃ҶаⴈỻˈӽՊᴹ䘉ṧⲴ
ᛵߥˈ䘉ⴈỻࡽ䶒Ⲵỻተᓄ䈕㔉Ҹᖸ儈Ⲵ䇴ԧˈཡ䍕Ⲵ৏ഐ൘Ҿਾᶕ㌏㌅Ⲵ䎠⌅Ǆ
ቭ㇑ՠ䇑ѝ䰤ỻተ䇝㓳٬ާᴹ޵൘Ⲵ⁑㋺ᙗˈնԔӪ᛺䇦ⲴᱟᴹањㆰঅⲴᯩ⌅তਆᗇ
Ҷ㢟ྭ㔃᷌Ǆ䘉⿽ᯩ⌅ሩҾԫօѝ䰤ỻተbⲴ䇝㓳٬Vtrain(b)ㅹҾ Vˆ (Successor(b))ˈަѝ Vˆ ᱟ
ᆖҐಘ䟷⭘ⲴሩVⲴ䘁լˈSuccessor(b) 㺘⽪bѻਾ޽䖞ࡠ〻ᒿ䎠ỻᰦⲴỻⴈ⣦ᘱ˄ҏቡᱟ〻
ᒿ䎠Ҷа↕઼ሩ᡻എᓄа↕ਾⲴỻተ˅Ǆ
䘉⿽ՠ䇑䇝㓳٬Ⲵᯩ⌅ਟ㻛ᖂ㓣Ѫ˖
䇝㓳٬ՠ䇑⌅ࡉ
Vtrain(b) Vˆ (Successor(b)) ˄1.1˅
ᡆ䇨䘉ⴻ䎧ᶕᴹ⛩⿫ཷˈਚ֯⭘ᖃࡽⲴ Vˆ ᶕՠ䇑䇝㓳٬ˈ䘉а䇝㓳٬৸㻛⭘ᶕᴤᯠ Vˆ Ǆ
ն䈧⌘᜿ˈᡁԜᱟ൘⭘ਾ㔝ỻተSuccessor(b)Ⲵՠ䇑٬ᶕՠ䇑ỻተbⲴ٬Ǆࠝⴤ㿹ˈᡁԜਟԕ
ⴻࡠ䎺᧕䘁⑨ᠿ㔃ᶏⲴỻተⲴ Vˆ 䎺䎻ੁ㋮⺞Ǆһᇎкˈ൘⢩ᇊᶑԦл˄ሶ൘ㅜ 13 ㄐ䇘䇪˅
䘉⿽สҾሩਾ㔗ỻተ䘋㹼ՠ䇑Ⲵ䘝ԓՠ䇑䇝㓳٬Ⲵᯩ⌅ˈᐢ㻛䇱᰾ਟԕ䘁Ѿᆼ㖾ൠ᭦ᮋࡠ
Vtrainՠ䇑٬Ǆ
1.2.4.2 ᵹ‫ٲ‬䈹᮪
࢙лⲴһᛵቡᱟѪ䘉њᆖҐ㇇⌅䘹ᤙᴰ䘲ਸ䇝㓳ṧֻ{<b, Vtrain(b)>}ⲴᵳwiǄㅜа↕ᗵ享
ᇊѹᴰ֣ᤏਸ˄best fit˅䇝㓳ᮠᦞⲴਜ਼ѹǄа⿽ᑨ⭘Ⲵᯩ⌅ᱟᢺᴰ֣Ⲵ‫ٷ‬䇮˄ᡆᵳੁ䟿䳶ਸ˅
ᇊѹѪ֯䇝㓳٬઼‫ٷ‬䇮 Vˆ 亴⍻ࠪⲴ٬䰤Ⲵ䈟ᐞᒣᯩEᴰሿǄ
E{ ¦
b ,V train ( b ) ! 䇝㓳ṧֻ
(Vtrain ( b ) Vˆ ( b )) 2
㠣↔ˈᡁԜⲴⴞḷቡᱟራ᢮ᵳ٬˄ㅹԧൠˈራ᢮ Vˆ ˅ˈ֯ሩҾ㿲⍻ࡠⲴ䇝㓳ᮠᦞ E ٬ᴰ
ሿॆǄㅜ 6 ㄐሶ䇘䇪൘ӰѸᶑԦлˈᴰሿॆ䈟ᐞᒣᯩ઼ㅹԧҾራ᢮㔉ᇊ㿲⍻䇝㓳ᮠᦞлⲴᴰ
ਟ㜭‫ٷ‬䇮Ǆ
ᐢ㓿⸕䚃аӋ㇇⌅ਟԕᗇࡠ㓯ᙗ࠭ᮠⲴᵳ֯↔ᇊѹⲴ E ᴰሿॆǄ൘䘉䟼䴰㾱ањ㇇⌅ˈ
ᆳਟԕ൘ᴹҶᯠⲴ䇝㓳ṧֻᰦ䘋а↕᭩䘋ᵳ٬ˈᒦфᆳሩՠ䇑Ⲵ䇝㓳ᮠᦞѝⲴᐞ䭉ᴹྭⲴ‫ڕ‬
༞ᙗǄањ䘉ṧⲴ㇇⌅㻛〠֌ᴰሿ൷ᯩ⌅˄least mean squares˅ˈᡆਛ LMS 䇝㓳⌅ࡉǄሩҾ
⇿а䇝㓳ṧֻˈᆳᢺᵳ٬ੁ߿ሿ䘉њ䇝㓳ᮠᦞ䈟ᐞⲴᯩੁ⮕ᗞ䈳ᮤǄྲㅜ 4 ㄐ䇘䇪Ⲵ䛓ṧˈ
䘉њ㇇⌅ਟ㻛ⴻ֌ሩਟ㜭Ⲵ‫ٷ‬䇮˄ᵳ٬˅オ䰤䘋㹼䲿ᵪⲴởᓖл䱽ᩌ㍒ˈԕ֯䈟ᐞᒣᯩ઼ E
ᴰሿॆǄLMS ㇇⌅ᱟ䘉ṧᇊѹⲴ˖
LMS ᵳ٬ᴤᯠ⌅ࡉ
ሩҾ⇿ањ䇝㓳ṧֻ<bˈVtrain(b)>
x ֯⭘ᖃࡽⲴᵳ䇑㇇ Vˆ (b)
x ሩ⇿ањᵳ٬wi䘋㹼ྲлᴤᯠ
wiĕwi+(Vtrain(b)- Vˆ (b)) xi
䘉䟼¨ᱟањሿⲴᑨᮠ˄∄ྲ 0.1˅⭘ᶕ䈳ᮤᵳ٬ᴤᯠⲴᑵᓖǄѪҶⴤ㿲ൠ⨶䀓䘉њᵳ
٬ᴤᯠ⌅ࡉⲴᐕ֌৏⨶ˈ䈧⌘᜿ᖃ䈟ᐞ˄Vtrain(b)- Vˆ (b)˅Ѫ 0 ᰦˈᵳнՊ㻛᭩ਈǄᖃ
(Vtrain(b)- Vˆ (b))Ѫ↓ᰦ˄ֻྲˈᖃ Vˆ (b)ཚվᰦ˅⇿ањᵳ٬Պṩᦞަሩᓄ⢩ᖱ٬໎࣐аᇊⲴ
∄ֻǄ䘉Պᨀॷ Vˆ (b)Ⲵ٬㘼߿ሿ䈟ᐞǄ⌘᜿ྲ᷌Ḁњ৲ᮠxiѪ 0ˈ䛓ѸᆳⲴ٬нՊഐ䘉њ䈟
ᐞ㘼᭩ਈˈ䘉ṧ‫֯ׯ‬ਚᴹ䛓Ӌ൘䇝㓳ṧֻⲴỻተѝ⺞ᇎࠪ⧠Ⲵ⢩ᖱⲴᵳ٬᡽㻛ᴤᯠǄԔӪਲ਼
᛺Ⲵᱟˈ൘аᇊⲴᶑԦлˈ䘉⿽ㆰঅⲴᵳ٬䈳ᮤᯩ⌅㻛䇱᰾ਟԕ᭦ᮋࡠVtrain ٬Ⲵᴰሿ䈟ᐞ
ᒣᯩ䙬䘁˄ቡ‫ۿ‬ㅜ 4 ㄐᡰ䇘䇪Ⲵ˅Ǆ
1.2.5 ᴶ㓾Ⲻ䇴䇗
㾯⌻䐣ỻᆖҐ㌫㔏Ⲵᴰ㓸䇮䇑ਟԕ㠚❦ൠ⭘ഋњ␵ᾊⲴ〻ᒿ⁑ඇᶕ᧿䘠ˈ䘉Ӌ⁑ඇ൘ᖸ
ཊᆖҐ㌫㔏ѝᱟṨᗳ㓴ԦǄ䘉ഋњ⁑ඇ㻛ᖂ㓣൘മ 1-1 ѝˈᆳԜᱟ˖
ᢗ㹼㌫㔏˄Performance system˅ˈ䘉њ⁑ඇᱟ⭘ᆖՊⲴⴞḷ࠭ᮠᶕ䀓ߣ㔉ᇊⲴԫ࣑ˈ
൘↔ቡᱟሩᔸ㾯⌻䐣ỻǄᆳᢺᯠ䰞仈˄ᯠаⴈỻ˅Ⲵᇎֻ֌Ѫ䗃‫ˈޕ‬ӗ⭏а㓴䀓ㆄ䐟㓯˄ሩ
ᔸশਢ䇠ᖅ˅֌Ѫ䗃ࠪǄ൘䘉䟼ˈᢗ㹼㌫㔏䟷⭘Ⲵ䘹ᤙла↕䎠⌅Ⲵㆆ⮕ᱟ⭡ᆖࡠⲴ䇴ՠ࠭
ᮠ Vˆ ᶕߣᇊⲴǄᡰԕᡁԜᵏᖵᆳⲴᙗ㜭Պ䲿⵰䇴ՠ࠭ᮠⲴᰕ⳺߶⺞㘼ᨀ儈Ǆ
插图——原书页码： 12
Experiment Generator-䈅傼⭏ᡀಘ
New Problem(initial game board)-ᯠ䰞仈˄ࡍ࿻ỻተ˅
Performance System-ᢗ㹼㌫㔏
Solution trace(game history)-䀓ㆄ䐟㓯˄ሩᔸশਢ˅
Critic-䢤ᇊಘ
Training examples-䇝㓳ṧֻ
Generalizer-⌋ॆಘ
Hypothesis-‫ٷ‬䇮
മ 1-1 㾯⌻䐣ỻᆖҐ〻ᒿⲴᴰ㓸䇮䇑
䢤ᇊಘ˄Critic˅ˈᆳԕሩᔸⲴ䐟㓯ᡆশਢ䇠ᖅ֌Ѫ䗃‫ˈޕ‬䗃ࠪⴞḷ࠭ᮠⲴа㌫ࡇ䇝㓳ṧ
ֻǄྲമᡰ⽪ˈ⇿ањ䇝㓳ṧֻሩᓄ䐟㓯ѝⲴḀњỻⴈ⣦ᘱ઼ⴞḷ࠭ᮠ㔉䘉њṧֻⲴ䇴ՠ٬
VtrainǄ൘ᡁԜⲴֻᆀѝˈ䢤ᇊಘሩᓄᔿ 1.1 㔉ࠪⲴ䇝㓳⌅ࡉǄ
⌋ॆಘ˄Generalizer˅ˈᆳԕ䇝㓳ṧֻ֌Ѫ䗃‫ˈޕ‬䗃ࠪањ‫ٷ‬䇮ˈ֌Ѫᆳሩⴞḷ࠭ᮠⲴ
ՠ䇑ǄᆳӾ⢩ᇊⲴ䇝㓳ṧֻѝ⌋ॆˈ⥌⍻ања㡜࠭ᮠˈ֯ަ㜭ཏ㾶ⴆ䘉Ӌṧֻԕ৺ṧֻѻ
ཆⲴᛵᖒǄ൘ᡁԜⲴֻᆀѝˈ⌋ॆಘሩᓄLMS㇇⌅ˈ䗃ࠪ‫ٷ‬䇮ᱟ⭘ᆖҐࡠⲴᵳ٬w0 ,..., w6᧿
ˆ
䘠Ⲵ࠭ᮠ V Ǆ
ᇎ傼⭏ᡀಘ˄Experiment Generator˅ˈᆳԕᖃࡽⲴ‫ٷ‬䇮˄ᖃࡽᆖࡠⲴ࠭ᮠ˅֌Ѫ䗃‫ˈޕ‬
䗃ࠪањᯠⲴ䰞仈˄ֻྲˈᴰࡍⲴỻተ˅‫׋‬ᢗ㹼㌫㔏৫᧒㍒ǄᆳⲴ䀂㢢ᱟ᥁䘹ᯠⲴ㓳Ґ䰞仈ˈ
ԕ֯ᮤњ㌫㔏ⲴᆖҐ䙏⦷ᴰབྷॆǄ൘ᡁԜⲴֻᆀѝˈᇎ傼⭏ᡀಘ䟷⭘Ҷ䶎ᑨㆰঅⲴㆆ⮕˖ᆳ
ᙫᱟ㔉ࠪањ਼ṧⲴࡍ࿻ỻተᶕᔰ࿻ᯠⲴаⴈỻǄᴤᆼழⲴㆆ⮕ਟ㜭㠤࣋Ҿ㋮ᗳ䇮䇑ỻᆀս
㖞ԕ᧒㍒ỻⴈオ䰤Ⲵ⢩ᇊ४ฏǄ
ᙫփᶕⴻˈᡁԜѪ㾯⌻䐣ỻ〻ᒿ֌Ⲵ䇮䇑ቡᱟӗ⭏ᢗ㹼㌫㔏ǃ䢤ᇊಘǃ⌋ॆಘ઼ᇎ傼⭏
ᡀಘⲴ⢩ᇊᇎֻǄᖸཊᵪಘᆖҐ㌫㔏䙊ᑨਟԕ⭘䘉ഋњ䙊⭘⁑ඇᶕ࡫⭫Ǆ
䇮䇑㾯⌻䐣ỻ〻ᒿⲴ⍱〻㻛ᖂ㓣൘മ 1-2 ѝǄ䘉њ䇮䇑ᐢ㓿൘ࠐᯩ䶒ᢺᆖҐԫ࣑䲀ࡦ൘

䖳ሿⲴ㤳ത޵Ǆ㾱ᆖҐⲴ⸕䇶㊫ර㻛䲀ࡦѪањঅаⲴ㓯ᙗ䇴ՠ࠭ᮠǄ㘼ф䘉њ䇴ՠ࠭ᮠ㻛
䲀ࡦѪӵ‫׍‬䎆Ҿ‫ޝ‬њỻⴈ⢩ᖱǄྲ᷌ⴞḷ࠭ᮠⵏⲴਟ㺘⽪Ѫ䘉Ӌ⢩ᇊ৲ᮠⲴ㓯ᙗ㓴ਸˈ䛓Ѹ
〻ᒿᆖࡠ䘉њⴞḷ࠭ᮠⲴਟ㜭ᙗᖸབྷǄ৽ѻˈᴰཊਚᐼᵋᆳᆖࡠањਸ⨶Ⲵ䘁լˈഐѪањ
〻ᒿᖃ❦н㜭ᆖՊᆳṩᵜн㜭㺘⽪Ⲵь㾯Ǆ
插图——原书页码： 13
Determine Type of Training Experience- ߣᇊ䇝㓳㓿傼ᖒᔿ
Games against experts- оуᇦሩᔸ
Games against self- о㠚ᐡሩᔸ
Table of correct moves- ↓⺞䎠ᆀⲴ㺘Ṭ
Determine Target Function- ߣᇊⴞḷ࠭ᮠ
Board->move- ỻⴈo䎠ᆀ
Board->value- ỻⴈo࠶٬
Determine Representation of Learned Function-ߣᇊⴞḷ࠭ᮠⲴ㺘⽪
Polynomial- ཊ亩ᔿ
Linear function of six features- ‫ޝ‬њ৲ᮠⲴ㓯ᙗ࠭ᮠ
Artificial neural network- Ӫᐕ⾎㓿㖁㔌
Determine Learning Algorithm- ߣᇊᆖҐ㇇⌅
Gradient descent- ởᓖл䱽
Linear programming- 㓯ᙗ㿴ࡂ
Completed Design- ᆼᡀⲴ䇮䇑
മ 1-2 㾯⌻䐣ỻᆖҐ〻ᒿⲴ䇮䇑䗷〻ᾲ䘠
ᡁԜ‫ٷ‬ᇊⵏᇎ࠭ᮠ V Ⲵਸ⨶Ⲵ䘁լ⺞ᇎਟ㻛㺘⽪Ѫ䘉⿽ᖒᔿǄ䛓Ѹ䰞仈ਈᡀ䘉⿽ᆖҐ
ᢰᵟᱟ੖⺞‫؍‬㜭ਁ⧠ањਸ⨶Ⲵ䘁լǄㅜ 13 ㄐᨀ‫׋‬Ҷа⿽⨶䇪࠶᷀ˈ㺘᰾ሩҾḀӋ㊫රⲴ
ᩌ㍒䰞仈ˈ൘⴨ᖃѕṬⲴࡽᨀлˈ䘉⿽ᯩ⌅⺞ᇎ᭦ᮋࡠᵏᵋⲴ䇴ՠ࠭ᮠǄᖸᒨ䘀ˈᇎ䐥㓿傼
㺘᰾䘉⿽ᆖҐ䇴ՠ࠭ᮠⲴ䙄ᖴ㓿ᑨᱟᡀ࣏Ⲵˈ⭊㠣൘㜭㻛䇱᰾Ⲵᛵᖒѻཆҏᱟྲ↔Ǆ
ᐢ㓿䇮䇑Ⲵ〻ᒿ㜭ᆖᗇ䏣ཏྭ㘼ࠫ䍕Ӫ㊫Ⲵ㾯⌻䐣ỻߐߋੇ˛ᡆ䇨н㜭Ǆ䜘࠶ൠˈ䘉ᱟ
ഐѪ Vˆ Ⲵ㓯ᙗ࠭ᮠ㺘⽪ཚㆰঅԕ㠤Ҿн㜭ᖸྭᦅ᥹䘉⿽ỻⲴᗞ࿉ѻ༴Ǆ❦㘼ˈྲ᷌㔉оа
њᴤᆼழⲴⴞḷ࠭ᮠ㺘⽪⌅ˈ䘉⿽䙊⭘Ⲵ䙄ᖴһᇎкਟԕ䶎ᑨᡀ࣏ǄֻྲˈTesauro(1992, 1995)
ᣕ੺ҶᆖҐл㾯⌻ৼ䱶ỻⲴ〻ᒿⲴ㊫լ䇮䇑ˈᯩ⌅ᱟᆖҐањ䶎ᑨ㊫լⲴỻተ䇴ՠ࠭ᮠǄᆳ
Ⲵ〻ᒿ֯⭘Ӫᐕ⾎㓿‫ݳ‬㖁㔌㺘⽪ᆖࡠⲴ䇴ՠ࠭ᮠˈᆳ㘳㲁ሩỻተⲴᆼᮤ᧿䘠㘼нᱟỻⴈⲴࠐ
њ৲ᮠǄ㓿শҶаⲮз⅑ԕкⲴ㠚ᡁ⭏ᡀⲴ䇝㓳∄䎋ਾˈԆⲴ〻ᒿ㜭ཏ઼а⍱ⲴӪ㊫㾯⌻ৼ
䱶ỻ䘹᡻аҹ儈лǄ
ᖃ❦䘈ਟ㜭Ѫ㾯⌻䐣ỻᆖҐԫ࣑䇮䇑ᖸཊަԆⲴ㇇⌅Ǆֻྲˈа⿽ਟ㜭ਚㆰঅൠᆈۘ䇝
㓳ṧֻˈ❦ਾ৫ራ᢮‫؍‬ᆈⲴĀᴰ᧕䘁Ⲵāᛵᖒᶕ३䝽ᯠⲴᛵߥ˄ᴰ䘁䛫㇇⌅ˈㅜ 8 ㄐ˅Ǆᡆ
㘵ਟԕӗ⭏བྷ䟿‫ى‬䘹Ⲵ㾯⌻䐣ỻ〻ᒿˈᒦ䇙ᆳԜ⴨ӂ∄䎋ˈ‫⮉؍‬ᴰᡀ࣏Ⲵ〻ᒿᒦ䘋а↕⭘⁑
ᤏ䘋ॆⲴᯩᔿᶕษ㛢ᡆਈᔲᆳԜ˄䚇Ր㇇⌅ˈㅜ 9 ㄐ˅ǄӪ㊫լѾ䚥ᗚਖа⿽䙄ᖴራ᢮ᆖҐ
ㆆ⮕ˈԆԜ࠶᷀ᡆੁ㠚ᐡ䀓䟺∄䎋ѝ⻠ࡠⲴᡀ䍕Ⲵ৏ഐ˄สҾ䀓䟺ⲴᆖҐˈㅜ 11 ㄐ˅Ǆк䶒
Ⲵ䇮䇑ᱟ䘉Ӌ⿽㊫ѝⲴањㆰঅⲴ㇇⌅ˈᆳᱟѪҶ㔉ᡁԜӺਾⲴ䪸ሩ⢩ᇊ㊫࡛Ⲵԫ࣑ⲴᆖҐ
ᯩ⌅Ⲵ䇮䇑ྐᇊส⹰Ǆ
1.3 ᵰಞᆜҖⲺжӑ㿸⛯ૂ䰤从
൘ᵪಘᆖҐᯩ䶒ˈањᴹ᭸Ⲵ㿲⛩ᱟᵪಘᆖҐ䰞仈㓿ᑨᖂ㔃Ҿᩌ㍒䰞仈ˈণሩ䶎ᑨབྷⲴ
‫ٷ‬䇮オ䰤䘋㹼ᩌ㍒ˈԕ⺞ᇊᴰ֣ᤏਸ㿲ሏࡠⲴᮠᦞ઼ᆖҐಘᐢᴹ⸕䇶Ⲵ‫ٷ‬䇮Ǆֻྲˈ㘳㲁а
лк䶒Ⲵ㾯⌻䐣ỻᆖҐ〻ᒿ䗃ࠪⲴ‫ٷ‬䇮オ䰤Ǆ䘉њ‫ٷ‬䇮オ䰤वਜ਼ᡰᴹਟ⭡ᵳw0ࡠw6Ⲵн਼٬
Ⲵ䇴ՠ࠭ᮠǄҾᱟᆖҐಘⲴԫ࣑ቡᱟᩌ㍒䘉њབྷⲴオ䰤ˈራ᢮о䇝㓳ᮠᦞᴰ֣ᤏਸⲴ‫ٷ‬䇮Ǆ
䪸ሩᤏਸᵳ٬ⲴLMS㇇⌅䙊䗷䘝ԓ䈳ᮤᵳ٬ᇎ⧠Ҷ䘉њⴞⲴˈ⇿ᖃ‫ٷ‬䇮Ⲵ䇴ՠ࠭ᮠ亴⍻ࠪ
ањо䇝㓳ᮠᦞᴹ‫ٿ‬ᐞⲴ٬ᰦቡሩ⇿њᵳ٬䘋㹼ṑ↓ǄᖃᆖҐಘ㘳㲁Ⲵ‫ٷ‬䇮㺘⽪ᇊѹҶањ
䘎㔝Ⲵ৲ᮠॆⲴ▌൘‫ٷ‬䇮オ䰤ᰦˈ䘉њ㇇⌅ᖸᴹ᭸Ǆ
ᵜҖⲴᖸཊㄐ㢲㔉ࠪҶሩаӋสᵜ㺘⽪˄ֻྲˈ㓯ᙗ࠭ᮠǃ䙫䗁᧿䘠ǃߣㆆṁǃӪᐕ⾎
㓿‫ݳ‬㖁㔌˅ᇊѹⲴ‫ٷ‬䇮オ䰤Ⲵᩌ㍒㇇⌅Ǆ䘉Ӌн਼Ⲵ‫ٷ‬䇮㺘⽪⌅䘲ਸҾᆖҐн਼Ⲵⴞḷ࠭ᮠǄ
ሩҾަѝⲴ⇿а⿽‫ٷ‬䇮㺘⽪⌅ˈሩᓄⲴᆖҐ㇇⌅ਁᥕн਼޵൘㔃ᶴⲴՈ࣯ᶕ㓴㓷ሩ‫ٷ‬䇮オ䰤
Ⲵᩌ㍒Ǆ
㠚࿻㠣㓸ˈᵜҖ䜭䍟ク⵰䘉⿽ᢺᆖҐ䰞仈㿶Ѫᩌ㍒䰞仈Ⲵⴻ⌅ˈӾ㘼䙊䗷ᩌ㍒ㆆ⮕઼ᆖ
Ґಘ᧒㍒Ⲵᩌ㍒オ䰤Ⲵ޵൘㔃ᶴᶕ࡫⭫ᆖҐᯩ⌅ǄᡁԜҏՊਁ⧠ˈ䘉⿽㿲⛩ሩҾᖒᔿॆൠ࠶
᷀㾱ᩌ㍒Ⲵ‫ٷ‬䇮オ䰤Ⲵབྷሿǃਟ࡙⭘Ⲵ䇝㓳ṧֻⲴᮠ䟿ԕ৺ањо䇝㓳ᮠᦞа㠤Ⲵ‫ٷ‬䇮㜭⌋
ॆࡠᵚ㿱ᇎֻⲴ㖞ؑᓖ䘉й㘵ѻ䰤Ⲵ‫ޣ‬㌫䶎ᑨᴹ⭘Ǆ
1.3.1 ᵰಞᆜҖⲺ䰤从
㾯⌻䐣ỻֻᆀᨀࠪҶᵪಘᆖҐᯩ䶒ᖸཊᲞ䙽䰞仈ǄᵪಘᆖҐ䘉䰘ᆖ、ˈ઼ᵜҖⲴ㔍བྷ䜘
࠶ˈ䜭㠤࣋Ҿഎㆄ㊫լл䶒Ⲵ䰞仈˖
x Ӿ⢩ᇊⲴ䇝㓳ᮠᦞᆖҐа㡜Ⲵⴞḷ࠭ᮠᆈ൘ӰѸṧⲴ㇇⌅˛ྲ᷌ᨀ‫׋‬Ҷ‫ݵ‬
䏣Ⲵ䇝㓳ᮠᦞˈӰѸṧⲴᶑԦлՊ֯⢩ᇊⲴ㇇⌅᭦ᮋࡠᵏᵋⲴ࠭ᮠ˛ଚњ㇇
⌅ሩଚӋ䰞仈઼㺘⽪Ⲵᙗ㜭ᴰྭǄ
x ཊቁ䇝㓳ᮠᦞᱟ‫ݵ‬䏣Ⲵ˛ᘾṧ᢮ࡠᆖҐࡠⲴ‫ٷ‬䇮Ⲵ㖞ؑᓖо䇝㓳ᮠᦞⲴᮠ
䟿৺ᨀ‫׋‬㔉ᆖҐಘⲴ‫ٷ‬䇮オ䰤⢩ᙗѻ䰤Ⲵа㡜‫ޣ‬㌫˛
x ᆖҐಘᤕᴹⲴ‫ݸ‬傼⸕䇶ᱟᘾṧᕅሬӾṧֻ䘋㹼⌋ॆⲴ䗷〻Ⲵ˛ᖃ‫ݸ‬傼⸕䇶
ӵӵᱟ䘁լ↓⺞ᰦˈᆳԜՊᴹᑞࣙੇ˛
x ሩҾ䘹ᤙᴹ⭘Ⲵਾ㔝䇝㓳㓿傼ˈӰѸṧⲴㆆ⮕ᴰྭ˛䘉њㆆ⮕Ⲵ䘹ᤙՊᘾṧ
ᖡ૽ᆖҐ䰞仈Ⲵ༽ᵲᙗ˛
x ᘾṧᢺᆖҐԫ࣑ㆰॆѪањᡆཊњ࠭ᮠ䙬䘁䰞仈˛ᦒа⿽ᯩᔿˈ㌫㔏䈕䈅മ
ᆖҐଚӋ࠭ᮠ˛䘉њ䗷〻ᵜ䓛㜭㠚ࣘॆੇ˛
x ᆖҐಘᘾṧ㠚ࣘൠ᭩ਈ㺘⽪⌅ᶕᨀ儈㺘⽪઼ᆖҐⴞḷ࠭ᮠⲴ㜭࣋˛
1.4 ྸ֋䰻䈱ᵢҜ
䘉ᵜҖӻ㓽ҶᵪಘᆖҐⲴѫ㾱㇇⌅઼䙄ᖴ˗н਼ᆖҐԫ࣑ਟ㹼ᙗ઼⢩ᇊ㇇⌅㜭࣋Ⲵ⨶䇪
㔃᷌˗ԕ৺ᵪಘᆖҐᓄ⭘Ҿ䀓ߣ⧠ᇎ䰞仈ⲴֻᆀǄਚ㾱ਟ㜭ˈ਴ㄐⲴ߉֌䜭࣋ҹо䰵䈫亪ᒿ
ᰐ‫ޣ‬Ǆ❦㘼аӋ⴨ӂ‫׍‬䎆ᙗᱟнਟ䚯‫Ⲵݽ‬Ǆྲ᷌ᵜҖ㻛⭘֌ᮉ、Җˈᡁᔪ䇞俆‫ݸ‬ᆼᡀㅜа઼
ㅜҼㄐˈ։л਴ㄐสᵜਟԕԕԫ᜿亪ᒿ䰵䈫Ǆ䮯ᓖѪањᆖᵏⲴᵪಘᆖҐ䈮〻ਟԕवᤜࡽг
ㄐԕ৺仍ཆⲴࠐњᴰᝏ‫ޤ‬䏓Ⲵㄐ㢲Ǆл䶒ㆰ㾱⍿㿸ал਴ㄐǄ
x ㅜ 2 ㄐवᤜสҾㅖਧ઼䙫䗁㺘⽪ⲴᾲᘥᆖҐǄҏ䇘䇪Ҷ‫ٷ‬䇮Ⲵа㡜ࡠ⢩↺‫ٿ‬
ᒿ㔃ᶴˈԕ৺ᆖҐѝᕅ‫ޕ‬ᖂ㓣‫ٿ‬㖞Ⲵᗵ㾱ᙗǄ
x ㅜ 3 ㄐवᤜߣㆆṁᆖҐ઼䗷ᓖᤏਸ䇝㓳ᮠᦞⲴ䰞仈Ǆ䘉аㄐҏࢆ᷀Ҷྕൾ࿶
ࡳ࠰üü䈕৏ࡉᔪ䇞൘оᮠᦞа㠤Ⲵ‫ٷ‬䇮ѝ䘹ᤙᴰ⸝‫ٷ‬䇮Ǆ
x ㅜ 4 ㄐवᤜӪᐕ⾎㓿㖁㔌Ⲵ⸕䇶ˈ⢩࡛ᱟ⹄ウᐢѵⲴ৽ੁՐ᫝㇇⌅ˈԕ৺ở
ᓖл䱽Ⲵа㡜ᯩ⌅Ǆ䘉аㄐवਜ਼ањ䈖㓶ⲴสҾ⾎㓿㖁㔌ⲴӪ㝨䇶࡛ᇎֻˈ
䈕ֻᆀ䴰㾱Ⲵᮠᦞ઼㇇⌅ਟԕ൘з㔤㖁кᗇࡠǄ
x ㅜ 5 ㄐ㔉ࠪҶᶕ㠚㔏䇑઼ՠ䇑⨶䇪Ⲵส⹰ᾲᘥˈ⵰䟽Ҿ֯⭘ᴹ䲀Ⲵṧᵜᮠᦞ
䇴ՠ‫ٷ‬䇮Ⲵ㋮ᓖǄ䘉аㄐवਜ਼Ҷ⭘Ҿՠ䇑‫ٷ‬䇮㋮ᓖⲴ㖞ؑオ䰤ˈ઼ሩн਼ᆖ
Ґ㇇⌅Ⲵ㋮ᓖ䘋㹼∄䖳Ⲵᯩ⌅Ǆ
x ㅜ 6 ㄐӻ㓽ᵪಘᆖҐⲴ䍍ਦᯟ㿲⛩ǄᰒवᤜҶ֯⭘䍍ਦᯟ࠶᷀࡫⭫䶎䍍ਦᯟ
ᆖҐ㇇⌅ˈ৸वᤜҶⴤ᧕༴⨶ᾲ⦷Ⲵ䍍ਦᯟ㇇⌅Ǆ䘉аㄐवᤜањᓄ⭘䍍ਦ
ᯟ࠶㊫ಘᶕ࠶㊫᮷ᵜ᮷ẓⲴ䈖㓶ֻᆀˈᡰ䴰Ⲵᮠᦞ઼䖟Ԧਟԕ൘з㔤㖁кᗇ
ࡠǄ
x ㅜ 7 ㄐ㾶ⴆҶ䇑㇇ᆖҐ⨶䇪ˈवᤜਟ㜭䘁լ↓⺞˄Probably Approximately
CorrectˈPAC˅ᆖҐ⁑ර઼ࠪ䭉⭼䲀˄Mistake-Bound˅ᆖҐ⁑රǄᵜㄐ䇘䇪
Ҷ㚄ਸཊњᆖҐᯩ⌅Ⲵ࣐ᵳཊᮠ˄Weighted Majority˅㇇⌅Ǆ
x ㅜ 8 ㄐ᧿䘠ҶสҾᇎֻⲴᆖҐᯩ⌅ˈवᤜᴰ䘁䛫ᆖҐˈተ䜘࣐ᵳഎᖂˈ઼ส
ҾṸֻⲴ᧘⨶Ǆ
x ㅜ 9 ㄐ䇘䇪Ҷṩᦞ⭏⢙䘋ॆᔪ⁑ⲴᆖҐ㇇⌅ˈवᤜ䚇Ր㇇⌅઼䚇Ր㕆〻Ǆ
x ㅜ 10 ㄐ㾶ⴆҶа㓴ᆖҐ㿴ࡉ䳶ਸⲴ㇇⌅ˈवᤜᆖҐа䱦 Horn ᆀਕⲴᖂ㓣䙫
䗁㕆〻ᯩ⌅Ǆ
x ㅜ 11 ㄐवਜ਼ҶสҾ䀓䟺ⲴᆖҐˈণа⿽֯⭘ԕࡽⲴ⸕䇶䀓䟺㿲ሏࡠⲴᇎֻˈ
❦ਾṩᦞ䘉Ӌ䀓䟺⌋ॆⲴᆖҐᯩ⌅Ǆ
x ㅜ 12 ㄐ䇘䇪ҶᢺԕࡽⲴ䘁լ⸕䇶㔃ਸ䘋⧠ᴹⲴ䇝㓳ᮠᦞѝԕᨀ儈ᆖҐ㋮ᓖ
Ⲵᯩ⌅Ǆ൘ަѝㅖਧ㇇⌅઼⾎㓿㖁㔌㇇⌅䜭ᴹ䇘䇪Ǆ
x ㅜ 13 ㄐ䇘䇪Ҷ໎ᕪᆖҐǄ䘉⿽ᯩ⌅ᱟѪҶ༴⨶ᶕ㠚䇝㓳ؑ᚟ѝⲴ䰤᧕Ⲵᡆ
ᔦ䘏Ⲵ৽侸Ǆᵜㄐࡽ䶒ᨀ৺ⲴлỻᆖҐ〻ᒿᱟ໎ᕪᆖҐⲴањㆰঅⲴֻᆀǄ
⇿ㄐⲴ㔃ቮवਜ਼Ҷᡰ㾶ⴆⲴѫ㾱ᾲᘥⲴሿ㔃ǃ䘋а↕䰵䈫Ⲵ৲㘳઼Ґ仈ǄަԆሩㄐ㢲Ⲵ
ᴤᯠˈवᤜᮠᦞ䳶઼㇇⌅Ⲵᇎ⧠ˈ䜭ਟӾ㖁൰ http://www.cs.cmu.edu/~tom/mlbook.html 䇯䰞
ࡠǄ
1.5 ቅ㔉ૂ㺛‫ݻ‬䈱⢟
ᵪಘᆖҐ㠤࣋Ҿ⹄ウᔪ・㜭ཏṩᦞ㓿傼㠚ᡁᨀ儈༴⨶ᙗ㜭Ⲵ䇑㇇ᵪ〻ᒿǄᵜㄐⲴ㾱⛩व
ᤜ˖
x ᵪಘᆖҐ㇇⌅൘ᖸཊᓄ⭘亶ฏ㻛䇱᰾ᴹᖸབྷⲴᇎ⭘ԧ٬ǄᆳԜ൘ԕлᯩ䶒⢩
࡛ᴹ⭘˖˄a˅ᮠᦞ᥆ᧈ䰞仈ˈণӾབྷ䟿ᮠᦞѝਁ⧠ਟ㜭वਜ਼൘ަѝⲴᴹԧ
٬Ⲵ㿴ᖻ˄ֻྲˈӾᛓ㘵ᮠᦞᓃѝ࠶᷀⋫⯇Ⲵ㔃᷌ˈᡆ㘵Ӿ䍒࣑ᮠᦞѝᗇࡠ
ؑ⭘䍧ⅮⲴᲞ䙽㿴ࡉ˅˗˄b˅൘ḀӋഠ䳮Ⲵ亶ฏѝˈӪԜਟ㜭䘈нާᴹᔰ
ਁࠪ儈᭸Ⲵ㇇⌅ᡰ䴰Ⲵ⸕䇶˄∄ྲˈӾമ‫ۿ‬ᓃѝ䇶࡛ࠪӪ㝨˅˗˄c˅䇑㇇
ᵪ〻ᒿᗵ享ࣘᘱൠ䘲ᓄਈॆⲴ亶ฏ˄ֻྲˈ൘৏ᯉ‫׋‬㔉ਈॆⲴ⧟ຳл䘋㹼⭏
ӗ䗷〻᧗ࡦˈᡆ䘲ᓄњӪ䰵䈫‫ޤ‬䏓Ⲵਈॆ˅Ǆ
x ᵪಘᆖҐӾн਼Ⲵᆖ、੨᭦ᾲᘥˈवᤜӪᐕᲪ㜭ˈᾲ⦷઼㔏䇑ˈ䇑㇇༽ᵲᙗˈ
ؑ᚟䇪ˈᗳ⨶ᆖ઼⾎㓿⭏⢙ᆖǃ᧗ࡦ䇪ǃԕ৺ଢᆖǄ
x ањᆼᮤᇊѹⲴᆖҐ䰞仈䴰㾱ањ᰾⺞⭼ᇊⲴԫ࣑ǃᙗ㜭ᓖ䟿ḷ߶ԕ৺䇝㓳
㓿傼ⲴᶕⓀǄ
x ᵪಘᆖҐ㇇⌅Ⲵ䇮䇑䗷〻ѝवਜ਼䇨ཊ䘹ᤙˈवᤜ䘹ᤙ䇝㓳㓿傼Ⲵ㊫රǃ㾱ᆖ
ҐⲴⴞḷ࠭ᮠǃ䈕ⴞḷ࠭ᮠⲴ㺘⽪ᖒᔿǃԕ৺Ӿ䇝㓳ṧֻѝᆖҐⴞḷ࠭ᮠⲴ
㇇⌅Ǆ
x ᆖҐⲴ䗷〻ণᩌ㍒Ⲵ䗷〻ˈᩌ㍒वਜ਼ਟ㜭‫ٷ‬䇮Ⲵオ䰤ˈ֯ᗇࡠⲴ‫ٷ‬䇮ᴰㅖਸ
ᐢᴹⲴ䇝㓳ṧ઼ֻަԆ‫ݸ‬傼Ⲵ㓖ᶏᡆ⸕䇶ǄᵜҖⲴབྷ䜘࠶޵ᇩത㔅⵰ᩌ㍒਴
⿽‫ٷ‬䇮オ䰤˄ֻྲˈवਜ਼ᮠ٬࠭ᮠǃ⾎㓿㖁㔌ǃߣㆆṁǃㅖਧ㿴ࡉⲴオ䰤˅
Ⲵн਼ᆖҐᯩ⌅ˈ઼⨶䇪к䘉Ӌᩌ㍒ᯩ⌅൘ӰѸᶑԦлՊ᭦ᮋࡠᴰ֣‫ٷ‬䇮Ǆ
ᴹᖸཊ‫ޣ‬ҾᵪಘᆖҐᴰᯠ⹄ウᡀ᷌ⲴՈ⿰䍴Ⓚਟ‫׋‬䰵䈫Ǆ⴨‫Ⲵޣ‬ᵲᘇवᤜǉᵪಘᆖҐǊ
˄Machine Learning˅ˈǉ⾎㓿䇑㇇Ǌ˄Neural Computation˅ ˈǉ⾎㓿㖁㔌Ǌ˄Neural Networks˅ ˈ
ǉ㖾ഭ㔏䇑ॿՊᵏ࠺Ǌ˄Journal of the American Statistical Association˅઼ǉIEEE ⁑ᔿ䇶઼࡛
ᵪಘᲪ㜭ᆖᣕǊ˄IEEE Transactions on Pattern Analysis and Machine Intelligence˅Ǆҏᴹབྷ䟿
Ⲵᒤ Պ㾶ⴆҶᵪಘᆖҐⲴ਴њᯩ䶒ˈवᤜഭ䱵ᵪಘᆖҐՊ䇞(ICML)ˈ⾎㓿ؑ᚟༴⨶㌫㔏
(NIPS)ˈ䇑㇇ᆖҐ⨶䇪Պ䇞(CCLT)ˈഭ䱵䚇Ր㇇⌅Պ䇞(ICGA)ˈഭ䱵⸕䇶ਁ⧠઼ᮠᦞ᥆ᧈ
Պ䇞(ICKDD)ˈ⅗⍢ᵪಘᆖҐՊ䇞(ECML)ㅹǄ
Ґ仈
1.1 㔉ࠪй⿽ᵪಘᆖҐᯩ⌅䘲ਸⲴᓄ⭘ˈй⿽н䘲ਸⲴᓄ⭘Ǆ᥁䘹ᵜҖᵚᨀ৺Ⲵᓄ⭘ˈ
ᒦሩ⇿њᓄ⭘ԕаਕ䈍ᶕ䇴ԧǄ
1.2 ᥁䘹аӋᵜҖᵚᨀࡠⲴᆖҐԫ࣑Ǆ⭘㤡᮷߉а⇥䈍䶎↓ᔿൠ࣐ԕ᧿䘠Ǆ޽ቭਟ㜭㋮
⺞ൠ᧿䘠ࠪᆳⲴԫ࣑ǃᙗ㜭㺑䟿ḷ߶઼䇝㓳㓿傼Ǆᴰਾˈ㔉ࠪ㾱ᆖҐⲴⴞḷ࠭ᮠ઼ᆳⲴ㺘⽪Ǆ
䇘䇪䘉њԫ࣑䇮䇑ѝ㘳㲁Ⲵѫ㾱ᣈѝǄ
1.3 䇱᰾ᵜㄐ᧿䘠ⲴLMSᵳᴤᯠ⌅ࡉ䟷⭘Ҷởᓖл䱽ᯩ⌅֯䈟ᐞᒣᯩᴰሿॆǄ⺞࠷ൠ
䇢ˈ‫ۿ‬᮷ѝ䛓ṧᇊѹᯩᐞEǄ❦ਾ䇑㇇EሩᵳwiⲴሬᮠˈަѝ‫ٷ‬ᇊ Vˆ (b) о᮷ѝᇊѹⲴаṧˈ
wE
ᱟањ㓯ᙗ࠭ᮠǄởᓖл䱽ᱟ䙊䗷о ᡀ∄ֻൠᴤᯠ⇿њᵳ٬ᇎ⧠ⲴǄᡰԕˈᗵ享䇱᰾
wwi
ሩҾᡰ䙷ࡠⲴ⇿ањ䇝㓳ṧֻˈLMS䇝㓳⌅ࡉ䜭ᱟ᤹䘉њ∄ֻᶕ᭩ਈᵳ٬Ǆ
1.4 മ 1-1 ѝᇎ傼⭏ᡀಘ⁑ඇਟ䟷⭘ަԆаӋㆆ⮕Ǆ⺞࠷ൠ䇢ˈ㘳㲁ᇎ傼⭏ᡀಘ⭘л䶒

Ⲵㆆ⮕ᨀࠪᯠⲴỻተ˖
x ӗ⭏䲿ᵪⲴਸ⌅Ⲵỻተ
x Ӿࡽ䶒Ⲵሩᔸѝ᥁䘹ањỻተˈ❦ਾ䎠а↕к⅑⋑ᴹ䎠Ⲵỻ㘼ӗ⭏ᯠⲴỻተ
x а⿽֐㠚ᐡ䇮䇑Ⲵㆆ⮕
䇘䇪䘉Ӌㆆ⮕ⲴՈ࣓Ǆྲ᷌䇝㓳ṧֻⲴᮠ䟿ᱟപᇊⲴˈଚањ᭸᷌ᴰྭ˛‫ٷ‬ᇊᙗ㜭㺑䟿
ḷ߶ᱟ൘ц⭼䭖ḷ䎋к䎒ỻᴰཊǄ
1.5 ֯⭘㊫լҾ㾯⌻䐣ỻ䰞仈Ⲵ㇇⌅ˈᇎ⧠ањᴤㆰঅⲴtic-tac-toe⑨ᠿķǄᢺᆖҐࡠⲴ
࠭ᮠ Vˆ 㺘⽪Ѫ㠚䘹Ⲵỻተ৲ᮠⲴ㓯ᙗ㓴ਸǄ㾱䇝㓳䘉њ〻ᒿˈਟԕ䇙ᆳ઼ᆳⲴਖањᤧ䍍
৽༽∄䎋ˈਾ㘵֯⭘᡻ᐕᔪ・Ⲵപᇊ䇴ՠ࠭ᮠǄ⭘മ㺘㔈ࠪ֐Ⲵ〻ᒿ㜌࡙ⲴⲮ࠶∄ˈሩᓄҾ
䇝㓳⅑ᮠǄ
৲㘳᮷⥞
ķ
䈁⌘˖䈕⑨ᠿỻⴈѪ 3X3 ᯩṬˈৼᯩӔӂ㩭ᆀˈ俆‫ݸ‬ᇎ⧠㠚ᯩйᆀ䘎а㓯㘵㜌Ǆ
ㅢ2ㄖ ᾸᘫᆜҖૂж㡢ࡦ⢯⇀ᓅ
从特殊的训练样例中归纳出一般函数是机器学习的中心问题。本章介绍概念学
习：给定某一类别的若干正例和反例，从中获得该类别的一般定义。概念学习也可
被看作一个搜索问题，它在预定义的假设空间中搜索假设，使其与训练样例有最佳
的拟合度。多数情形下，为了高效的搜索，可以利用假设空间中一种自然形成的结
构——即一般到特殊偏序结构。本章展示了几种概念学习算法，并讨论了这些算法
能收敛得到正确假设的条件。这里还分析了归纳学习的本质，以及任意程序能从训
练数据中泛化的理由。
2.1 ԁ㔃
䇨ཊᵪಘᆖҐ䰞仈⎹৺ࡠӾ⢩↺䇝㓳ṧֻѝᗇࡠа㡜ᾲᘥǄ∄ྲӪԜнᯝᆖҐⲴаӋа
㡜ᾲᘥ઼㊫࡛वᤜ˖呏㊫ǃ⊭䖖ǃऔཻⲴᆖҐㅹǄ⇿њᾲᘥਟ㻛ⴻ֌ањሩ䊑ᡆһԦ䳶ਸˈ
ᆳᱟӾᴤབྷⲴ䳶ਸѝ䘹ਆⲴᆀ䳶˄ྲӾࣘ⢙Ⲵ䳶ਸѝ䘹ਆ呏㊫˅ˈᡆ㘵ᱟ൘䘉њ䖳བྷ䳶ਸѝ
ᇊѹⲴᐳቄ࠭ᮠ˄ྲ൘ࣘ⢙䳶ਸѝᇊѹⲴ࠭ᮠˈᆳሩ呏㊫ӗ⭏ true ᒦሩަԆࣘ⢙ӗ⭏ false˅Ǆ
ᵜㄐ㘳㲁Ⲵ䰞仈ᱟˈ㔉ᇊаṧֻ䳶ਸԕ৺⇿њṧֻᱟ੖኎ҾḀаᾲᘥⲴḷ⌘ˈᘾṧ㠚ࣘ
ᾲᘥᆖҐ˄concept learning˅ˈᡆ〠Ӿṧֻѝ䙬
᧘ᯝࠪ䈕ᾲᘥⲴа㡜ᇊѹǄ䘉а䰞仈㻛〠Ѫᾲ
䘁ᐳቄ٬࠭ᮠǄ
定义： ᾲᘥᆖҐᱟᤷӾᴹ‫ޣ‬Ḁњᐳቄ࠭ᮠⲴ䗃‫ޕ‬䗃ࠪ䇝㓳ṧֻѝˈ᧘ᯝࠪ䈕ᐳቄ࠭
ᮠǄ
2.2 жѠᾸᘫᆜҖԱࣗ
ѪҶ㢟ྭൠ⨶䀓ᾲᘥᆖҐˈ㘳㲁ањᾲᘥᆖҐⲴֻᆀˈⴞḷᾲᘥᱟ˖ĀAldo 䘋㹼≤к䘀
ࣘⲴᰕᆀāǄ㺘 2-1 ᧿䘠Ҷа㌫ࡇᰕᆀⲴṧֻˈ⇿њṧֻ㺘⽪Ѫ኎ᙗⲴ䳶ਸǄ኎ᙗ EnjoySport
㺘⽪䘉аཙ Aldo ᱟ੖ҀҾ䘋㹼≤к䘀ࣘǄ䘉њԫ࣑ⲴⴞⲴᱟˈสҾḀཙⲴ਴኎ᙗˈԕ亴⍻
ࠪ䈕ཙ EnjoySport Ⲵ٬Ǆ
㺘 2-1 ⴞḷᾲᘥ EnjoySport Ⲵ↓઼ֻ৽ֻ
Example Sky AirTemp Humidity Wind Water Forecast EnjoySport

1 Sunny Warm Normal Strong Warm Same Yes
2 Sunny Warm High Strong Warm Same Yes
3 Rainy Cold High Strong Warm Change No
4 Sunny Warm High Strong Cool Change Yes
൘䘉⿽ᛵߥлˈ䟷ਆӰѸṧⲴᖒᔿᶕ㺘⽪‫ٷ‬䇮઒˛ਟԕ‫ݸ‬㘳㲁ањ䖳ѪㆰঅⲴᖒᔿˈণ
ᇎֻⲴ਴኎ᙗ㓖ᶏⲴਸਆᔿǄ൘䘉䟼ˈਟԔ⇿њ‫ٷ‬䇮Ѫ 6 њ㓖ᶏⲴੁ䟿ˈ䘉Ӌ㓖ᶏᤷᇊҶ኎
ᙗ SkyǃAirTempǃHumidityǃWindǃWater ઼ Forecast Ⲵ٬Ǆ⇿њ኎ᙗਟਆ٬Ѫ˖
z ⭡Ā?ā㺘⽪ԫ᜿٬
z ᰾⺞ᤷᇊⲴ኎ᙗ٬˄ྲ AirTemp=Warm˅
z ⭡Āā㺘⽪н᧕ਇԫօ٬
ྲ᷌ḀӋᇎֻ x ┑䏣‫ٷ‬䇮 h Ⲵᡰᴹ㓖ᶏˈ䛓Ѹ h ሶ x ࠶㊫Ѫ↓ֻˈ˄h(x)=1 ˅Ǆ∄ྲˈ

Ѫࡔᇊ Aldo ਚ൘ሂߧ઼▞⒯Ⲵᰕᆀ䟼䘋㹼≤к䘀ࣘ˄ᒦоަԆ኎ᙗᰐ‫˅ޣ‬ ˈ䘉ṧⲴ‫ٷ‬䇮ਟ㺘
⽪Ѫл䶒Ⲵ㺘䗮ᔿ˖
<?, Cold, High, ?, ?, ?>
ᴰа㡜Ⲵ‫ٷ‬䇮ᱟ⇿аཙ䜭ᱟ↓ֻˈਟ㺘⽪Ѫ˖
<?, ?, ?, ?, ?, ?>
㘼ᴰ⢩↺Ⲵ‫ٷ‬䇮ণ⇿аཙ䜭ᱟ৽ֻˈ㺘⽪Ѫ˖
<, , , , , >
㔬кᡰ䘠ˈEnjoySport 䘉њᾲᘥᆖҐԫ࣑䴰㾱ᆖҐⲴᱟ֯ EnjoySport=Yes Ⲵᰕᆀˈᒦሶ

ަ㺘⽪Ѫ኎ᙗ㓖ᶏⲴਸਆᔿǄа㡜䈤ᶕˈԫօᾲᘥᆖҐԫ࣑㜭㻛᧿䘠Ѫ˖ᇎֻⲴ䳶ਸǃᇎֻ
䳶 ਸ к Ⲵ ⴞḷ ࠭ ᮠ ǃ ‫ى‬䘹 ‫ ٷ‬䇮 Ⲵ 䳶ਸ ԕ ৺ 䇝㓳ṧ ֻ Ⲵ 䳶 ਸǄ ԕ 䘉 ⿽ а㡜 ᖒ ᔿ ᇊ ѹⲴ
EnjoySport ᾲᘥᆖҐԫ࣑㿱㺘 2-2Ǆ
㺘 2-2 EnjoySport ᾲᘥᆖҐԫ࣑
ᐢ⸕˖
ᇎֻ䳶 X˖ਟ㜭Ⲵᰕᆀˈ⇿њᰕᆀ⭡л䶒Ⲵ኎ᙗ᧿䘠˖
Sky˄ਟਆ٬Ѫ SunnyˈCloudy ઼ Rainy˅
AirTemp˄ਟਆ٬Ѫ Warm ઼ Cold˅
Humidity˄ਟਆ٬Ѫ Normal ઼ High˅
Wind˄ਟਆ٬Ѫ Strong ઼ Weak˅
Water˄ਟਆ٬Ѫ Warm ઼ Cool˅
Forecast˄ਟਆ٬Ѫ Same ઼ Change˅
‫ٷ‬䇮䳶 H˖⇿њ‫ٷ‬䇮᧿䘠Ѫ 6 њ኎ᙗ SkyˈAirTempˈHumidityˈWindˈWater ઼ Forecast Ⲵ٬㓖
ᶏⲴਸਆǄ㓖ᶏਟԕѪĀ?ā˄㺘⽪᧕ਇԫ᜿٬˅ˈĀā˄㺘⽪ᤂ㔍ᡰᴹ٬˅ˈᡆа⢩ᇊ٬Ǆ
ⴞḷᾲᘥ c: EnjoySport: Xė{0, 1}
䇝㓳ṧֻ䳶 D˖ⴞḷ࠭ᮠⲴ↓઼ֻ৽ֻ˄㿱㺘 2-1˅
≲䀓˖
H ѝⲴа‫ٷ‬䇮 hˈ֯ሩҾ X ѝԫ᜿ xˈh(x)=c(x)Ǆ
2.2.1 ᵥ䈣ᇐѿ
൘ᵜҖѝˈᡁԜ֯⭘ԕлⲴᵟ䈝ᶕ䇘䇪ᾲᘥᆖҐ䰞仈Ǆᾲᘥᇊѹ൘ањᇎ ᇎֻ˄instance˅
䳶ਸѻкˈ䘉њ䳶ਸ㺘⽪Ѫ XǄ൘ᵜֻѝˈX ᱟᡰᴹਟ㜭Ⲵᰕᆀˈ⇿њᰕᆀ⭡ SkyǃAirTempǃ
HumidityǃWindǃWater ઼ Forecast ‫ޝ‬њ኎ᙗ㺘⽪ǄᖵᆖҐⲴᾲᘥᡆ࠭ᮠ〠Ѫⴞ
ⴞḷᾲᘥ (target
concept)ˈ䇠֌ cǄа㡜ᶕ䈤ˈc ਟԕᱟᇊѹ൘ᇎֻ X кⲴԫ᜿ᐳቄ࠭ᮠˈণ c:Xė{0, 1}Ǆ൘
䘉њֻᆀ䟼ˈⴞḷᾲᘥሩᓄҾ኎ᙗ EnjoySport Ⲵ٬ˈᖃ EnjoySport=Yes ᰦ c(x)=1ˈᖃ
EnjoySport=No ᰦ c(x)=0Ǆ
൘ᆖҐⴞḷᾲᘥᰦˈᗵ享ᨀ‫׋‬а྇䇝䇝㓳ṧֻ˄training examples˅ˈ⇿њṧֻѪ X ѝⲴа
њᇎֻ x ԕ৺ᆳⲴⴞḷᾲᘥ٬ c(x)˄ྲ㺘 2-1 ѝⲴ䇝㓳ṧֻ˅ǄሩҾ c(x)=1 Ⲵᇎֻ㻛〠Ѫ↓ ↓
ֻ(positive example)ˈᡆ〠ѪⴞḷᾲᘥⲴᡀઈǄሩҾ c(x)=0 ⲴᇎֻѪ৽ ৽ֻ(negative example)ˈ
ᡆ〠Ѫ䶎ⴞḷᾲᘥᡀઈǄ㓿ᑨਟԕ⭘ᒿ‫<ڦ‬x,c(x)>ᶕ᧿䘠䇝㓳ṧֻˈ㺘⽪ަवਜ਼Ҷᇎֻ x ઼
ⴞḷᾲᘥ٬ c(x)Ǆㅖਧ D ⭘ᶕ㺘⽪䇝㓳ṧֻⲴ䳶ਸǄ
аᰖ㔉ᇊⴞḷᾲᘥ c Ⲵ䇝㓳ṧֻ䳶ˈᆖҐಘ䶒ѤⲴ䰞仈ቡᱟ‫ٷ‬䇮ᡆՠ䇑 cǄ֯⭘ㅖਧ H

ᶕ㺘⽪ᡰ
ᡰᴹਟ㜭‫ٷ‬䇮(all possible hypotheses)Ⲵ䳶ਸˈ䘉њ䳶ਸ޵᡽ᱟѪ⺞ᇊⴞḷᾲᘥᡰ㘳㲁
Ⲵ㤳തǄ䙊ᑨ H ‫׍‬䇮䇑㘵ᡰ䘹ᤙⲴ‫ٷ‬䇮㺘⽪㘼ᇊǄH ѝ⇿њⲴ‫ٷ‬䇮 h 㺘⽪ X кᇊѹⲴᐳቄ
࠭ᮠˈণ h:Xė{0,1}ǄᵪಘᆖҐⲴⴞḷቡᱟራ᢮ањ‫ٷ‬䇮 hˈ֯ሩҾ X ѝⲴᡰᴹ xˈh(x)=c(x)Ǆ
2.2.2 ᖈ㓩ᆜҖ‫ٽ‬䇴
ᵪಘᆖҐⲴԫ࣑ᱟ൘ᮤњᇎֻ䳶ਸ X к⺞ᇊоⴞḷᾲᘥ c ⴨਼Ⲵ‫ٷ‬䇮 hˈ❦㘼ᡁԜሩҾ

c ӵᴹⲴؑ᚟ਚᱟᆳ൘䇝㓳ṧֻкⲴ٬Ǆഐ↔ˈᖂ㓣ᆖҐ㇇⌅ᴰཊਚ㜭‫؍‬䇱䗃ࠪⲴ‫ٷ‬䇮㜭о
䇝㓳ṧֻ⴨ᤏਸǄྲ᷌⋑ᴹᴤཊⲴؑ᚟ˈᡁԜਚ㜭‫ٷ‬ᇊˈሩҾᵚ㿱ᇎֻᴰྭⲴ‫ٷ‬䇮ቡᱟо䇝
㓳ᮠᦞᴰ֣ᤏਸⲴ‫ٷ‬䇮Ǆ䘉ᱟᖂ㓣ᆖҐⲴањสᵜ‫ٷ‬ᇊˈᵜҖѝሶሩ↔‫ڊ‬ᴤཊⲴ䱀䘠Ǆ䘉䟼
ᡁԜㆰঅᨀ৺ˈ൘ㅜ 5ǃ6ǃ7 ㄐሶᴤᖒᔿॆ઼ᇊ䟿ൠᇑᇊ઼࠶᷀䘉а‫ٷ‬ᇊǄ
ᖂ㓣ᆖҐ‫ٷ‬䇮 ԫа‫ٷ‬䇮ྲ᷌൘䏣ཏབྷⲴ䇝㓳ṧֻ䳶ѝᖸྭൠ䙬䘁ⴞḷ࠭ᮠˈᆳҏ㜭൘
ᵚ㿱ᇎֻѝᖸྭൠ䙬䘁ⴞḷ࠭ᮠǄ
2.3 ֒Ѱᩒ㍘ⲺᾸᘫᆜҖ
ᾲᘥᆖҐਟԕⴻ֌ᱟањᩌ㍒Ⲵ䗷〻ˈ㤳തᱟ‫ٷ‬䇮Ⲵ㺘⽪ᡰ䳀ਜ਼ᇊѹⲴᮤњオ䰤Ǆᩌ㍒
ⲴⴞḷᱟѪҶራ᢮㜭ᴰྭൠᤏਸ䇝㓳ṧֻⲴ‫ٷ‬䇮Ǆᗵ享⌘᜿ࡠˈᖃ‫ٷ‬䇮Ⲵ㺘⽪ᖒᔿ䘹ᇊਾˈ
䛓Ѹҏቡ䳀ਜ਼ൠѪᆖҐ㇇⌅⺞ᇊҶᡰᴹ‫ٷ‬䇮Ⲵオ䰤Ǆ䘉Ӌ‫ٷ‬䇮ᱟᆖҐ〻ᒿᡰ㜭㺘⽪Ⲵˈҏᱟ
ᆳ㜭ཏᆖҐⲴǄ㘳㲁൘ EnjoySport ᆖҐԫ࣑ѝⲴᇎֻ䳶ਸ X ઼‫ٷ‬䇮䳶ਸ HǄྲ᷌኎ᙗ Sky
ᴹ 3 ⿽ਟ㜭Ⲵ٬ˈ㘼 AirTempǃHumidityǃWindǃWater ઼ Forecast 䜭ਚᴹє⿽ਟ㜭٬ˈࡉ
ᇎֻオ䰤 X वਜ਼Ҷ 3×2×2×2×2×2=96 ⿽н਼ⲴᇎֻǄ㊫լⲴ䇑㇇ਟᗇˈ൘‫ٷ‬䇮オ䰤 H ѝᴹ
䈝⌅н਼(syntactically distinct)Ⲵ‫ٷ‬䇮Ǆ❦㘼ˈ⌘᜿ࡠवਜ਼ᴹㅖਧⲴ
5×4×4×4×4×4=5120 ⿽䈝
‫ٷ‬䇮ԓ㺘オᇎֻ䳶ਸˈণᆳԜሶ⇿њᇎֻ䜭࠶㊫Ѫ৽ֻǄഐ↔ˈ䈝ѹн਼(semantically distinct)
Ⲵ‫ٷ‬䇮ਚᴹ 1+4×3×3×3×3×3=973 њǄ䘉䟼Ⲵ EnjoySport ֻᆀᱟањ䶎ᑨㆰঅⲴᆖҐԫ࣑ˈ
ᆳⲴ‫ٷ‬䇮オ䰤⴨ሩ䖳ሿфᴹ䲀Ǆཊᮠᇎ䱵ⲴᆖҐԫ࣑वਜ਼ᴤབྷⲴǃᴹᰦᱟᰐ䲀Ⲵ‫ٷ‬䇮オ䰤Ǆ
ྲ᷌ᢺᆖҐⴻ֌ᱟањᩌ㍒䰞仈ˈ䛓Ѹᖸ㠚❦ˈሩᆖҐ㇇⌅Ⲵ⹄ウ䴰㾱㘳ḕ‫ٷ‬䇮オ䰤ᩌ
㍒Ⲵн਼ㆆ⮕Ǆ⢩࡛ᕅ䎧ᡁԜ‫ޤ‬䏓Ⲵ㇇⌅ᓄ㜭ᴹ᭸ൠᩌ㍒䶎ᑨབྷⲴᡆᰐ䲀Ⲵ‫ٷ‬䇮オ䰤ˈԕ᢮
ࡠᴰ֣ᤏਸ䇝㓳ᮠᦞⲴ‫ٷ‬䇮Ǆ
2.3.1 ‫ٽ‬䇴Ⲻж㡢ࡦ⢯⇀ᓅ
䇨ཊᾲᘥᆖҐ㇇⌅ѝˈᩌ㍒‫ٷ‬䇮オ䰤Ⲵᯩ⌅‫׍‬䎆Ҿަѝа⿽ᖸᴹ⭘Ⲵ㔃ᶴ˖‫ٷ‬䇮Ⲵа㡜
ࡠ⢩↺ᒿ‫ޣ‬㌫Ǆ࡙⭘‫ٷ‬䇮オ䰤Ⲵ䘉⿽㠚❦㔃ᶴˈᡁԜਟԕ൘ᰐ䲀Ⲵ‫ٷ‬䇮オ䰤ѝ䘋㹼ᖫᓅⲴᩌ
㍒ˈ㘼н䴰㾱᰾⺞ൠࡇѮᡰᴹⲴ‫ٷ‬䇮ǄѪ䈤᰾а㡜ࡠ⢩↺ᒿˈ㘳㲁ԕлєњ‫ٷ‬䇮˖
h1=<Sunny, ?, ?, Strong, ?, ?>
h2=<Sunny, ?, ?, ?, ?, ?>
ଚӋᇎֻਟ㻛h1઼h2ࡂ࠶Ѫ↓ֻ˛⭡Ҿh2वਜ਼Ⲵᇎֻ㓖ᶏ䖳ቁˈᆳࡂ࠶ࠪⲴ↓ֻҏ䖳ཊǄ
ᇎ䱵кˈԫօ㻛h1ࡂ࠶Ѫ↓ֻⲴᇎֻ䜭Պ㻛h2ࡂ࠶Ѫ↓ֻˈഐ↔ˈᡁԜ䈤h2∄h1ᴤа㡜Ǆ
ⴤ㿲кⲴĀ∄ĂĂᴤа㡜ā䘉⿽‫ޣ‬㌫ਟԕྲл㋮⺞ᇊѹǄ俆‫ˈݸ‬ሩXѝԫ᜿ᇎֻx઼Hѝ
ԫ ᜿ ‫ ٷ‬䇮 h ˈ ᡁ Ԝ 䈤 x ┑ 䏣 h ᖃ ф ӵ ᖃ h(x)=1 Ǆ ⧠ ൘ ԕ ᇎ ֻ 䳶 ਸ Ⲵ ᖒ ᔿ ᇊ ѹ а њ
more-general-than-or-equal-toⲴ‫ޣ‬㌫˖㔉ᇊ‫ٷ‬䇮hj઼hkˈhj more-general-than-or-equal-to hkˈ
ᖃфӵᖃԫ᜿ањ┑䏣hkⲴᇎ਼ֻᰦҏ┑䏣hjǄ
定义： Ԕhj ઼hk Ѫ൘XкᇊѹⲴᐳቄ࠭ᮠǄᇊѹањmore-general-than-or-equal-to‫ޣ‬

㌫ˈ䇠‫ڊ‬ıgǄ〠hjıg hkᖃфӵᖃ
(xęX)[(hk(x)=1)ė(hj(x)=1)]
ᴹ ᗵ 㾱㘳㲁 а ‫ ٷ‬䇮 ѕ Ṭ ൠ ∄ ਖ а ‫ ٷ‬䇮 ᴤ а 㡜 Ⲵ ᛵ ᖒ Ǆ ഐ ↔ ˈ ᡁ Ԝ 䈤 hj ѕ Ṭ Ⲵ
more-general-than hk˄߉֌hj˚ghk˅ˈᖃфӵᖃ(hjıghk)ġ(hkıghj)Ǆᴰਾˈ䘈ਟԕᇊѹ䘶ੁ
Ⲵ‫ޣ‬㌫Ā∄ĂĂᴤ⢩↺āѪhj more-specific-than hkˈᖃhk more-general-than hjǄ
插图——原书页码：25
Instances: ᇎֻ䳶
Hypotheses˖‫ٷ‬䇮䳶
Specific˖⢩↺
General˖а㡜
മ 2-1 ᇎֻǃ‫ٷ‬䇮઼ more-general-than ‫ޣ‬㌫

ᐖ䗩ⲴᯩṶԓ㺘ᡰᴹᇎֻⲴ䳶ਸXˈਣ䗩ⲴᯩṶԓ㺘ᡰᴹ‫ٷ‬䇮䳶ਸHǄਣ䗩Ⲵ⇿њ‫ٷ‬䇮ሩᓄᐖ䗩Xѝ
Ḁњᆀ䳶üüণ㻛↔‫ٷ‬䇮ࡂ࠶Ѫ↓ֻⲴ䳶ਸǄ䘎᧕‫ٷ‬䇮Ⲵ㇝ཤԓ㺘more-general-than‫ޣ‬㌫Ǆ㇝ཤᡰ
ᤷѪ䖳⢩↺Ⲵ‫ٷ‬䇮Ǆ⌘᜿ࡠh2ሩᓄⲴᇎֻᆀ䳶वਜ਼Ҷh1ሩᓄⲴᇎֻᆀ䳶ˈഐ↔h2 more-general-than
h 1Ǆ
Ѫ䈤᰾䘉Ӌᇊѹˈ㘳㲁EnjoySportֻᆀѝⲴh1ǃh2ǃh3ˈྲമ 2-1 ᡰ⽪Ǆ䘉йњ‫ٷ‬䇮ᱟྲ

օ⭡ıg‫ޣ‬㌫⴨‫ޣ‬㚄䎧ᶕⲴ˛ྲࡽᡰ䘠ˈh2∄h1ᴤа㡜ᱟഐѪ⇿њ┑䏣h1Ⲵᇎֻ䜭┑䏣h2Ǆ⴨
լⲴˈh2ҏ∄h3ᴤа㡜Ǆ⌘᜿h1઼h3ѻ䰤⴨ӂѻ䰤нᆈ൘ıg‫ޣ‬㌫ˈ㲭❦┑䏣䘉єњ‫ٷ‬䇮Ⲵᇎ
ֻᴹӔਐˈն⋑ᴹањ䳶ਸᆼ‫ޘ‬वਜ਼ਖањ䳶ਸǄ⌘᜿ıg઼˚g‫ޣ‬㌫Ⲵᇊѹ⤜・ҾⴞḷᾲᘥǄ
ᆳԜਚ‫׍‬䎆Ҿ┑䏣䘉єњ‫ٷ‬䇮Ⲵᇎֻˈ㘼оଚӋᇎֻ┑䏣ⴞḷᾲᘥᰐ‫ޣ‬Ǆ⭘ᖒᔿॆⲴ䈝䀰ᶕ
䈤ˈıg‫ޣ‬㌫ᇊѹҶ‫ٷ‬䇮オ䰤HкⲴањ‫ٿ‬ᒿ˄ণ䘉њ‫ޣ‬㌫ᱟ㠚৽ǃ৽ሩ〠઼Ր䙂Ⲵ˅ Ǆ‫ٿ‬ᒿ
‫ޣ‬㌫Ⲵਜ਼ѹ˄ሩᓄҾ‫ޘ‬ᒿ˅ᱟˈਟ㜭ᆈ൘h1઼h3䘉ṧⲴ‫ٷ‬䇮ሩˈ (h1ıgh3)㘼ф (h3ıgh1)Ǆ
ıg‫ޣ‬㌫ᖸ䟽㾱ˈഐѪᆳ൘‫ٷ‬䇮オ䰤Hкሩԫ
ԫ᜿ᾲᘥᆖҐ䰞仈ᨀ‫׋‬Ҷа⿽ᴹ⭘Ⲵ㔃ᶴǄ
ਾ䶒Ⲵㄐ㢲ሶ䱀䘠ᾲᘥᆖҐ㇇⌅ྲօ࡙⭘䘉а‫ٿ‬ᒿ㔃ᶴˈԕᴹ᭸ൠᩌ㍒‫ٷ‬䇮オ䰤Ǆ
2.4 Find-Sφሱᢴᶷཝ⢯⇀‫ٽ‬䇴
ྲօ֯⭘ more-general-than ‫ٿ‬ᒿᶕᩌ㍒о䇝㓳ṧֻ⴨а㠤Ⲵ‫ٷ‬䇮˛а⿽࣎⌅ᱟӾ H ѝ

ᴰ⢩↺‫ٷ‬䇮ᔰ࿻ˈ❦ਾ൘䈕‫ٷ‬䇮㾶ⴆ↓ֻཡ䍕ᰦሶަа㡜ॆ˄ᖃа‫ٷ‬䇮㜭↓⺞ൠࡂ࠶ањ↓
ֻᰦˈ〠䈕‫ٷ‬䇮Ā㾶ⴆā䈕↓ֻ˅Ǆ֯⭘‫ٿ‬ᒿᇎ⧠Ⲵ Find-S ㇇⌅Ⲵ㋮⺞᧿䘠㿱㺘 2-3Ǆ
㺘 2-3 Find-S ㇇⌅
1. ሶ h ࡍ࿻ॆѪ H ѝᴰ⢩↺‫ٷ‬䇮
2. ሩ⇿њ↓ֻ x
ሩhⲴ⇿њ኎ᙗ㓖ᶏai
ྲ᷌ x┑䏣ai
䛓Ѹ н‫ڊ‬ԫօһ
੖ࡉ ሶhѝaiᴯᦒѪx┑䏣Ⲵ㍗䛫Ⲵᴤа㡜㓖ᶏ
3. 䗃ࠪ‫ٷ‬䇮 h
Ѫ䈤᰾䘉а㇇⌅ˈ‫ٷ‬ᇊ㔉ҸᆖҐಘⲴа㌫ࡇ䇝㓳ṧֻྲ㺘 2-1 ᡰ⽪ǄFind-S Ⲵㅜа↕ᱟ

ሶ h ࡍ࿻ॆѪ H ѝᴰ⢩↺‫ٷ‬䇮˖
hĕ<, , , , , >
൘ᢛ᧿ࡠ㺘 2-1 ѝㅜањ䇝㓳ṧֻᰦˈᆳࡊྭᱟњ↓ֻǄᖸ␵ᾊˈ䘉ᰦⲴ h ཚ⢩↺ҶǄ

h ѝⲴ⇿ањ㓖ᶏ䜭н㻛䈕ṧֻ┑䏣ˈഐ↔ˈ⇿њ኎ᙗ䜭㻛ᴯᦒᡀ㜭ᤏਸ䈕ֻⲴ㍗䛫Ⲵᴤ
а㡜Ⲵ٬㓖ᶏˈҏቡᱟ䘉аṧֻⲴ኎ᙗ٬ᵜ䓛˖
hĕ<Sunny, Warm, Normal, Strong, Warm, Same>
䘉њ h ӽᰗཚ⢩↺Ҷˈᆳᢺ䲔ҶㅜањṧֻԕཆⲴᡰᴹᇎֻ䜭ࡂ࠶Ѫ৽ֻǄла↕ˈㅜ
2 њ䇝㓳ṧֻ˄ӽ❦Ѫ↓ֻ˅䘛֯䈕㇇⌅䘋а↕ሶ h ⌋ॆǄ䘉⅑֯⭘Ā?āԓᴯ h ѝн㜭┑
䏣ᯠṧֻⲴ኎ᙗ٬ǄѻਾⲴ‫ٷ‬䇮ਈѪ˖
hĕ<Sunny, Warm, ?, Strong, Warm, Same>
❦ਾ༴⨶ㅜйњ䇝㓳ṧֻˈ䘉䟼ᱟањ৽ֻˈh нਈǄᇎ䱵кˈFind-S ㇇⌅ㆰঅൠᘭ ᘭ⮕

⇿ањ৽ֻʽ䘉аᔰ࿻լѾᴹ⛩ཷᙚǄ⌘᜿䘉ᰦ‫ٷ‬䇮 h ӽ❦оᯠⲴ৽ֻа㠤˄ণ h 㜭ሶ↔ֻ
↓⺞ൠࡂ࠶Ѫ৽ֻ˅ˈഐ↔н䴰㾱ሩ h ֌ԫօᴤ᭩Ǆа㡜ᛵߥлˈਚ㾱ᡁԜ‫ٷ‬ᇊ‫ٷ‬䇮オ䰤 H
⺞ᇎवਜ਼ⵏ↓Ⲵⴞḷᾲᘥ cˈ㘼ф䇝㓳ṧֻнवਜ਼䭉䈟ˈ䛓ѸᖃࡽⲴ‫ٷ‬䇮 h н䴰㾱ഐ৽ֻࠪ
⧠㘼ᴤ᭩Ǆ৏ഐ൘Ҿᖃࡽ‫ٷ‬䇮 h ᱟ H ѝоᡰ㿲ሏࡠⲴ↓ֻ⴨а㠤Ⲵᴰ⢩↺Ⲵ‫ٷ‬䇮ˈ⭡Ҿ‫ٷ‬
ᇊⴞḷᾲᘥ c ൘ H ѝˈ㘼фᆳаᇊᱟоᡰᴹ↓ֻа㠤Ⲵˈ䛓Ѹ c аᇊ∄ h ᴤа㡜Ǆ㘼ⴞḷ
ᾲᘥ c нՊ㾶ⴆањ৽ֻˈഐ↔ h ҏнՊ˄⭡ more-general-than Ⲵᇊѹ˅
Ǆഐ↔ˈሩ৽ֻˈh
н䴰㾱֌ࠪԫօ‫؞‬᭩Ǆ
᧕⵰ᆼᡀ Find-S ㇇⌅ˈㅜഋњ↓ֻ֯ᗇ h ᴤа㡜˖
hĕ<Sunny, Warm, ?, Strong, ?, ?>
Find-S ㇇⌅╄⽪Ҷа⿽࡙⭘ more-general-than ‫ٿ‬ᒿᶕᩌ㍒‫ٷ‬䇮オ䰤Ⲵᯩ⌅Ǆ䘉аᩌ㍒⋯

⵰‫ٿ‬ᒿ䬮ˈӾ䖳⢩↺Ⲵ‫ٷ‬䇮䙀⑀䖜〫ࡠ䖳а㡜Ⲵ‫ٷ‬䇮Ǆമ 2-2 䈤᰾Ҷ൘ᇎ઼ֻ‫ٷ‬䇮オ䰤ѝⲴ
䘉⿽ᩌ㍒䗷〻Ǆ൘⇿а↕ˈ‫ٷ‬䇮ਚ൘䴰㾱㾶ⴆᯠⲴ↓ֻᰦ㻛⌋ॆǄഐ↔ˈ⇿а↕ᗇࡠⲴ‫ٷ‬䇮ˈ
䜭ᱟ൘䛓а⛩ко䇝㓳ṧֻа㠤Ⲵᴰ⢩↺Ⲵ‫ٷ‬䇮Ǆ䘉ҏᱟަ਽ᆇ Find-S Ⲵ⭡ᶕǄᾲᘥᆖҐ
Ⲵᙍᜣ൘䇨ཊн਼Ⲵ㇇⌅ѝ⭘ࡠˈᆳԜ֯⭘Ҷ਼ṧⲴ more-general-than ‫ٿ‬ᒿǄа䜘࠶㇇⌅൘
ᵜㄐ䇘䇪ˈਖаӋ᭮൘ㅜ 10 ㄐǄ
Instances: ᇎֻ䳶
Hypotheses˖‫ٷ‬䇮䳶
Specific˖⢩↺
General˖а㡜
മ 2-2 Find-S ѝⲴ‫ٷ‬䇮オ䰤ᩌ㍒

ᩌ㍒ᔰ࿻ҾHѝᴰ⢩↺Ⲵ‫ٷ‬䇮h0ˈ❦ਾṩᦞ䇝㓳ṧֻ䙀⑀а㡜ॆ˄h1ࡠh4˅Ǆ൘ᇎֻオ䰤മѝˈ↓
ֻ㻛ḷԕĀ+āˈ৽ֻḷԕĀ-āˈ㘼⋑ᴹवਜ਼൘䇝㓳ṧֻѝⲴᇎֻࡉԕᇎᗳശ⛩㺘⽪Ǆ
Find-S ㇇⌅Ⲵ‫ޣ‬䭞⢩⛩൘Ҿ˖ሩԕ኎ᙗ㓖ᶏⲴਸਆᔿ᧿䘠Ⲵ‫ٷ‬䇮オ䰤˄ྲ EnjoySport

ѝⲴ H˅ˈFind-S ‫؍‬䇱䗃ࠪѪ H ѝо↓ֻа㠤Ⲵᴰ⢩↺Ⲵ‫ٷ‬䇮Ǆਚ㾱↓⺞Ⲵⴞḷᾲᘥवਜ਼൘
H ѝˈᒦф䇝㓳ᮠᦞ䜭ᱟ↓⺞Ⲵˈᴰ㓸Ⲵ‫ٷ‬䇮ҏоᡰᴹ৽ֻа㠤Ǆ❦㘼ˈ䘉аᆖҐ㇇⌅ӽᆈ
൘аӋᵚ䀓ߣⲴ䰞仈˖
z ᆖҐ䗷〻ᱟ੖᭦ᮋࡠҶ↓⺞Ⲵⴞḷᾲᘥ˛㲭❦ Find-S ᢮ࡠҶо䇝㓳ᮠᦞа㠤Ⲵ
‫ٷ‬䇮ˈն⋑࣎⌅⺞ᇊᆳᱟ੖᢮ࡠҶᜏаਸ䘲Ⲵ‫ٷ‬䇮˄ণⴞḷᾲᘥᵜ䓛˅ˈᡆᱟ
੖䘈ᴹަԆਟ㜭Ⲵ‫ٷ‬䇮ǄᡁԜᐼᵋ㇇⌅⸕䚃ᆳ㜭੖᭦ᮋࡠⴞḷᾲᘥˈྲ᷌н㜭ˈ
㠣ቁ㾱᧿䘠ࠪ䘉⿽н⺞ᇊᙗǄ
z ѪӰѸ㾱⭘ᴰ⢩↺Ⲵ‫ٷ‬䇮Ǆྲ᷌ᴹཊњо䇝㓳ṧֻа㠤Ⲵ‫ٷ‬䇮ˈFind-S ਚ㜭᢮
ࡠᴰ⢩↺ⲴǄѪӰѸᡁԜ‫ྭٿ‬ᴰ⢩↺Ⲵ‫ٷ‬䇮ˈ㘼н䘹ᴰа㡜‫ٷ‬䇮ˈᣁᡆа㡜〻
ᓖսҾє㘵ѻ䰤ⲴḀњ‫ٷ‬䇮Ǆ
z 䇝㓳ṧֻᱟ੖⴨ӂа㠤˛൘ཊᮠᇎ䱵ⲴᆖҐ䰞仈ѝˈ䇝㓳ᮠᦞѝᑨࠪ⧠ḀӋ䭉
䈟ᡆಚ༠ˈ䘉ṧⲴна㠤Ⲵ䇝㓳䳶ሶѕ䟽⹤ൿ Find-S ㇇⌅ˈഐѪᆳᘭ⮕Ҷᡰᴹ
৽ֻǄᡁԜᵏᵋⲴ㇇⌅㠣ቁ㜭Ự⍻ࠪ䇝㓳ᮠᦞⲴна㠤ᙗˈᒦфᴰྭ㜭ᇩ㓣䘉
ṧⲴ䭉䈟Ǆ
z ྲ᷌ᴹཊњᶱབྷ⢩↺‫ٷ‬䇮ᘾѸ࣎˛൘ EnjoySport ԫ࣑Ⲵ‫ٷ‬䇮䈝䀰 H ѝˈᙫᴹа
њᜏаⲴᴰ⢩↺‫ٷ‬䇮о䇝㓳ᮠᦞа㠤Ǆ❦㘼ˈሩަԆаӋ‫ٷ‬䇮オ䰤˄ਾ䶒ሶ䇘
䇪ࡠ˅ਟ㜭ᴹཊњᶱབྷ⢩↺‫ٷ‬䇮Ǆ䘉⿽ᛵߥлˈFind-S ᗵ享㻛ᢙኅˈԕ‫ݱ‬䇨ަ
൘䘹ᤙᘾṧ⌋ॆ‫ٷ‬䇮Ⲵ䐟ᖴкഎⓟˈԕᇩ㓣ⴞḷ‫ٷ‬䇮սҾ‫ٿ‬ᒿ㔃ᶴⲴਖа࠶᭟
кⲴਟ㜭ᙗǄᴤ䘋а↕ˈᡁԜਟԕᇊѹањнᆈ൘ᶱབྷ⢩↺‫ٷ‬䇮Ⲵ‫ٷ‬䇮オ䰤ˈ
❦㘼䘉ᱟањᴤ⨶䇪ᙗⲴ䰞仈㘼нᱟᇎ䐥䰞仈˄㿱Ґ仈 2.7˅
2.5 ਎ශグ䰪ૂُ䘿⎾䲚㇍⌋
ᵜ㢲᧿䘠ⲴᱟᾲᘥᆖҐⲴਖа⿽䙄ᖴণ‫ى‬ ‫ى‬䘹⎸䲔㇇⌅˄Candidate-Elimination˅Ǆᆳ㜭䀓
ߣ Find-S ѝⲴ㤕ᒢн䏣ѻ༴ǄFind-S 䗃ࠪⲴ‫ٷ‬䇮ਚᱟ H ѝ㜭ཏᤏਸ䇝㓳ṧֻⲴཊњ‫ٷ‬䇮ѝ
ⲴањǄ㘼൘‫ى‬䘹⎸䲔㇇⌅ѝˈ䗃ࠪⲴᱟо䇝㓳ṧֻа㠤Ⲵᡰᴹ‫ٷ‬䇮Ⲵ䳶ਸǄԔӪ᛺ཷൠᱟˈ
‫ى‬䘹⎸䲔㇇⌅൘᧿䘠䘉а䳶ਸᰦн䴰㾱᰾⺞ࡇѮަᡰᴹᡀઈǄ䘉ҏᖂ࣏Ҿ more-general-than
‫ٿ‬ᒿ㔃ᶴǄ൘䘉䟼䴰㾱㔤ᣔања㠤‫ٷ‬䇮䳶ਸⲴㆰ⌱㺘⽪ˈ❦ਾ൘䙷ࡠᯠⲴ䇝㓳ṧֻᰦ䙀↕
㋮ॆ䘉а㺘⽪Ǆ
‫ى‬䘹⎸䲔㇇⌅Ⲵᓄ⭘ᴹ˖Ӿॆᆖ䍘䉡࠶᷀˄chemical mass spectroscopy˅ѝᆖҐ㿴ࡉᙗ

(Mitchell 1979)˗઼ᆖҐ੟ਁᔿᩌ㍒Ⲵ᧗ࡦ㿴ࡉ(Mitchell et al. 1983)Ǆ❦㘼ˈ‫ى‬䘹⎸䲔㇇⌅઼
Find-S ㇇⌅Ⲵᇎ䱵ᓄ⭘䜭ਇࡠ䲀ࡦˈഐѪᆳԜ൘䇝㓳ᮠᦞਜ਼ᴹಚ༠ᰦᙗ㜭䖳ᐞǄ൘䘉䟼ӻ㓽
‫ى‬䘹⎸䲔㇇⌅ⲴⴞⲴˈᱟѪҶส⹰ⲴᵪಘᆖҐ⨶䇪ᨀ‫׋‬ањ㢟ྭⲴᾲᘥṶᷦǄᵜㄐަ։䜘࠶
ሶኅ⽪䘉а㇇⌅৺⴨‫Ⲵޣ‬䰞仈ǄӾлаㄐᔰ࿻ሶ㘳ሏ䶒ሩᴹಚ༠ᮠᦞᰦᴤᑨ⭘ⲴᆖҐ㇇⌅Ǆ
2.5.1 㺞⽰
‫ى‬䘹⎸䲔㇇⌅ራ᢮ᡰᴹо䇝㓳ṧֻа㠤Ⲵ‫ٷ‬䇮ǄѪ㋮⺞᧿䘠䘉а㇇⌅ˈ䘉䟼‫ݸ‬ᕅ‫ޕ‬аӋ
สᵜⲴᇊѹǄ俆‫ˈݸ‬ᡁԜ〠ањ‫ٷ‬䇮ᱟо䇝㓳ṧֻа
а㠤Ⲵ(consistent)ˈᖃᆳ㜭↓⺞࠶㊫䘉Ӌ
ṧֻǄ
定义： ањ‫ٷ‬䇮 h о䇝㓳ṧֻ䳶ਸ D а㠤(consistent)ˈᖃфӵᖃሩ D ѝ⇿ањṧֻ

<x,c(x)>ˈh(x)=c(x)Ǆ
Consistent(h,D)į(<x,c(x)> ę D) h(x)=c(x)
⌘᜿䘉䟼ᇊѹⲴа а㠤оࡽ䶒ᇊѹⲴ┑
┑䏣ᴹ‫ޣ‬䭞Ⲵн਼Ǆањṧֻ x ൘ h(x)=1 ᰦ〠Ѫ┑
┑
䏣‫ٷ‬䇮 hˈн䇪 x ᱟⴞḷᾲᘥⲴ↓ֻ䘈ᱟ৽ֻǄ❦㘼ˈ䘉аṧֻᱟ੖о h а㠤оⴞḷᾲᘥᴹ
‫ˈޣ‬ণᱟ੖ h(x)=c(x)Ǆ
‫ى‬䘹⎸䲔㇇⌅㜭ཏ㺘⽪о䇝㓳ṧֻа㠤Ⲵᡰᴹ‫ٷ‬䇮Ǆ൘‫ٷ‬䇮オ䰤ѝⲴ䘉аᆀ䳶㻛〠Ѫ‫ޣ‬
ਈරオ䰤(version space)ˈഐѪᆳवਜ਼ⲴᱟⴞḷᾲᘥⲴᡰᴹਸ
Ҿ‫ٷ‬䇮オ䰤 H ઼䇝㓳ṧֻ D Ⲵਈ
⨶ⲴਈරǄ
ਈරオ䰤(version space)ˈḷ䇠ѪVSH,DˈᱟH
定义： ‫ޣ‬Ҿ‫ٷ‬䇮オ䰤H઼䇝㓳ṧֻ䳶DⲴਈ
ѝо䇝㓳ṧֻDа㠤Ⲵᡰᴹ‫ٷ‬䇮ᶴᡀⲴᆀ䳶Ǆ
VSH,Dį{hęH|Consistent(h,D)}
2.5.2 ࡍ㺞੄⎾䲚㇍⌋
ᱮ❦ˈ㺘⽪ਈරオ䰤Ⲵа⿽ᯩ⌅ᱟࡇࠪަᡰᴹᡀઈǄ䘉ṧਟӗ⭏ањㆰঅⲴ㇇⌅ˈ〠Ѫ
ࡇ㺘ਾ⎸䲔˄List-Then-Eliminate˅㇇⌅Ǆަᇊѹ㿱㺘 2-4Ǆ
㺘 2-4 ࡇ㺘ਾ⎸䲔㇇⌅
ࡇ㺘ਾ⎸䲔㇇⌅
1. ਈරオ䰤 VersionSpaceĕवਜ਼ H ѝᡰᴹ‫ٷ‬䇮Ⲵࡇ㺘
2. ሩ⇿њ䇝㓳ṧֻ<x, c(x)>
Ӿਈරオ䰤ѝ〫䲔ᡰᴹ h(x)Įc(x)Ⲵ‫ٷ‬䇮 h
3. 䗃ࠪ VersionSpace ѝⲴ‫ٷ‬䇮ࡇ㺘
ࡇ㺘ਾ⎸䲔㇇⌅俆‫ݸ‬ሶਈරオ䰤ࡍ࿻ॆѪवਜ਼ H ѝᡰᴹ‫ٷ‬䇮ˈ❦ਾӾѝ৫䲔оԫа䇝
㓳ṧֻна㠤Ⲵ‫ٷ‬䇮Ǆवਜ਼‫ى‬䘹‫ٷ‬䇮Ⲵਈරオ䰤䲿⵰㿲ሏࡠ䎺ᶕ䎺ཊⲴṧֻ㘼㕙߿ˈⴤࡠਚ
࢙ањ˄⨶ᜣᛵߥл˅оᡰᴹṧֻа㠤Ⲵ‫ٷ‬䇮Ǆ䘉ਟ㜭ቡᱟᡰ㾱ⲴⴞḷᾲᘥǄྲ᷌⋑ᴹ‫ݵ‬䏣
Ⲵᮠᦞ֯ਈරオ䰤㕙߿ࡠਚᴹањ‫ٷ‬䇮ˈ䛓Ѹ䈕㇇⌅ሶ䗃ࠪањ䳶ਸˈ䘉њ䳶ਸѝᡰᴹⲴ‫ٷ‬
䇮о䇝㓳ṧֻ䜭а㠤Ǆ
৏ࡉкˈਚ㾱‫ٷ‬䇮オ䰤ᱟᴹ䲀Ⲵˈቡਟ֯⭘ࡇ㺘ਾ⎸䲔㇇⌅ǄᆳާᴹᖸཊՈ⛩ˈྲ㜭‫؍‬
䇱ᗇࡠᡰᴹо䇝㓳ᮠᦞа㠤Ⲵ‫ٷ‬䇮Ǆնᱟˈ䘉а㇇⌅䶎ᑨ✖⩀ൠࡇࠪҶ H ѝᡰᴹ‫ٷ‬䇮ˈ䘉
ሩҾབྷཊᮠᇎ䱵Ⲵ‫ٷ‬䇮オ䰤ᱟн⧠ᇎⲴ㾱≲Ǆ
2.5.3 ਎ශグ䰪Ⲻᴪㆶ᱄㺞⽰
‫ى‬䘹⎸䲔㇇⌅ок䶒Ⲵࡇ㺘ਾ⎸䲔㇇⌅䚥ᗚ਼ṧⲴ৏ࡉǄ❦㘼ˈᆳ֯⭘а⿽ᴤㆰ᰾Ⲵਈ
රオ䰤Ⲵ㺘⽪⌅Ǆ൘↔ˈਈරオ䰤㻛㺘⽪ѪᆳⲴᴰа㡜Ⲵ઼ᴰ⢩↺ⲴᡀઈǄ䘉ӋᡀઈᖒᡀҶ
а㡜઼⢩↺䗩⭼Ⲵ䳶ਸˈ䘉Ӌ䗩⭼൘ᮤњ‫ٿ‬ᒿ㔃ᶴѝࡂ࠶ࠪਈරオ䰤Ǆ
മ 2-3 ਈරオ䰤৺ަа㡜઼⢩↺䗩⭼䳶ਸ
ਈරオ䰤ѝवਜ਼ҶᡰᴹⲴ 6 њ‫ٷ‬䇮ˈնਟԕㆰঅൠ⭘ S ઼ G ᶕ㺘⽪Ǆ㇝ཤ㺘⽪ᇎֻ䰤Ⲵ
more-general-than ‫ޣ‬㌫Ǆ䘉њਈරオ䰤ሩᓄҾ㺘 2-1 ѝ᧿䘠Ⲵ EnjoySport ᾲᘥᆖҐ䰞仈৺ަ䇝㓳ṧ
ֻǄ
Ѫ䈤᰾ਈරオ䰤Ⲵ䘉⿽㺘⽪ˈ޽а⅑㘳㲁㺘 2-2 ѝ᧿䘠Ⲵ EnjoySport ᾲᘥᆖҐ䰞仈Ǆሩ

Ҿ㺘 2-1 ѝ㔉ᇊⲴ 4 њ䇝㓳ṧֻˈFind-S 䗃ࠪ‫ٷ‬䇮˖
h˙<Sunny, Warm, ?, Strong, ?, ?>
ᇎ䱵кˈ䘉ਚᱟ H ѝо䇝㓳ṧֻа㠤Ⲵᡰᴹ 6 њ‫ٷ‬䇮ѻаǄᡰᴹ 6 њ‫ٷ‬䇮൘മ 2-3 ѝ

㺘⽪ࠪǄᆳԜᶴᡀҶо䈕ᮠᦞ䳶ਸ઼‫ٷ‬䇮㺘⽪⴨ሩᓄⲴਈරオ䰤Ǆ6 њ‫ٷ‬䇮ѻ䰤Ⲵ㇝ཤ㺘⽪
ᇎֻ䰤Ⲵ more-general-than ‫ޣ‬㌫Ǆ‫ى‬䘹⎸䲔㇇⌅䙊䗷֯⭘ᴰа㡜ᡀઈ˄൘മ 2-3 ѝḷѪ G˅
઼ᴰ⢩↺ᡀઈ˄മѝḷѪ S˅ᶕ㺘⽪ਈරオ䰤Ǆਚ㔉ᇊ䘉єњ䳶ਸ S ઼ GˈቡਟԕࡇѮࠪਈ
රオ䰤ѝⲴᡰᴹᡀઈˈᯩ⌅ᱟ֯⭘а㡜ࡠ⢩↺‫ٿ‬ᒿ㔃ᶴᶕ⭏ᡀ S ઼ G 䳶ਸѻ䰤Ⲵᡰᴹ‫ٷ‬䇮Ǆ
ਟԕⴤ㿲ൠⴻࠪˈ֯⭘ᴰа㡜઼ᴰ⢩↺䳶ਸ㺘⽪ਈරオ䰤Ⲵ֌⌅ᱟਸ⨶ⲴǄл䶒ᡁԜ㋮
⺞ൠᇊѹ S ઼ G 䘉єњ䗩⭼䳶ਸˈᒦф䇱᰾ᆳԜ⺞ᇎԓ㺘Ҷਈරオ䰤Ǆ
定义： ‫ޣ‬Ҿ‫ٷ‬䇮オ䰤 H ઼䇝㓳ᮠᦞ D Ⲵа а㡜䗩⭼˄General boundary˅Gˈᱟ൘ H ѝ

о D ⴨а㠤Ⲵᶱབྷа㡜˄maximally general˅ᡀઈⲴ䳶ਸǄ
Gį{ gęH | Consistent(g, D)ġ(g´ęH)[(g´ ˚g g) ġConsistent(g´, D)]}
定义： ‫ޣ‬Ҿ‫ٷ‬䇮オ䰤 H ઼䇝㓳ᮠᦞ D Ⲵ⢩ ⢩↺䗩⭼˄Specific boundary˅Sˈᱟ൘ H ѝ

о D ⴨а㠤Ⲵᶱབྷ⢩↺˄maximally specific˅ᡀઈⲴ䳶ਸǄ
Sį{ sęH | Consistent(s, D)ġ(s´ęH)[(s˚g s´) ġConsistent(s´, D)]}
ਚ㾱䳶ਸ G ઼ S 㻛㢟ྭൠᇊѹҶ˄㿱Ґ仈 2.7˅

ˈᆳԜቡᆼ‫ޘ‬㿴ᇊҶਈරオ䰤Ǆ䘉䟼䘈
ਟԕ䇱᰾ˈਈරオ䰤Ⲵ⺞࠷㓴ᡀᱟ˖G ѝवਜ਼Ⲵ‫ٷ‬䇮䳶ˈS ѝवਜ਼Ⲵ‫ٷ‬䇮䳶ˈԕ৺ G ઼ S
ѻ䰤‫ٿ‬ᒿ㔃ᶴᡰ㿴ᇊⲴ‫ٷ‬䇮Ǆ
ᇊ⨶ 2-1 ਈරオ䰤㺘⽪ᇊ⨶ǄԔ X Ѫаԫ᜿Ⲵᇎֻ䳶ਸˈH оѪ X кᇊѹⲴᐳቄ‫ٷ‬䇮Ⲵ䳶ਸǄ

Ԕ c: Xė{0, 1}Ѫ X кᇊѹⲴԫаⴞḷᾲᘥˈᒦԔ D Ѫԫа䇝㓳ṧֻⲴ䳶ਸ{<x, c(x)>}ǄሩᡰᴹⲴ Xˈ
HˈcˈD ԕ৺㢟ྭᇊѹⲴ S ઼ G˖
VSH,D = { hęH | (sęS) (gęG) (gıghıgs)}
䇱᰾˖Ѫ䇱᰾䈕ᇊ⨶ਚ䴰䇱᰾˖(1)⇿ањ┑䏣кᔿਣ䗩Ⲵh䜭൘VSH,Dѝˈ(2) VSH,DⲴ⇿њᡀઈ䜭
┑䏣ㅹᔿਣ䗩ǄѪ䇱᰾˄1˅ˈԔgѪGѝԫ᜿ањᡀઈˈsѪSѝԫаᡀઈˈhѪHⲴԫаᡀઈ㘼фgıgh
ıgsǄ⭡SⲴᇊѹˈsᗵ享㻛DѝᡰᴹⲴ↓ֻ┑䏣ǄഐѪhıg sˈ hҏ㻛Dѝᡰᴹ↓ֻ┑䏣Ǆ⴨լൠˈ⭡G
Ⲵᇊѹˈgᗵ享н㻛Dѝԫа৽ֻ┑䏣ˈф⭡Ҿ gıg hˈhҏн㻛Dѝᡰᴹ৽ֻ┑䏣Ǆ⭡Ҿ h㻛Dѝᡰᴹ
↓ֻ┑䏣фн㻛ަѝᡰᴹ৽ֻ┑䏣ˈഐ↔hоDа㠤ˈഐ↔hᱟVSH,DⲴᡀઈǄ䘉䇱᰾Ҷ↕僔˄1˅Ǆ˄2˅
Ⲵ䇘䇪〽ᗞᴹӋ༽ᵲˈਟԕ֯⭘৽䇱⌅ˈ‫ٷ‬ᇊVSH,DѝḀаhн┑䏣ㅹᔿਣ䗩ˈ䛓Ѹሶӗ⭏⸋⴮˄㿱Ґ
仈 2.6˅Ǆ
2.5.4 ُ䘿⎾䲚ᆜҖ㇍⌋
‫ى‬䘹⎸䲔㇇⌅䇑㇇ࠪⲴਈරオ䰤ˈवਜ਼ H ѝᡰᴹо䇝㓳ṧֻⲴ㿲ሏࡠⲴᒿࡇа㠤Ⲵ‫ٷ‬
䇮Ǆᔰ࿻ˈਈරオ䰤㻛ࡍ࿻ॆѪ H ѝᡰᴹ‫ٷ‬䇮Ⲵ䳶ਸǄণሶ G 䗩⭼䳶ਸࡍ࿻ॆѪ H ѝᴰа
㡜Ⲵ‫ٷ‬䇮˖
G0ĕ{<?, ?, ?, ?, ?, ?>}
ᒦሶ S 䗩⭼䳶ਸࡍ࿻ॆѪᴰ⢩↺‫ٷ‬䇮˖
S0ĕ{<, , , , , >}
䘉єњ䗩⭼䳶ਸवਜ਼Ҷᮤњ‫ٷ‬䇮オ䰤ǄഐѪHѝᡰᴹ‫ٷ‬䇮䜭∄S0ᴤа㡜ˈф∄G0ᴤ⢩↺Ǆ
㇇⌅൘༴⨶⇿њ䇝㓳ṧֻᰦˈS઼G䗩⭼䳶ਸ࠶࡛㻛⌋ॆ઼⢩ॆˈӾਈරオ䰤ѝ䙀↕⎸৫о
ṧֻна㠤Ⲵ‫ٷ‬䇮Ǆ൘ᡰᴹ䇝㓳ṧֻ༴⨶ᆼਾˈᗇࡠⲴਈරオ䰤ቡवਜ਼Ҷᡰᴹоṧֻа㠤Ⲵ
‫ٷ‬䇮ˈ㘼фਚवਜ਼䘉ṧⲴ‫ٷ‬䇮Ǆ䘉а㇇⌅൘㺘 2-5 ѝ᧿䘠˖
㺘 2-5 ֯⭘ਈරオ䰤Ⲵ‫ى‬䘹⎸䲔㇇⌅
⌘᜿↓઼ֻ৽ֻᱟᘾṧ਼ᰦᖡ૽ S ઼ G ⲴǄ
ሶ G 䳶ਸࡍ࿻ॆѪ H ѝᶱབྷа㡜‫ٷ‬䇮
ሶ S 䳶ਸࡍ࿻ॆѪ H ѝᶱབྷ⢩↺‫ٷ‬䇮
ሩ⇿њ䇝㓳ṧֻ dˈ䘋㹼ԕл᫽֌˖
ྲ᷌ d ᱟа↓ֻ
Ӿ G ѝ〫৫ᡰᴹо d на㠤Ⲵ‫ٷ‬䇮
ሩ S ѝ⇿њо d на㠤Ⲵ‫ٷ‬䇮 s
Ӿ S ѝ〫৫ s
ᢺ s ⲴᡰᴹⲴᶱሿ⌋ॆᔿ h ࣐‫ ࡠޕ‬S ѝˈަѝ h ┑䏣
h о d а㠤ˈ㘼ф G ⲴḀњᡀઈ∄ h ᴤа㡜
Ӿ S ѝ〫৫ᡰᴹ䘉ṧⲴ‫ٷ‬䇮˖ᆳ∄ S ѝਖа‫ٷ‬䇮ᴤа㡜
ྲ᷌ d ᱟањ৽ֻ
Ӿ S ѝ〫৫ᡰᴹо d на㠤Ⲵ‫ٷ‬䇮
ሩ G ѝ⇿њо d на㠤Ⲵ‫ٷ‬䇮 g
Ӿ G ѝ〫৫ g
ᢺ g ⲴᡰᴹⲴᶱሿ⢩ॆᔿ h ࣐‫ ࡠޕ‬G ѝˈަѝ h ┑䏣
h о d а㠤ˈ㘼ф S ⲴḀњᡀઈ∄ h ᴤ⢩↺
Ӿ G ѝ〫৫ᡰᴹ䘉ṧⲴ‫ٷ‬䇮˖ᆳ∄ G ѝਖа‫ٷ‬䇮ᴤ⢩↺
⌘᜿㇇⌅ѝⲴ᫽֌ˈवᤜሩ㔉ᇊ‫ٷ‬䇮Ⲵᶱሿ⌋ॆᔿ઼ᶱሿ⢩ॆᔿⲴ䇑㇇ˈᒦ⺞ᇊ䛓Ӌ䶎
ᶱሿ઼䶎ᶱབྷⲴ‫ٷ‬䇮ǄާփⲴᇎ⧠ᖃ❦‫׍‬䎆Ҿᇎ઼ֻ‫ٷ‬䇮Ⲵ㺘⽪ᯩᔿǄ❦㘼ˈਚ㾱䘉Ӌ᫽֌
㻛㢟ྭൠᇊѹҶˈ䈕㇇⌅ቡਟᓄ⭘Ҿԫ᜿ᾲᘥᆖҐ઼ԫ᜿‫ٷ‬䇮オ䰤Ǆ൘ԕлሶᇎ䱵╄⽪㇇⌅
Ⲵ䘀㹼↕僔ˈӾѝਟԕⴻࡠ൘ EnjoySport 䘉њֻᆀѝˈ䘉Ӌ᫽֌ᱟᘾṧᇎ⧠ⲴǄ
2.5.5 ㇍⌋Ⲻ⽰ׁ
മ 2-4 ╄⽪Ҷ‫ى‬䘹⎸䲔㇇⌅ᓄ⭘ࡠ㺘 2-1 ѝཤєњ䇝㓳ṧֻᰦⲴ䘀㹼↕僔Ǆྲкᡰ䘠ˈ

䗩⭼䳶ਸ‫ݸ‬㻛ࡍ࿻ॆѪG0઼S0ˈ࠶࡛ԓ㺘Hѝᴰа㡜઼ᴰ⢩↺Ⲵ‫ٷ‬䇮Ǆ
Training examples: 䇝㓳ṧֻ
മ 2-4 ‫ى‬䘹⎸䲔㇇⌅↕僔 1
S0઼G0ѪᴰࡍⲴ䗩⭼䳶ਸˈ࠶࡛ሩᓄᴰ⢩↺઼ᴰа㡜‫ٷ‬䇮Ǆ䇝㓳ṧֻ 1 ઼ 2 ֯ᗇS䗩⭼ਈᗇᴤа㡜ˈ
ྲFind-S㇇⌅ѝаṧǄ䘉ӋṧֻሩG䗩⭼⋑ᴹᖡ૽Ǆ
ᖃㅜањ䇝㓳ṧֻࠪ⧠ᰦ˄䘉䟼Ѫа↓ֻ˅ˈ‫ى‬䘹⎸䲔㇇⌅ỰḕS䗩⭼ˈᒦਁ⧠ᆳ䗷Ҿ⢩
↺ҶüüഐѪᆳн㜭㾶ⴆ䈕↓ֻǄ䘉а䗩⭼ቡ㻛‫؞‬᭩Ѫ㍗䛫ᴤа㡜Ⲵ‫ٷ‬䇮ˈԕ㾶ⴆᯠⲴṧֻǄ
‫؞‬᭩ਾⲴ䗩⭼൘മ 2-4 ѝᱮ⽪ѪS1ǄG䗩⭼н䴰㾱‫؞‬᭩ˈഐѪG0㜭ཏ↓⺞ൠ㾶ⴆ䈕ṧֻǄᖃ
༴⨶ㅜҼњ䇝㓳ṧֻᰦ˄ҏᱟˉ↓ֻ˅ˈ਼ṧൠˈ䴰㾱ሶS䘋а↕⌋ॆࡠS2 ˈGӽᰗнਈ
˄G2=G1=G0˅Ǆ⌘᜿ሩཤєњ↓ֻⲴ༴⨶䶎ᑨ㊫լҾFind-S㇇⌅Ǆ
൘ཤє↕Ⲵ㇇⌅ѝˈ↓ֻ֯ᗇਈරオ䰤ⲴS䗩⭼䙀⑀⌋ॆǄ㘼৽ֻᢞ╄Ⲵ䀂㢢ᚠྭ⴨৽ˈ
֯ᗇG䗩⭼䙀⑀⢩ॆǄ㘳㲁ㅜйњ䇝㓳ṧֻˈྲമ 2-5 ᡰ⽪Ǆ䘉а৽ֻᱮ⽪ˈG䗩⭼䗷Ҿа
㡜ҶǄҏቡᱟ䈤ˈGѝⲴ‫ٷ‬䇮䭉䈟ൠሶ䈕ֻࡔᇊѪ↓ֻҶǄഐ↔G䗩⭼ѝⲴ‫ٷ‬䇮ᗵ享㻛⢩ॆˈ
֯ᆳ㜭ሩᯠⲴ৽ֻ↓⺞࠶㊫Ǆྲമ 2-5 ᡰ⽪ˈ䘉䟼ᴹࠐ⿽ਟ䘹Ⲵᶱሿᴤ⢩↺Ⲵ‫ٷ‬䇮Ǆ䘉Ӌ‫ޘ‬
䜭ᡀѪᯠⲴG3䗩⭼䳶ਸⲴᡀઈǄ
മ 2-5 ‫ى‬䘹⎸䲔㇇⌅↕僔 2
ṧֻ 3 ᱟа৽ֻˈᆳᢺG2䗩⭼⢩ॆѪG3Ǆ⌘᜿൘G3ѝᴹཊњਟ䘹Ⲵᶱབྷа㡜‫ٷ‬䇮Ǆ
ᴹ 6 њ኎ᙗਟԕ⭘ᶕ֯G2⢩ॆˈѪӰѸਚᴹ 3 њ൘G3ѝ઒˛∄ྲh=<?, ?, Normal, ?, ?, ?>

ᱟG2Ⲵањᶱሿ⢩ॆᔿˈᆳ㜭ཏሶᯠⲴṧֻ↓⺞ൠࡂ࠶Ѫ৽ֻˈնᆳн൘G3ѝǄሶ䘉а‫ٷ‬
䇮ᧂ䲔൘ཆⲴ৏ഐᱟˈᆳоԕࡽ䙷ࡠⲴ↓ֻна㠤Ǆ൘㇇⌅ѝਚᱟㆰঅൠࡔᯝhᒦн∄ᖃࡽ
⢩↺䗩⭼S2ᴤа㡜Ǆᇎ䱵кਈරオ䰤ⲴS䗩⭼ᖒᡀҶԕᖰ↓ֻⲴ᪈㾱䈤᰾ˈᆳਟԕ⭘ᶕࡔᯝ
ԫօ㔉ᇊⲴ‫ٷ‬䇮ᱟ੖оԕᖰṧֻа㠤Ǆṩᦞᇊѹˈԫօ∄Sᴤа㡜Ⲵ‫ٷ‬䇮㜭ཏ㾶ⴆᡰᴹS㜭㾶
ⴆⲴṧֻˈণԕᖰⲴᡰᴹ↓ֻǄ਼ṧˈG䗩⭼䈤᰾Ҷԕᖰᡰᴹ৽ֻⲴؑ᚟Ǆԫօ∄Gᴤ⢩↺
Ⲵ‫ٷ‬䇮㜭‫؍‬䇱оᡰᴹ৽ֻ⴨а㠤Ǆ䘉ᱟഐѪṩᦞᇊѹˈԫа‫ٷ‬䇮нՊ㾶ⴆGᡰн㜭㾶ⴆⲴṧ
ֻǄ
ㅜഋњ䇝㓳ṧֻˈྲമ 2-6 ᡰ⽪ˈ֯ਈරオ䰤Ⲵ S 䗩⭼ᴤа㡜ॆǄᆳҏሬ㠤 G 䗩⭼ѝ

Ⲵањᡀઈ㻛ࡐ䲔ˈഐѪ䘉њᡀઈн㜭㾶ⴆᯠⲴ↓ֻǄᴰਾ䘉аࣘ֌ᶕ㠚Ҿ㺘 2-5 ㇇⌅ѝĀྲ
᷌ d ᱟа↓ֻāлⲴㅜа↕僔ǄѪ⨶䀓䘉а↕Ⲵ৏ഐˈ䴰㾱㘳㲁ѪӰѸна㠤Ⲵ‫ٷ‬䇮㾱Ӿ G
ѝ〫৫Ǆ⌘᜿䘉а‫ٷ‬䇮н㜭޽㻛⢩ॆˈഐѪ䘉ṧᆳሶн㜭㾶ⴆᯠⲴṧֻǄᆳҏн㜭㻛⌋ॆˈ
ഐѪ᤹➗ G Ⲵᇊѹˈԫօᴤа㡜Ⲵ‫ٷ‬䇮㠣ቁՊ㾶ⴆањ৽ֻǄ䘉ṧˈ䘉а‫ٷ‬䇮ᗵ享Ӿ G ѝ
〫৫ˈҏ⴨ᖃҾ〫৫Ҷਈරオ䰤Ⲵ‫ٿ‬ᒿ㔃ᶴѝⲴаᮤњ࠶᭟Ǆ
插图——原书页码：36 к
മ 2-6 ‫ى‬䘹⎸䲔㇇⌅↕僔 3
↓ֻ֯S䗩⭼ᴤа㡜ˈӾS3ਈѪS4ǄG3Ⲵањᡀઈҏᗵ享㻛ࡐ䲔ˈഐѪᆳн޽∄S4䗩⭼ᴤа㡜Ǆ
൘༴⨶ᆼ䘉 4 њṧֻਾˈ䗩⭼䳶ਸS4઼G4ࡂ࠶ࠪⲴਈරオ䰤वਜ਼Ҷоṧֻа㠤Ⲵᡰᴹ‫ٷ‬
䇮Ⲵ䳶ਸǄᮤњਈරオ䰤ˈवਜ਼䛓Ӌ⭡S4઼G4⭼ᇊⲴ‫ٷ‬䇮䜭൘മ 2-7 ѝ⽪ࠪǄ䘉аਈරオ䰤
н‫׍‬䎆Ҿ䇝㓳ṧᵜࠪ⧠Ⲵ⅑ᒿ˄ഐѪᴰ㓸ᆳवਜ਼Ҷо䇝㓳ṧֻ䳶а㠤Ⲵᡰᴹ‫ٷ‬䇮˅Ǆྲ᷌ᨀ
‫׋‬ᴤཊⲴ䇝㓳ᮠᦞˈS઼G䗩⭼ሶ㔗㔝অ䈳〫ࣘᒦ⴨ӂ䶐䘁ˈࡂ࠶ࠪ䎺ᶕ䎺ሿⲴਈරオ䰤ᶕǄ
插图——原书页码：36 л
മ 2-7 EnjoySport ᾲᘥᆖҐ䰞仈ѝⲴᴰ㓸Ⲵਈරオ䰤
2.6 ީӄ਎ශグ䰪ૂُ䘿⎾䲚Ⲻ䈪᱄
2.6.1 ُ䘿⎾䲚㇍⌋ᱥੜՐ᭬ᮑࡦ↙⺤Ⲻ‫ٽ‬䇴
⭡‫ى‬䘹⎸䲔㇇⌅ᗇࡠⲴਈරオ䰤㜭ཏ᭦ᮋࡠ᧿䘠ⴞḷᾲᘥⲴ‫ٷ‬䇮ⲴᶑԦᱟ˄1˅൘䇝㓳
ṧֻѝ⋑ᴹ䭉䈟˄2˅൘ H ѝ⺞ᇎवਜ਼᧿䘠ⴞḷᾲᘥⲴ↓⺞‫ٷ‬䇮Ǆᇎ䱵кˈྲ᷌䙷ࡠᯠⲴ䇝
㓳ṧֻˈਟԕⴁ⍻ਈරオ䰤ԕࡔᇊަоⵏ↓Ⲵⴞḷᾲᘥѻ䰤ᱟ੖䘈ᴹ࠶↗ˈԕ৺Ѫ㋮⺞⺞ᇊ
ⴞḷᾲᘥ䘈䴰㾱ཊቁ䇝㓳ṧֻǄᖃ S ઼ G 䗩⭼䳶ਸ᭦ᮋࡠঅњⲴਟ⺞ᇊⲴ‫ٷ‬䇮ᰦˈⴞḷᾲ
ᘥ᡽ⵏ↓㧧ᗇǄ
ྲ᷌䇝㓳ᮠᦞѝवਜ਼䭉䈟Պᘾṧ˛∄ྲˈԕкֻᆀѝㅜҼњṧֻ㻛䭉䈟ൠḷ⽪Ѫа৽
ֻǄ䘉⿽ᛵߥлˈᖸнᒨˈ㇇⌅㛟ᇊՊӾਈරオ䰤ѝࡐ䲔↓⺞ⲴⴞḷᾲᘥǄഐѪᆳՊࡐ䲔ᡰ
ᴹоṧֻна㠤Ⲵ‫ٷ‬䇮ˈᡰԕ൘䙷ࡠ䘉а䭉䈟Ⲵ৽ֻᰦˈ㇇⌅ሶӾਈරオ䰤ѝ〫৫↓⺞Ⲵⴞ
ḷᾲᘥǄᖃ❦ˈྲ᷌㔉ᇊ䏣ཏⲴ䇝㓳ᮠᦞˈᴰ㓸ˈᡁԜՊਁ⧠ S ઼ G 䗩⭼᭦ᮋᗇࡠањオ
Ⲵਈරオ䰤ˈӾ㘼ᗇ⸕䇝㓳ᮠᦞᴹ䈟ǄオⲴਈරオ䰤㺘⽪ H ѝ⋑
⋑ᴹ‫ٷ‬䇮㜭ཏоṧֻа㠤Ǆ
⴨լⲴᛵᖒՊࠪ⧠൘ਖа⿽⧟ຳѝ˖ᖃ䇝㓳ṧֻ↓⺞ˈնⴞḷᾲᘥн㜭⭡‫ٷ‬䇮㺘⽪ᯩᔿᡰ᧿
䘠˄∄ྲⴞḷᾲᘥᱟḀࠐњ኎ᙗ⢩ᖱⲴ᷀ਆˈ㘼‫ٷ‬䇮オ䰤ਚ᭟ᤱਸਆⲴᖒᔿ˅ǄԕਾᡁԜሶ
䈖㓶㘳㲁䘉Ӌਟ㜭ᙗǄⴞࡽˈᡁԜਚ㘳㲁ṧֻᮠᦞᱟ↓⺞Ⲵᒦфⴞḷᾲᘥ⺞ᇎ൘‫ٷ‬䇮オ䰤ѝǄ
2.6.2 сж↛䴶㾷ӶѾṭⲺ䇣㓹ṭׁ
ࡠ䘉䟼ᡁԜ䜭‫ٷ‬ᇊ䇝㓳ṧֻ⭡Ḁњཆ䜘Ⲵᯭᮉ㘵ᨀ‫׋‬Ǆ‫ٷ‬ᜣᆖҐಘਟԕѫᇠᇎ傼䘋〻ˈ
ла↕ᆳ㾱㠚ᐡ䘹ᤙањᇎֻˈ❦ਾӾཆ⭼˄㠚❦⭼ᡆањᯭᮉ㘵˅㧧ᗇ䈕ᇎֻⲴ↓⺞࠶㊫
㔃᷌Ǆ䘉а൪Ჟਟ࠶Ѫє⿽ᛵߥˈа⿽ᱟᆖҐಘ൘㠚❦⭼ѝ䘋㹼ᇎ傼˄ྲ䙐аᓗᯠẕ❦ਾ䇙
㠚❦⭼ߣᇊަᱟ੖⢒പ˅ˈᡆ൘ањᯭᮉ㘵ᤷሬлᆖҐ˄ᨀࠪаᓗᯠẕằⲴ䇮䇑ˈ❦ਾ䇙ᯭ
ᮉ㘵ᶕࡔᯝᆳᱟ੖⢒പ˅ ḕ䈒˄query˅ᶕԓ㺘ᆖҐಘᔪ・Ⲵ䘉њᇎֻˈ❦ਾ⭡
ǄᡁԜ䘉䟼⭘ḕ
ཆ⭼ᶕሩᆳ࠶㊫Ǆ
޽⅑㘳㲁മ 2-3 ѝᡰ⽪ⲴӾ EnjoySport Ⲵ 4 њṧֻѝᆖҐࡠⲴਈරオ䰤Ǆ䘉ᰦᆖҐಘᘾ

ṧ㜭ᨀࠪањ䖳ྭⲴḕ䈒˛а㡜ᛵߥлᘾṧ䟷ਆа⿽ྭⲴḕ䈒ㆆ⮕˛ᱮ❦ᆖҐಘᓄ䈅മ൘
ᖃࡽਈරオ䰤ѝ䘹ᤙ‫ٷ‬䇮ˈԕ䘋а↕ࡂ࠶䈕オ䰤Ǆഐ↔ˈ䴰㾱䘹ᤙⲴᇎֻ䴰┑䏣˖ᆳ㜭㻛ਈ
රオ䰤ѝаӋ‫ٷ‬䇮࠶㊫Ѫ↓ֻˈਖаӋ࠶㊫Ѫ৽ֻǄަѝањ䘉ṧⲴᇎֻᱟ˖
<Sunny, Warm, Normal, Light, Warm, Same>
⌘᜿䘉аᇎֻ┑䏣ਈරオ䰤Ⲵ 6 њ‫ٷ‬䇮ѝⲴ 3 њǄྲ᷌ᯭᮉ㘵ሶᇎֻࡂ࠶Ѫ↓ֻˈਈර

オ䰤Ⲵ S 䗩⭼ቡ䴰㾱㻛⌋ॆǄ⴨৽ˈྲ᷌ᯭᮉ㘵ࡂ࠶ަѪ৽ֻˈG 䗩⭼䴰㾱㻛⢩ॆǄᰐ䇪ଚ
⿽ᛵߥˈᵪಘሶ㜭ཏᆖࡠᴤཊⲴ⸕䇶ˈԕ⺞ᇊⴞḷᾲᘥˈᒦሶਈරオ䰤㕙ሿࡠ৏ᶕⲴаॺǄ
а㡜ᶕ䈤ˈᾲᘥᆖҐⲴᴰՈḕ䈒ㆆ⮕ˈᱟӗ⭏ᇎֻԕ┑䏣ᖃࡽਈරオ䰤ѝབྷ㠤ॺᮠⲴ‫ٷ‬
䇮Ǆྲ↔ˈਈරオ䰤Ⲵབྷሿਟԕ൘䙷ࡠ⇿њᯠṧֻᰦ߿ॺˈ↓⺞Ⲵⴞḷᾲᘥቡਟ൘ªlog2|VS|º
⅑ᇎ傼ਾᗇࡠǄ䘉ᴹ⛩䊑⧙Ā20 䰞ā⑨ᠿˈ䙊䗷䰞仈Ⲵᱟˋ੖എㆄ䙀⑀㧧ᗇ䰞仈Ⲵᴰ㓸ㆄ
Ṹˈ⧙ 20 䰞⑨ᠿⲴㆆ⮕ᱟᨀⲴ䰞仈ᴰྭ㜭ᢺ‫ى‬䘹ㆄṸ߿ॺǄ㲭❦൘മ 2-3 Ⲵਈරオ䰤ѝˈ
ᡁԜਟԕ⭏ᡀањᇎֻሶަ㋮⺞ൠ࠶ॺǄնа㡜ᛵߥлˈਟ㜭ᰐ⌅ᶴ䙐ࠪ䘉ṧⲴ㋮⺞࠶ॺⲴ
ᇎֻǄ䘉ṧˈḕ䈒Ⲵᮠⴞਟ㜭ՊཊҾªlog2|VS|ºǄ
2.6.3 ᙄṭֵ⭞уᇂ‫ޞ‬ᆜҖᾸᘫ
൘к䶒Ⲵֻᆀѝˈྲ᷌䲔Ҷ 4 њṧֻѻཆ⋑ᴹᴤཊⲴ䇝㓳ṧֻˈնᵪಘ⧠൘㾱ሩᵚ㿱䗷
Ⲵᇎֻ䘋㹼࠶㊫Ǆ㲭❦മ 2-3 Ⲵਈරオ䰤ѝӽवਜ਼ཊњ‫ٷ‬䇮ˈণⴞḷᾲᘥ䘈ᵚᆼ‫ޘ‬ᆖҐࡠˈ
ӽ❦ᴹਟ㜭ሩᯠṧֻ䘋㹼аᇊਟؑᓖⲴ࠶㊫ǄѪ⽪㤳䘉а䗷〻ˈ‫ٷ‬ᇊᵪಘ䴰㾱ሩ㺘 2-6 ѝⲴ
4 њᯠᇎֻ䘋㹼࠶㊫Ǆ
㺘 2-6 ᖵ࠶㊫Ⲵᯠᇎֻ
Instance Sky AirTemp Humidity Wind Water Forecast EnjoySport
A Sunny Warm Normal Strong Cool Change ?
B Rainy Cold Normal Light Warm Same ?
C Sunny Warm Normal Light Warm Same ?
D Sunny Cold Normal Strong Warm Same ?
⌘᜿ˈ㲭❦ᇎֻ A н൘䇝㓳ṧֻѝˈնᖃࡽਈරオ䰤ѝˈ⇿њ‫ٷ‬䇮˄㿱മ 2-3˅䜭ሶަ

࠶㊫Ѫ↓ֻǄ⭡Ҿਈරオ䰤Ⲵᡰᴹ‫ٷ‬䇮а㠤਼᜿ᇎֻ A Ѫ↓ֻˈᆖҐಘሶ A ࡂ࠶Ѫ↓ֻⲴ
ਟؑᓖˈоਚᴹঅњⲴⴞḷᾲᘥᰦаṧǄн㇑ਈරオ䰤ѝଚњ‫ٷ‬䇮ᴰ㓸ᡀѪⴞḷᾲᘥˈᆳ䜭
Պሶަࡂ࠶Ѫ↓ֻǄ䘋а↕ˈᡁԜ⸕䚃н䴰㾱ࡇѮਈරオ䰤ѝᡰᴹⲴ‫ٷ‬䇮ˈቡਟ⸕䚃⇿њ‫ٷ‬
䇮䜭Պሶަࡂ࠶Ѫ↓ֻǄ䘉аᶑԦᡀ・ᖃфӵᖃᇎֻ┑䏣 S Ⲵ⇿њᡀઈ˄ѪӰѸ˛˅ Ǆ৏ഐ
ᱟਈරオ䰤ѝⲴަԆ⇿њ‫ٷ‬䇮ˈ䜭㠣ቁ∄ S ⲴḀњᡀઈᴤа㡜Ǆ⭡ᡁԜⲴ more-general-than
Ⲵᇊѹˈྲ᷌ᯠⲴᇎֻ┑䏣 S Ⲵᡰᴹᡀઈˈᆳаᇊҏ┑䏣䘉Ӌᴤа㡜Ⲵ‫ٷ‬䇮Ǆ
⴨լൠˈᇎֻ B 㻛ਈරオ䰤ѝⲴ⇿њ‫ٷ‬䇮ࡂ࠶Ѫ৽ֻǄᡰԕ䘉њᇎֻਟ㻛᭮ᗳൠࡂ࠶
Ѫ৽ֻˈণ֯ᾲᘥᱟнᆼ‫ޘ‬ᆖҐⲴǄሩ䘉аᶑԦⲴ⍻䈅Ⲵᴹ᭸ᯩ⌅ᱟˈࡔᯝᇎֻн┑䏣 G
ѝⲴᡰᴹᡀઈ˄ѪӰѸ˛˅Ǆ
ᇎֻ C Ⲵᛵߥᴹᡰн਼Ǆਈරオ䰤ѝॺᮠⲴ‫ٷ‬䇮ࡂ࠶ަѪ↓ֻˈॺᮠࡂ࠶Ѫ৽ֻǄഐ
↔ˈᆖҐಘᰐ⌅ਟؑൠ࠶㊫䘉аṧֻˈ䲔䶎ᨀ‫׋‬ᴤཊⲴ䇝㓳ṧֻǄ⌘᜿ࡠˈᇎֻ C оࡽа
㢲ᨀࠪⲴањᴰՈḕ䈒⴨਼Ǆ䘉ᱟਟԕ亴㿱ⲴˈഐѪᴰᴹ࠶㊫↗ѹᙗⲴᇎֻҏаᇊᴰ㜭ᨀ‫׋‬
ᯠⲴ࠶㊫ؑ᚟Ǆ
ᴰਾˈᇎֻ D ൘ਈරオ䰤ѝ㻛єњ‫ٷ‬䇮࠶Ѫ↓ֻˈ㻛ަԆ 4 њ‫ٷ‬䇮࠶Ѫ৽ֻǄ䘉њֻ

ᆀⲴ࠶㊫ਟؑᓖ∄ᇎֻ A ઼ B 㾱ሿǄᣅ⾘䘹Ѯ㾱‫ੁٮ‬Ҿ৽ֻ࠶㊫ˈᡰԕᡁԜਟԕ䗃ࠪᤕᴹ
ᴰབྷ⾘ᮠⲴ࠶㊫ˈ䘈ਟ䱴ᑖањਟؑᓖ∄ֻԕ㺘᰾ᣅ⾘Ⲵ‫ੁٮ‬〻ᓖǄ൘ㅜ 6 ㄐሶ䇘䇪ࡠˈྲ
᷌‫ٷ‬ᇊ H ѝᡰᴹ‫ٷ‬䇮ᱟᴹ⴨ㅹⲴ‫ݸ‬傼ᾲ⦷ˈ䛓Ѹᣅ⾘Ⲵᯩ⌅㜭ᗇࡠᯠᇎֻⲴᴰਟ㜭࠶㊫Ǆ
䘋а↕Ⲵˈᣅ↓ֻ⾘‫ٷ‬䇮ᡰঐⲴ∄ֻਟ㿶Ѫ˖൘㔉ᇊ䇝㓳ᮠᦞᰦˈᇎֻѪ↓ֻⲴਟ㜭ᙗǄ
2.7 ᖈ㓩‫څ‬㖤
ྲкᡰ䘠ˈ൘㔉ᇊ↓⺞Ⲵ䇝㓳ṧֻᒦф‫؍‬䇱ࡍ࿻‫ٷ‬䇮オ䰤वਜ਼ⴞḷᾲᘥᰦˈ‫ى‬䘹⎸䲔㇇
⌅ਟԕ᭦ᮋࡠⴞḷᾲᘥǄྲ᷌ⴞḷᾲᘥн൘‫ٷ‬䇮オ䰤ѝᘾѸ࣎˛ᱟ੖ਟ䇮䇑аवਜ਼ᡰᴹ‫ٷ‬䇮
Ⲵオ䰤ᶕ䀓ߣ䘉аഠ䳮˛‫ٷ‬䇮オ䰤ⲴབྷሿሩҾ㇇⌅᧘ᒯࡠᵚ㿱ᇎֻⲴ㜭࣋ᴹӰѸᖡ૽˛‫ٷ‬
䇮オ䰤Ⲵབྷሿሩᡰ䴰䇝㓳ṧֻⲴᮠ䟿ᴹӰѸᖡ૽˛䘉Ӌ䜭ᱟᖂ㓣᧘⨶ѝⲴаӋสᵜ䰞仈Ǆ䘉
䟼ᡁԜ൘‫ى‬䘹⎸䲔㇇⌅ѝ㘳ሏ䘉Ӌ䰞仈Ǆ❦㘼ਟԕⴻࡠˈ䘉䟼Ⲵ࠶᷀ѝᗇࡠⲴ㔃䇪ਟԕᓄ⭘
Ҿԫ᜿ⲴᾲᘥᆖҐ㌫㔏Ǆ
2.7.1 жѠᴿ‫ٽⲺڅ‬䇴グ䰪
ྲ᷌ᜣ‫؍‬䇱‫ٷ‬䇮オ䰤वਜ਼ⴞḷᾲᘥˈањ᰾ᱮⲴᯩ⌅ᱟᢙབྷ‫ٷ‬䇮オ䰤ˈ֯⇿њਟ㜭Ⲵ‫ٷ‬
䇮䜭वਜ਼൘޵Ǆ޽а⅑֯⭘ EnjoySport 䘉њֻᆀˈަѝᡁԜሶ‫ٷ‬䇮オ䰤䲀ࡦѪਚवਜ਼኎ᙗ٬
ⲴਸਆǄ⭡Ҿ䘉а䲀ࡦˈ‫ٷ‬䇮オ䰤н㜭ཏ㺘⽪ᴰㆰঅⲴ᷀ਆᖒᔿⲴⴞḷᾲᘥˈྲĀSky=Sunny
ᡆ Sky=CloudyāǄᇎ䱵кˈྲ᷌㔉ᇊԕлйњ䇝㓳ṧֻˈᆳԜᶕ㠚Ҿ䈕᷀ਆᔿ‫ٷ‬䇮ˈᡁԜⲴ
㇇⌅ሶᗇࡠањオⲴਈරオ䰤Ǆ
Example Sky AirTemp Humidity Wind Water Forecast EnjoySport
1 Sunny Warm Normal Strong Cool Change Yes
2 Cloudy Warm Normal Strong Cool Change Yes
3 Rainy Warm Normal Strong Cool Change No
ѻᡰԕнᆈ൘о䘉 3 њṧֻа㠤Ⲵ‫ٷ‬䇮Ⲵ৏ഐᱟˈоཤєњṧֻа㠤ˈᒦф㜭൘㔉ᇊ‫ٷ‬
䇮オ䰤 H ѝ㺘⽪Ⲵᴰ⢩↺Ⲵ‫ٷ‬䇮ᱟ˖
S2: <?, Warm, Nornal, Strong, Cool, Change>
䘉а‫ٷ‬䇮㲭❦ᱟ H ѝоṧֻа㠤Ⲵᴰ⢩↺Ⲵ‫ٷ‬䇮ˈᆳӽ❦䗷Ҿа㡜ॆҶ˖ᆳሶㅜйњ
ṧֻ䭉䈟ൠࡂѪ↓ֻǄ䰞仈൘ҾˈᡁԜ֯ᆖҐಘ‫ੁٿ‬Ҿਚ㘳㲁ਸਆⲴ‫ٷ‬䇮ˈ䘉䟼䴰㾱㺘⽪㜭
࣋ᴤᕪⲴ‫ٷ‬䇮オ䰤Ǆ
2.7.2 ᰖ‫Ⲻڅ‬ᆜҖಞ
ᖸᱮ❦ˈѪҶ‫؍‬䇱ⴞḷᾲᘥ൘‫ٷ‬䇮オ䰤ѝˈ䴰㾱ᨀ‫׋‬ањ‫ٷ‬䇮オ䰤ˈᆳ㜭㺘䗮ᡰᴹⲴਟ ਟ
ᮉᦸᾲᘥ(every teachable concept)Ǆᦒ䀰ѻˈᆳ㜭ཏ㺘䗮ᇎֻ䳶 X Ⲵᡰᴹਟ㜭Ⲵᆀ䳶Ǆа㡜
ൠˈᡁԜᢺ䳶ਸ X ᡰᴹᆀ䳶Ⲵ䳶ਸ〠Ѫ X Ⲵᑲ ᑲ䳶˄power set˅Ǆ
ֻྲ൘EnjoySportᆖҐԫ࣑ѝˈ֯⭘ 6 ⿽኎ᙗ᧿䘠Ⲵᇎֻオ䰤XⲴབྷሿѪ 96Ǆ൘䘉аᇎ

ֻ䳶ਸкਟԕᇊѹཊቁᾲᘥ˛ᦒ䀰ѻˈXⲴᑲ䳶བྷሿᱟӰѸ˛а㡜䈤ᶕ൘䳶ਸXкᇊѹⲴ⴨
ᔲᆀ䳶ᮠⴞ˄ণXᑲ䳶Ⲵབྷሿ˅Ѫ 2|X|ˈަѝ|X|ᱟXⲴ‫ݳ‬㍐ᮠⴞǄഐ↔൘䘉аᇎֻオ䰤кਟᇊ
ѹ 296ˈᡆབྷ㓖ᱟ 1028њн਼Ⲵⴞḷᾲᘥˈ䘉ҏᱟᆖҐಘᡰ䴰㾱ᆖҐⲴⴞḷᾲᘥᮠⴞǄഎᗶ
2.3 㢲ѝਸਆ‫ٷ‬䇮オ䰤ਚ㜭㺘⽪ 973 њ‫ٷ‬䇮üüᇎ൘ᱟањ‫ٿ‬㖞ᖸབྷⲴ‫ٷ‬䇮オ䰤ʽ
⧠൘ሶ EnjoySport ᆖҐԫ࣑䟽ᯠᇊѹѪа⿽ᰐ‫Ⲵٿ‬ᖒᔿǄᯩ⌅ᱟᇊѹањᯠⲴ‫ٷ‬䇮オ䰤

H´ˈᆳ㜭㺘⽪ᇎֻⲴ⇿ањᆀ䳶ˈҏቡᱟᢺ H´ሩᓄࡠ X Ⲵᑲ䳶Ǆᇊѹ H´Ⲵа⿽࣎⌅ᱟˈ‫ݱ‬
䇨֯⭘ࡽ䶒Ⲵ‫ٷ‬䇮Ⲵԫ᜿᷀ਆǃਸਆ઼੖ᇊᔿǄֻྲⴞḷᾲᘥĀSky=Sunny ᡆ Sky=Cloudyā
ਟ㻛᧿䘠Ѫ˖
<Sunny, ?, ?, ?, ?, ?> Ģ <Cloudy, ?, ?, ?, ?, ?>
㔉ᇊ䘉ṧⲴ‫ٷ‬䇮オ䰤ˈᡁԜቡਟԕᆹ‫ޘ‬ൠ֯⭘‫ى‬䘹⎸䲔㇇⌅ˈ㘼нᗵᣵᗳᰐ⌅㺘䗮ⴞḷ
ᾲᘥǄ❦㘼ˈ㲭❦䘉њ‫ٷ‬䇮オ䰤ᧂ䲔Ҷ㺘䗮㜭࣋Ⲵ䰞仈ˈᆳ৸ӗ⭏ҶањᯠⲴǃ਼ṧഠ䳮Ⲵ
䰞仈˖ᾲᘥᆖҐ㇇⌅ሶᆼ‫ޘ‬ᰐ⌅Ӿ䇝㓳ṧֻѝ⌋ॆʽަ৏ഐྲлˈ‫ٷ‬ᇊᡁԜᨀ‫׋‬Ҷ 3 њ↓ֻ
˄x1ˈx2ˈx3˅ԕ৺єњ৽ֻ˄x4ˈx5˅㔉ᆖҐಘǄ䘉ᰦˈਈරオ䰤ⲴS䗩⭼वਜ਼Ⲵ‫ٷ‬䇮↓ྭ
ᱟйњ↓ֻⲴ᷀ਆ˖
S: { (x1Ģx2Ģx3) }
ഐѪ䘉ᱟ㜭㾶ⴆ 3 њ↓ֻⲴᴰ⢩↺‫ٷ‬䇮Ǆ⴨լൠˈG 䗩⭼ሶ⭡䛓Ӌࡊྭ㜭ᧂ䲔ᦹ৽ֻⲴ

䛓Ӌ‫ٷ‬䇮㓴ᡀǄ
G: { (x4Ģx5)}
䰞仈൘Ҿˈ䘉а䶎ᑨާᴹ㺘䗮࣋Ⲵ‫ٷ‬䇮㺘⽪ᯩ⌅ѝˈS 䗩⭼ᙫᱟㆰঅⲴᡰᴹ↓ֻ᷀ਆᔿˈ
G 䗩⭼ᙫᱟᡰᴹ৽ֻⲴ᷀ਆⲴ੖ᇊᔿǄ䘉ṧ㜭ཏ⭡ S ઼ G ᰐ↗ѹൠ࠶㊫Ⲵˈਚᴹᐢ㿱ࡠⲴ
䇝㓳ṧֻᵜ䓛Ǆ㾱ᜣ㧧ᗇঅњⲴⴞḷᾲᘥˈቡᗵ享ᨀ‫ ׋‬X ѝᡰᴹⲴᇎֻ֌Ѫ䇝㓳ṧֻǄ
ⴻ䎧ᶕ䚯‫ݽ‬䘉а䰞仈Ⲵᯩ⌅ਟԕ֯⭘↔䜘࠶ᆖҐⲴਈරオ䰤ˈ❦ਾྲ 2.6.3 㢲ѝ䛓ṧ⭡

ਈරオ䰤Ⲵᡰᴹᡀઈᣅ⾘ǄнᒨⲴᱟˈ㜭ཏӗ⭏а㠤ᣅ⾘Ⲵਚᴹ䛓Ӌᐢ㿱䗷Ⲵ䇝㓳ṧֻǄሩ
ަԆᡰᴹⲴᇎֻˈ䘋㹼ᣅ⾘⋑ᴹԫօ᭸᷌˖⇿ањᵚ㿱䗷Ⲵᇎֻ䜭Պ㻛ਈරオ䰤ѝࡊ ࡊྭॺᮠ
Ⲵ‫ٷ‬䇮ࡂ࠶Ѫ↓ֻˈ㘼㻛ਖаॺࡂ࠶Ѫ৽ֻ˄ѪӰѸ˛˅Ǆ৏ഐྲлˈ㤕 H ᱟ X Ⲵᑲ䳶ˈ㘼
x ᱟḀњᵚࠪ⧠䗷ⲴᇎֻˈࡉሩҾਈරオ䰤ѝа㾶ⴆ x Ⲵ‫ٷ‬䇮 hˈᗵ❦ᆈ൘ਖа‫ٷ‬䇮 h´ˈᆳ
о h ࠐѾ⴨ㅹਚн䗷ሩ x Ⲵ࠶㊫н਼Ǆ㘼фྲ᷌ h ൘ਈරオ䰤ѝˈ䛓Ѹ h´ҏ൘ˈഐѪᆳሩҾ
ᐢᖰ䇝㓳ṧֻⲴࡂ࠶о h ᆼ‫ޘ‬аṧǄ
2.7.3 ᰖ‫څ‬ᆜҖⲺᰖ⭞ᙝ
ԕкⲴ䇘䇪䈤᰾Ҷᖂ㓣᧘⨶Ⲵањสᵜ኎ᙗ˖ᆖҐಘྲ᷌нሩⴞḷᾲᘥⲴᖒᔿ‫ڊ‬亴‫Ⲵݸ‬
‫ٷ‬ᇊˈᆳӾṩᵜкᰐ⌅ሩᵚ㿱ᇎֻ䘋㹼࠶㊫Ǆᇎ䱵к൘ᡁԜ৏ᶕⲴ EnjoySport ԫ࣑ѝˈ‫ى‬䘹
⎸䲔㇇⌅㜭ཏӾ䇝㓳ṧֻѝ⌋ॆˈަᜏаⲴ৏ഐቡᱟᆳᱟᴹ‫ˈⲴٿ‬ᆳ䳀ਜ਼‫ٷ‬ᇊҶⴞḷᾲᘥਟ
ԕ⭡኎ᙗ٬Ⲵਸਆᶕ㺘⽪Ǆྲ᷌䘉а‫ٷ‬ᇊ↓⺞˄ᒦф䇝㓳ᮠᦞᰐ䭉˅ˈሩҾᯠᇎֻⲴ࠶㊫ҏ
Պᱟ↓⺞ⲴǄնྲ᷌䘉њ‫ٷ‬ᇊн↓⺞ˈ‫ى‬䘹⎸䲔㇇⌅㛟ᇊՊ䭉䈟ൠ࠶㊫ X ѝḀӋᇎֻǄ
⭡Ҿᖂ㓣ᆖҐ䴰㾱Ḁ⿽ᖒᔿⲴ亴‫ٷݸ‬ᇊˈᡆ〠Ѫᖂ㓣‫ٿ‬㖞˄Inductive bias˅*ˈᡁԜਟԕ
⭘ᖂ㓣‫ٿ‬㖞ᶕ᧿䘠н਼ᆖҐᯩ⌅Ⲵ⢩ᖱǄ⧠൘ᶕ㋮⺞ൠᇊѹᖂ㓣‫ٿ‬㖞Ǆ䘉䟼㾱㧧ਆⲴ‫ޣ‬䭞ᙍ
ᜣ൘ҾˈᆖҐಘ൘Ӿ䇝㓳ṧֻѝ⌋ॆᒦ᧘ᯝᯠᇎֻⲴ࠶㊫䗷〻ѝᡰ䟷⭘Ⲵㆆ⮕Ǆഐ↔ˈ㘳㲁
а㡜ᛵߥлԫ᜿ⲴᆖҐ㇇⌅Lˈԕ৺Ѫԫ᜿ⴞḷᾲᘥcᨀ‫Ⲵ׋‬ԫ᜿䇝㓳ᮠᦞDc={<x, c(x)>}Ǆ
䇝㓳䗷〻㔃ᶏਾˈL䴰㾱ሩᯠⲴᇎֻxi䘋㹼࠶㊫ǄԔL(xi, Dc)㺘⽪൘ሩ䇝㓳ᮠᦞDcᆖҐਾL䍻
ҸxiⲴ࠶㊫˄↓ֻᡆ৽ֻ˅
ˈᡁԜਟԕྲл᧿䘠Lᡰ䘋㹼Ⲵ䘉аᖂ㓣᧘⨶䗷〻˖
(Dcġxi) L(xi,Dc)
䘉䟼Ⲵ䇠ਧy z㺘⽪zӾyᖂ㓣᧘⨶ᗇࡠˈֻྲˈྲ᷌ԔLѪ‫ى‬䘹⎸䲔㇇⌅ˈDcѪ㺘 2-1

ѝⲴ䇝㓳ᮠᦞˈxiѪ㺘 2-6 ѝㅜањᇎֻˈࡉᖂ㓣᧘⨶ਟᗇࡠ㔃䇪L(xi, Dc)=(EnjoySport=yes)Ǆ
⭡ҾLᱟаᖂ㓣ᆖҐ㇇⌅ˈࡉа㡜ᛵߥлL(xi, Dc)䘉а᧘䇪ࠪⲴ㔃᷌↓⺞ᙗᰐ⌅䇱᰾˗ҏ
ቡᱟ䈤ˈ࠶㊫L(xi, Dc)ᒦ䶎Ӿ䇝㓳ᮠᦞDc઼ᯠᇎֻxiѝ╄㓾⍮⭏Ǆ❦㘼䰞仈ᱟˈ䴰㾱൘Dcġxi
к䱴࣐ᘾṧⲴࡽᨀˈԕ֯L(xi,Dc)㜭╄㓾⍮⭏ǄᡁԜᇊѹLⲴᖂ ᖂ㓣‫ٿ‬㖞Ѫ䘉Ӌ䱴࣐ࡽᨀⲴ䳶ਸǄ
ᴤ㋮⺞ൠ䈤ˈᡁԜᇊѹLⲴᖂ㓣‫ٿ‬㖞Ѫࡽᨀ䳶ਸBˈ֯ᡰᴹⲴᯠᇎֻxi┑䏣Ǆ
(BġDcġxi) ƅ L(xi, Dc)
䘉䟼Ⲵ䇠ਧ yƅz 㺘⽪ z Ӿ y ╄㓾⍮⭏˄follow deductivelyˈᡆ z ਟԕ⭡ y 䇱᰾ᗇࠪ˅Ǆ䘉

ṧˈᡁԜᇊѹᆖҐಘⲴᖂ㓣㕆ੁѪ䱴࣐Ⲵࡽᨀ䳶ਸ Bˈ䙊䗷 B ‫࠶ݵ‬ൠ֯ᖂ㓣᧘⨶⭡╄㓾᧘⨶
ᶕ䇪䇱Ǆԕлᱟ䈕ᇊѹⲴᙫ㔃˖
定义：㘳㲁ሩҾᇎֻ䳶ਸXⲴᾲᘥᆖҐ㇇⌅LǄԔcѪXкᇊѹⲴԫаᾲᘥˈᒦԔDc={<x,
c(x)>}ѪcⲴԫ᜿䇝㓳ṧֻ䳶ਸǄԔL(xi, Dc)㺘⽪㓿䗷ᮠᦞDcⲴ䇝㓳ਾˈL䍻Ҹᇎֻxi
Ⲵ࠶㊫ǄLⲴᖂ ᖂ㓣‫ٿ‬㖞ᱟᴰሿᯝ䀰䳶ਸBˈᆳ֯ԫ᜿ⴞḷᾲᘥc઼⴨ᓄⲴ䇝㓳ṧֻDc
┑䏣˖
(xięX)[ (BġDcġxi) ƅ L(xi, Dc)] (2.1)
䛓Ѹˈ‫ى‬䘹⎸䲔㇇⌅Ⲵᖂ㓣‫ٿ‬㖞ᱟӰѸ઒˛俆‫⺞ݸ‬ᇊ䘉а㇇⌅ⲴL(xi, Dc)˖㔉ᇊᮠᦞ䳶
Dcˈ‫ى‬䘹⎸䲔㇇⌅俆‫ݸ‬䇑㇇ਈරオ䰤VSH,Dcˈ❦ਾ൘ਈරオ䰤ᡰवਜ਼Ⲵ‫ٷ‬䇮ѝᣅ⾘ˈ䘋㹼ᯠ
ᇎֻxiⲴ࠶㊫Ǆ䘉䟼‫ٷ‬ᇊӗ⭏xiⲴ࠶㊫ⲴᶑԦᱟᣅ⾘а㠤Ѫ↓ᡆѪ䍏ˈ੖ࡉн䘋㹼࠶㊫Ǆ⧠
൘ᶕഎㆄӰѸᱟ‫ى‬䘹⎸䲔㇇⌅L(xi, Dc)Ⲵᖂ㓣‫ٿ‬㖞Ⲵ䰞仈˖ᖸㆰঅˈቡᱟcęH䘉њࡽᨀǄᴹ
Ҷ䘉аࡽᨀˈ‫ى‬䘹⎸䲔㇇⌅ᡰᢗ㹼Ⲵ⇿аᖂ㓣᧘⨶䜭ਟԕ㻛╄㓾䇪䇱Ǆ
⧠൘ⴻаⴻѪӰѸL(xi, Dc)䘉а࠶㊫ਟ⭡B={cęH}ǃᮠᦞDc઼ᇎֻxi╄㓾⍮⭏Ǆ俆‫ˈݸ‬
⌘᜿ྲ᷌‫ٷ‬ᇊcęHˈ䛓Ѹਟ╄㓾⍮⭏ࠪcęVSH,DcǄ䘉а⍮⭏ⲴᶑԦ䲔cęHˈ䘈वᤜਈරオ
䰤VSH,DcⲴᇊѹ˄ণHѝоDcа㠤Ⲵᡰᴹ‫ٷ‬䇮䳶ਸ˅ ˈԕ৺ሩDc={<x, c(x)>}Ⲵᇊѹ˄ণоⴞḷ
ᾲᘥа㠤Ⲵ䇝㓳ᮠᦞ˅Ǆަ⅑ˈ⭡ҾL(xi, Dc)ᱟа࠶㊫ˈᆳᇊѹѪਈරオ䰤ѝᡰᴹ‫ٷ‬䇮Ⲵа㠤
ᣅ⾘Ǆഐ↔ˈྲ᷌L䗃ࠪ࠶㊫L(xi, Dc)ˈ䛓ѸVSH,Dcѝ⇿а‫ٷ‬䇮ᗵሶӗ⭏਼ṧⲴ࠶㊫ˈवᤜ‫ٷ‬
䇮cęVSH,DcǄഐ↔c(xi)=L(xi, Dc)‫ى‬䘹⎸䲔㇇⌅Ⲵᖂ㓣‫ٿ‬㖞ᾲᤜ䈤᰾ྲл˖
*
䘉䟼Ⲵᵟ䈝ᖂ
ᖂ㓣‫ٿ‬㖞˄inductive bias˅н㾱઼㔏䇑ᆖѝᲞ䙽֯⭘Ⲵՠ
ՠ䇑‫ٿ‬ᐞ˄estimation bias˅␧⏶Ǆ
ՠ䇑‫ٿ‬ᐞሶ൘ㅜ 5 ㄐ䇘䇪Ǆ
‫ى‬䘹⎸䲔㇇⌅Ⲵᖂ㓣‫ٿ‬㖞˖ⴞḷᾲᘥ c वਜ਼൘㔉ᇊⲴ‫ٷ‬䇮オ䰤 H ѝǄ
Inductive system: ᖂ㓣㌫㔏
New instance: ᯠᇎֻ
Candidate Elimination Algorithm ‫ى‬䘹⎸䲔㇇⌅
Using Hypothesis Space H: ֯⭘‫ٷ‬䇮オ䰤 H
Classification of new instance, or “don't know”: ሩᯠᇎֻⲴ࠶㊫ˈᡆĀᰐ⌅࠶㊫ā
Equivalent deductive system: ㅹԧⲴ╄㓾㌫㔏
Assertion “H contains the target concept”: ᯝ䀰˖ĀH वਜ਼ⴞḷᾲᘥā
Theorem Prover:ᇊ⨶䇱᰾ಘ
Inductive bias made explicit: 㻛᰾⺞ॆⲴᖂ㓣‫ٿ‬㖞
മ 2-8 ⭘ㅹԧⲴ╄㓾㌫㔏ᶕ⁑ᤏᖂ㓣㌫㔏
֯⭘‫ٷ‬䇮オ䰤 H Ⲵ‫ى‬䘹⎸䲔㇇⌅Ⲵ䗃‫ޕ‬䗃ࠪ㹼ѪˈㅹԧҾ࡙⭘Ҷᯝ䀰ĀH वਜ਼ⴞḷᾲᘥāⲴ╄㓾
ᇊ⨶䇱᰾ಘǄ䈕ᯝ䀰ഐ↔㻛〠Ѫ‫ى‬䘹⎸䲔㇇⌅Ⲵᖂᖂ㓣‫ٿ‬㖞Ǆ⭘ᖂ㓣‫ٿ‬㖞ᶕ࡫⭫ᖂ㓣㌫㔏ˈਟԕ‫ׯ‬Ҿ
֯⭘ㅹԧⲴ╄㓾㌫㔏ᶕ⁑ᤏᆳԜǄ䘉ᨀ‫׋‬Ҷа⿽ሩᖂ㓣㌫㔏䘋㹼∄䖳Ⲵᯩ⌅ˈ䙊䗷ᆳԜӾ䇝㓳ᮠᦞ
ѝ⌋ॆⲴㆆ⮕Ǆ
മ 2-8 Ѫа⽪᜿മ䀓Ǆк䶒Ⲵമᱮ⽪Ⲵ‫ى‬䘹⎸䲔㇇⌅ᴹєњ䗃‫˖ޕ‬䇝㓳ṧ઼ֻᖵ࠶㊫Ⲵ
ᯠᇎֻǄл䶒ⲴമѪа╄㓾ᇊ⨶䇱᰾ಘˈᆳⲴ䗃‫ޕ‬वᤜ਼ṧⲴє㓴ᮠᦞˈ޽࣐кᯝ䀰ĀH व
ਜ਼ⴞḷᾲᘥā Ǆ䘉єњ㌫㔏ሩᡰᴹਟ㜭Ⲵ䇝㓳ṧֻ䗃‫઼ޕ‬ᯠᇎֻ䗃‫ޕ‬ӗ⭏਼ṧⲴ䗃ࠪǄᖃ❦ˈ
൘ᇊ⨶䇱᰾ಘѝᱮᔿ䗃‫Ⲵޕ‬ᖂ㓣‫ٿ‬㖞ਚᱟ䳀ਜ਼൘Ҷ‫ى‬䘹⎸䲔㇇⌅Ⲵԓ⸱ѝǄ൘Ḁ⿽᜿ѹкˈ
ᖂ㓣‫ٿ‬㖞ਚ൘ᡁԜⲴঠ䊑ѝᆈ൘ˈնᆳ⺞ᇎᱟ㜭㻛ᆼᮤᇊѹⲴᯝ䀰䳶ਸǄ
ሶᖂ㓣᧘⨶㌫㔏ⴻ֌ᱟवਜ਼Ҷᖂ㓣‫ٿ‬㖞ˈྭ༴൘Ҿᆳᨀ‫׋‬Ҷа⿽䶎〻ᒿॆⲴ᧿䘠᡻⇥ˈ
ԕ᧿䘠ᆖҐಘӾ㿲ሏࡠⲴᮠᦞѝ䘋㹼⌋ॆⲴㆆ⮕Ǆަ⅑ᆳ䘈ਟԕሩᖂ㓣‫ٿ‬㖞ᕪᓖн਼ⲴᆖҐ
ಘ䘋㹼∄䖳Ǆֻྲˈ㘳㲁ԕл 3 њᆖҐ㇇⌅ˈ᤹ަᴹ‫ٿ‬〻ᓖӾᕡࡠᕪ䘋㹼ᧂᒿ˖
1. ᵪỠᆖҐಘ˄Rote-Learner˅Ǆㆰঅൠሶ⇿њ㿲ሏࡠⲴ䇝㓳ṧֻᆈۘлᶕǄਾ㔝Ⲵᇎ
ֻⲴ࠶㊫䙊䗷൘޵ᆈѝ३䝽䘋㹼Ǆྲ᷌ᇎֻ൘޵ᆈѝ᢮ࡠҶˈᆈۘⲴ࠶㊫㔃᷌㻛䗃ࠪǄ
੖ࡉ㌫㔏ᤂ㔍䘋㹼࠶㊫Ǆ
2. ‫ى‬䘹⎸䲔㇇⌅ǄᯠⲴᇎֻਚ൘ਈරオ䰤ᡰᴹᡀઈ䜭䘋㹼਼ṧ࠶㊫ᰦ᡽䗃ࠪ࠶㊫㔃
᷌ˈ੖ࡉ㌫㔏ᤂ㔍࠶㊫Ǆ
3. Find-SǄྲࡽᡰ䘠ˈ䘉а㇇⌅ራ᢮о䇝㓳ṧֻа㠤Ⲵᴰ⢩↺Ⲵ‫ٷ‬䇮ˈᆳ⭘䘉а‫ٷ‬䇮
ᶕ࠶㊫ਾ㔝ᇎֻǄ
䇠ᗶᆖҐಘ⋑ᴹᖂ㓣‫ٿ‬㖞ǄሩҾᯠᇎֻᡰ‫࠶Ⲵڊ‬㊫㜭Ӿᐢ㿲ሏࡠⲴ䇝㓳ṧֻѝ╄㓾⍮
⭏ˈн䴰㾱䱴࣐ⲴࡽᨀǄ‫ى‬䘹⎸䲔㇇⌅ᴹ䖳ᕪⲴᖂ㓣‫ٿ‬㖞˖ণⴞḷᾲᘥ享൘‫ٷ‬䇮オ䰤ѝ㜭㺘
⽪Ǆ⭡Ҿᆳᱟᴹ‫ˈⲴٿ‬ᡰԕ㜭ཏሩ䇠ᗶᆖҐಘн㜭࠶㊫Ⲵᇎֻ䘋㹼࠶㊫Ǆᖃ❦࠶㊫Ⲵ↓⺞ᙗ
ҏᆼ‫׍ޘ‬䎆Ҿᖂ㓣‫ٿ‬㖞Ⲵ↓⺞ᙗǄFind-S㇇⌅ᴹᴤᕪⲴᖂ㓣‫ٿ‬㖞ˈ䲔Ҷ‫ٷ‬ᇊⴞḷᾲᘥ享൘‫ٷ‬
䇮オ䰤ѝˈᆳ䘈ᴹਖа仍ཆⲴᖂ㓣‫ٿ‬㖞ࡽᨀ˖ԫօᇎֻˈ䲔䶎ᆳⲴ䘶ᇎֻਟ⭡ަԆ⸕䇶䙫䗁
᧘ࠪˈ੖ࡉᆳѪ৽ֻǄ*
൘⹄ウަԆⲴᖂ㓣᧘⨶ᯩ⌅ᰦˈᴹᗵ㾱⢒䇠䘉⿽ᖂ㓣‫ٿ‬㖞Ⲵᆈ൘৺ަᕪᓖǄа⿽㇇⌅ྲ
᷌ᴹ‫ٿ‬ᙗ䎺ᕪˈ䛓ᆳⲴᖂ㓣㜭࣋䎺ᕪˈਟԕ࠶㊫ᴤཊⲴᵚ㿱ᇎֻǄḀӋᖂ㓣‫ٿ‬㖞ᱟሩ㊫࡛Ⲵ
‫ٷ‬ᇊˈԕ⺞ᇊⴞḷᾲᘥⲴ㤳തǄྲĀ‫ٷ‬䇮オ䰤 H वਜ਼ⴞḷᾲᘥāǄަԆⲴᖂ㓣‫ٿ‬㖞ਚᱟሩ‫ٷ‬
䇮䘋㹼ᧂᒿˈԕ᧿䘠‫ྭٿ‬〻ᓖˈ∄ྲĀ‫ੁٿ‬Ҿ⢩↺‫ٷ‬䇮ˈ㘼нᱟа㡜‫ٷ‬䇮ǄāḀӋ‫ٿ‬㖞䳀ਜ਼
൘ᆖҐಘѝнਟᴤ᭩ˈྲ䘉䟼ᡰ䇘䇪ⲴֻᆀǄ൘ㅜ 11 ઼ 12 ㄐਟԕⴻࡠ᰾⺞㺘⽪ᖂ㓣‫ٿ‬㖞Ⲵ
㌫㔏ˈᆳԜሶ‫ٿ‬㖞㺘⽪Ѫᯝ䀰Ⲵ䳶ਸᒦਟ⭡ᆖҐಘ᫽㓥Ǆ
ᵜㄐⲴ㾱⛩वᤜ˖
ᾲᘥᆖҐਟⴻ֌ᱟᩌ㍒亴ᇊѹ▌൘‫ٷ‬䇮オ䰤Ⲵ䗷〻Ǆ
‫ٷ‬䇮Ⲵа㡜ࡠ⢩↺‫ٿ‬ᒿ㔃ᶴਟԕᇊѹ൘ԫօᾲᘥᆖҐ䰞仈ѝˈᆳᨀ‫׋‬Ҷа⿽ᴹ⭘Ⲵ
㔃ᶴԕ‫ׯ‬Ҿ‫ٷ‬䇮オ䰤Ⲵᩌ㍒Ǆ
Find-S ㇇⌅֯⭘Ҷа㡜ࡠ⢩↺ᒿˈ൘‫ٿ‬ᒿ㔃ᶴⲴањ࠶᭟кᢗ㹼Ⲵа㡜ࡠ⢩↺ᩌ
㍒ˈԕራ᢮оṧֻа㠤Ⲵᴰ⢩↺‫ٷ‬䇮Ǆ
‫ى‬䘹⎸䲔㇇⌅࡙⭘а㡜ࡠ⢩↺ᒿˈ䙊䗷⑀䘋ൠ䇑㇇ᶱབྷ⢩↺‫ٷ‬䇮䳶ਸ S ઼ᶱབྷа㡜
‫ٷ‬䇮䳶ਸ G 䇑㇇ਈරオ䰤˄ণᡰᴹо䇝㓳ᮠᦞа㠤Ⲵ‫ٷ‬䇮䳶˅Ǆ
⭡Ҿ S ઼ G Ӿᮤњ‫ٷ‬䇮䳶ਸѝࡂ࠶ࠪҶо䇝㓳ᮠᦞа㠤Ⲵ䛓䜘࠶䳶ਸˈᆳԜᨀ‫׋‬
ҶሩⴞḷᾲᘥⲴн⺞ᇊᙗ᧿䘠Ǆਜ਼ᴹཊњ‫ٷ‬䇮Ⲵਈරオ䰤ਟԕ⭘ᶕࡔᯝᆖҐಘᱟ੖
ᐢ᭦ᮋࡠҶⴞḷᾲᘥ˗ࡔᯝ䇝㓳ᮠᦞᱟ੖на㠤˗ӗ⭏ḕ䈒ԕ䘋а↕㋮ॆਈරオ䰤˗
ԕ৺⺞ᇊᵚ㿱䗷Ⲵᇎֻᱟ੖㜭⭘нᆼ‫ޘ‬ᆖҐࡠⲴᾲᘥᶕᰐ↗ѹൠ࠶㊫Ǆ
ਈරオ䰤઼‫ى‬䘹⎸䲔㇇⌅Ѫ⹄ウᾲᘥᆖҐᨀ‫׋‬Ҷа⿽ᴹ⭘ⲴṶᷦˈ❦㘼䘉а㇇⌅㕪
ቁ励ἂᙗˈ⢩࡛ᱟ൘䙷ࡠᴹಚ༠Ⲵᮠᦞԕ৺ⴞḷᾲᘥᰐ⌅൘‫ٷ‬䇮オ䰤ѝ㺘⽪Ⲵᛵߥ
лǄㅜ 10 ㄐ᧿䘠Ҷࠐ⿽สҾа㡜ࡠ⢩↺ᒿ‫ޣ‬㌫ⲴᾲᘥᆖҐ㇇⌅ˈᆳԜ㜭ཏ༴⨶ᴹ
ಚ༠ᮠᦞǄ
*
⌘᜿ᴰਾ䶒䘉њᖂ㓣‫ٿ‬㖞‫ٷ‬ᇊˈᆳवਜ਼ҶḀ⿽唈䇔᧘⨶ˈᡆ䶎অ䈳᧘⨶Ǆ
ᖂ㓣ᆖҐ㇇⌅㜭ཏሩᵚ㿱ᮠᦞ䘋㹼࠶㊫ˈᱟഐѪᆳԜ൘䘹ᤙа㠤Ⲵ‫ٷ‬䇮ᰦ䳀ਜ਼Ⲵᖂ
㓣‫ٿ‬㖞Ǆ‫ى‬䘹⎸䲔㇇⌅ѝⲴ‫ٿ‬㖞Ѫ˖ⴞḷᾲᘥਟԕ൘‫ٷ‬䇮オ䰤ѝ᢮ࡠ˄cęH˅Ǆ䗃
ࠪⲴ‫ٷ‬䇮઼ሩਾ㔝ᇎֻⲴ࠶㊫ਟ⭡䘉аࡽᨀ৺䇝㓳ṧֻ╄㓾᧘ࠪǄ
ྲ᷌‫ٷ‬䇮オ䰤㻛ᢙኅˈ֯ሩᓄᇎֻ䳶Ⲵ⇿ањᆀ䳶˄ᇎֻⲴᑲ䳶˅䜭ᴹањ‫ٷ‬䇮ˈ
ሶ֯‫ى‬䘹⎸䲔㇇⌅ѝⲴᖂ㓣‫ٿ‬㖞⎸ཡǄ❦㘼ˈ䘉ҏሶ⎸䲔ަሩᯠᇎֻ࠶㊫Ⲵ㜭࣋Ǆ
ᰐ‫Ⲵٿ‬ᆖҐಘᰐ⌅ሩᵚ㿱ṧֻ䘋㹼ᖂ㓣Ǆ
ᾲᘥᆖҐԕ৺֯⭘а㡜ࡠ⢩↺ᒿⲴ⴨‫⹄ޣ‬ウ⭡ᶕҏѵǄBruner et al.˄1957˅䖳ᰙൠሩӪ
㊫ⲴᾲᘥᆖҐ֌ࠪ⹄ウˈ㘼 Hunt & Hovland˄1963˅䖳ᰙሶަ㠚ࣘॆǄWinston˄1970˅Ⲵ
ᴹ਽Ⲵঊ༛䇪᮷ѝሶᾲᘥᆖҐⴻ֌ᱟवਜ਼⌋ॆ઼⢩ॆ᫽֌Ⲵᩌ㍒䗷〻ǄPlotkin˄1970, 1971˅
䖳ᰙൠᨀ‫׋‬ҶᖒᔿॆⲴ more-general-than ‫ޣ‬㌫ˈԕ৺ањ⴨‫Ⲵޣ‬ᾲᘥT-वᇩ˄൘ㅜ 10 ㄐѝ
䇘䇪˅ǄSimon ઼ Lea˄1973˅ሶᆖҐⲴ䗷〻ⴻ֌ᱟ൘‫ٷ‬䇮オ䰤ѝᩌ㍒Ⲵ䗷〻ǄަԆаӋ䖳
ᰙⲴᾲᘥᆖҐ㌫㔏वᤜ˄Popplestone 1969˗Michalski 1973˗Buchanan 1974˗Vere 1975˗
Hayes-Roth 1974˅Ǆབྷ䟿ⲴสҾㅖਧ㺘⽪ⲴᾲᘥᆖҐ㇇⌅ᐢ㻛ᔰਁࠪᶕǄㅜ 10 ㄐ᧿䘠Ҷࠐ⿽
䘁ᵏⲴᾲᘥᆖҐ㇇⌅Ǆवᤜ⭘а䱦䙫䗁㺘⽪ⲴᾲᘥᆖҐ㇇⌅ˈሩᴹಚ༠ᮠᦞᴹ励ἂᙗⲴ㇇⌅ˈ
ԕ৺ᖃⴞḷᾲᘥᰐ⌅൘ᆖҐಘⲴ‫ٷ‬䇮オ䰤ѝ㺘⽪ᰦ㜭䖳ྭൠ䱽㓗ᆖҐⲴ㇇⌅Ǆ
ਈරオ䰤઼‫ى‬䘹⎸䲔㇇⌅⭡ Mitchell˄1977ˈ1982˅ᨀࠪˈ䘉а㇇⌅ᐢᓄ⭘Ҿ䍘䉡࠶᷀
˄mass spectroscopy˅ѝⲴ㿴ࡉ᧘⨶˄Mitchell 1979˅ԕ৺ᓄ⭘ҾᆖҐᩌ㍒᧗ࡦ㿴ࡉ˄Mitchell
ǄHaussler˄1988˅䇱᰾ণ֯ᖃ‫ٷ‬䇮オ䰤ਚवਜ਼ㆰঅⲴ⢩ᖱਸਆᰦˈа㡜䗩⭼Ⲵ
et al. 1983˅
བྷሿṩᦞ䇝㓳ṧֻⲴᮠⴞᤷᮠ໎䮯ǄSmith & Rosenbloom˄1990˅ᨀࠪሩ G 䳶ਸⲴ㺘⽪䘋㹼
ㆰঅⲴᴤ᭩ˈԕ᭩䘋ަ⢩ᇊᛵߥлⲴ༽ᵲᙗˈHirsh˄1992˅ᨀࠪ൘ḀӋᛵߥлнᆈۘ G 䳶
ਸᰦᆖҐ䗷〻ѪṧֻᮠⴞⲴཊ亩ᔿ࠭ᮠǄSubramanian & Feigenbaum˄1986˅䇘䇪Ҷ⢩ᇊᛵ
ߥл䙊䗷࠶䀓ਈරオ䰤ԕ⭏ᡀᴹ᭸ḕ䈒а⿽ᯩ⌅Ǆ‫ى‬䘹⎸䲔㇇⌅ⲴањᴰབྷⲴᇎ䱵䲀ࡦᱟᆳ
㾱≲䇝㓳ᮠᦞᱟᰐಚ༠ⲴǄMitchell˄1979˅᧿䘠Ҷ䈕㇇⌅Ⲵа⿽ᢙኅˈԕ༴⨶ਟ亴㿱Ⲵᴹ
䲀ᮠ䟿Ⲵ䈟࠶㊫ṧֻˈHirsh˄1990, 1994˅ᨀࠪа⿽㢟ྭⲴᢙኅԕ༴⨶ާᴹᇎᮠ٬኎ᙗⲴ䇝
㓳ṧֻѝⲴᴹ䲀ಚ༠ǄHirsh˄1990˅᧿䘠Ҷа⿽䙂໎ਈරオ䰤ਸᒦ㇇⌅ˈᆳሶ‫ى‬䘹⎸䲔㇇
⌅ᢙኅࡠ㜭༴⨶⭡н਼㊫රⲴ٬㓖ᶏ㺘⽪Ⲵ䇝㓳ؑ᚟Ǆᶕ㠚⇿њ㓖ᶏⲴؑ᚟⭡ਈරオ䰤ᶕ㺘
⽪ˈ❦ਾ⭘Ӕਐਈරオ䰤Ⲵ࣎⌅ਸᒦ䘉Ӌ㓖ᶏǄSebag˄1994, 1996˅ኅ⽪Ҷа⿽㻛〠Ѫ᷀ਆ
ਈරオ䰤Ⲵᯩ⌅ᶕӾᴹಚ༠ᮠᦞѝᆖҐ᷀ਆᾲᘥǄӾ⇿њ↓ֻѝᆖࡠањ࠶・Ⲵਈරオ䰤ˈ
❦ਾ⭘䘉н਼ਈරオ䰤䘋㹼ᣅ⾘ԕ࠶㊫ᯠᇎֻǄྩ൘ࠐњ䰞仈亶ฏ䘋㹼Ҷᇎ傼ˈᗇࠪྩⲴᯩ
⌅਼ަԆᒯ⌋֯⭘Ⲵᖂ㓣ᯩ⌅ᴹ਼ṧ㢟ྭⲴᙗ㜭ˈྲߣㆆṁ઼ k-䘁䛫ᯩ⌅Ǆ
Ґ仈
2.1 䀓䟺ѪӰѸ EnjoySport ᆖҐԫ࣑Ⲵ‫ٷ‬䇮オ䰤ⲴབྷሿѪ 973Ǆྲ᷌໎࣐а኎ᙗ

WaterCurrentˈਟਆ٬ LightǃModerate ઼ Strongˈ䛓Ѹਟ㜭Ⲵᇎֻᮠ઼ਟ㜭Ⲵ‫ٷ‬䇮ᮠሶՊ໎
࣐ཊቁ˛᧘ᒯࡠа㡜ˈ໎࣐аᯠ኎ᙗ Aˈᴹ k ⿽ਆ٬ˈᇎֻᮠ઼‫ٷ‬䇮ᮠሶՊ໎࣐ཊቁ˛
2.2 ൘‫ى‬䘹⎸䲔㇇⌅ѝˈྲ᷌䇝㓳ṧֻ᤹㺘 2-1 ѝⲴ䘶

䘶ᒿࠪ⧠ˈ䈧࠶↕㔉ࠪ S ઼ G 䗩⭼
䳶ਸǄ㲭❦н䇪ṧֻࠪ⧠亪ᒿྲօˈᴰ㓸Ⲵਈරオ䰤⴨਼˄ѪӰѸ˛˅ˈ൘ѝ䰤↕僔ѝᗇࡠ
Ⲵ S ઼ G ӽሶ‫׍‬䎆Ҿ䈕亪ᒿǄᱟ੖ᴹ࣎⌅ሩ䇝㓳ṧֻᧂᒿˈԕ֯ EnjoySport ֻᆀѝⲴᡰᴹ S
઼ G 䳶ਸⲴѝ䰤㔃᷌Ⲵབྷሿѻ઼Ѫᴰሿ˛
2.3 㔗㔝㘳㲁 EnjoySport ᆖҐԫ઼࣑ 2.2 㢲ѝ᧿䘠Ⲵ‫ٷ‬䇮オ䰤 HǄྲ᷌ᇊѹањᯠⲴ‫ٷ‬

䇮オ䰤 H´ˈᆳवਜ਼ H ѝᡰᴹ‫ٷ‬䇮Ⲵᡀሩ᷀ਆǄྲ H´ѝа‫ٷ‬䇮Ѫ˖
<?, Cold, High, ?, ?, ?>Ģ<Sunny, ?, High, ?, ?, Same>
䈅䐏䑚䘀㹼֯⭘䈕‫ٷ‬䇮オ䰤 H´Ⲵ‫ى‬䘹⎸䲔㇇⌅ˈ㔉ᇊⲴ䇝㓳ṧֻྲ㺘 2-1 ᡰ⽪˄䴰㾱

࠶↕ࡇࠪ S ઼ G 䳶ਸ˅Ǆ
2.4 ‫ٷ‬ᇊаᇎֻオ䰤वਜ਼ xˈy ᒣ䶒ѝⲴᮤᮠ⛩ˈ‫ٷ‬䇮䳶ਸ H Ѫ⸙ᖒ䳶Ǆᴤ㋮⺞ൠˈ‫ٷ‬

䇮ⲴᖒᔿѪ aİxİbˈcİyİdˈަѝ a,b,c,d Ѫԫ᜿ᮤᮠǄ
˄a˅㘳㲁ሩᓄҾлമᡰ⽪↓ֻ˄ˇ˅઼৽ֻ˄ˉ˅䳶ਸⲴਈරオ䰤ˈᆳⲴ S 䗩⭼ᱟӰ
Ѹ˛߉ࠪަѝⲴ‫ٷ‬䇮ᒦ൘മѝ⭫ࠪǄ
˄b˅ਈරオ䰤Ⲵ G 䗩⭼ᱟӰѸˈ߉ࠪަѝⲴ‫ٷ‬䇮ᒦ൘മѝ⭫ࠪǄ
˄c˅‫ٷ‬ᇊᆖҐಘਟᨀࠪањᯠᇎֻ(x, y)ˈᒦ㾱≲ᯭᮉ㘵䘋㹼࠶㊫ˈ䈅㔉ࠪањḕ䈒ˈ
ᰐ䇪ᯭᮉ㘵ᘾṧ࠶㊫䜭㜭‫؍‬䇱߿ሿਈරオ䰤Ǆ޽㔉ࠪањн㜭‫؍‬䇱Ⲵḕ䈒Ǆ
˄d˅֌Ѫᯭᮉ㘵ˈྲ᷌ᜣ䇙ᆖҐಘᆖҐа⢩ᇊⲴⴞḷᾲᘥ˄ྲ 3İxİ5ˈ2İyİ9˅
ˈ
Ѫ֯‫ى‬䘹⎸䲔㇇⌅ᆼ‫ޘ‬ᆖҐࡠⴞḷᾲᘥˈ䴰㾱ᨀ‫ⲴⲴ׋‬䇝㓳ṧֻᮠⴞᴰሿᱟཊቁǄ
2.5 䈧ⴻԕлⲴ↓઼ֻ৽ֻᒿֻˈᆳԜ᧿䘠ⲴᾲᘥᱟĀєњտ൘਼аᡯ䰤ѝⲴӪāǄ⇿њ
䇝㓳ṧֻ᧿䘠Ҷањᴹ ᴹᒿሩˈ⇿њӪ⭡ަᙗ࡛ǃཤਁ仌㢢˄black, brown ᡆ blonde˅ǃ䓛儈
˄tall, medium ᡆ short˅ԕ৺ഭ㉽˄US, French, German, Irish, Indian, Chinese ᡆ Portuguese˅Ǆ
+ < <male brown tall US>, <female black short US> >
+ < <male brown short French>, <female black short US> >
- < <female brown tall German>, <female black short Indian> >
+ < <male brown tall Irish>, <female brown short Irish> >
㘳㲁൘䘉ӋᇎֻкᇊѹⲴ‫ٷ‬䇮オ䰤Ѫ˖ަѝᡰᴹ‫ٷ‬䇮ԕаሩ 4 ‫ݳ‬㓴㺘⽪ˈަѝ⇿њ٬㓖
ᶏо EnjoySport ѝⲴ‫ٷ‬䇮㺘⽪⴨լˈਟԕѪ˖⢩ᇊ٬ǃĀ?āᡆ㘵ĀāǄֻྲˈл䶒Ⲵ‫ٷ‬䇮˖
< <male ? Tall ? > <female ? ? French> >
ᆳ㺘⽪Ҷᡰᴹ䘉ṧⲴᴹᒿሩ˖ㅜањӪѪ儈њ⭧ᙗ˄ഭ㉽઼ਁ㢢ԫ᜿˅ˈㅜҼњӪѪ⌅
ഭྣᙗ˄ਁ㢢઼䓛儈ԫ᜿˅Ǆ
˄a˅ṩᦞк䘠ᨀ‫Ⲵ׋‬䇝㓳ṧ઼ֻ‫ٷ‬䇮㺘⽪ˈ᡻ࣘᢗ㹼‫ى‬䘹⎸䲔㇇⌅Ǆ⢩࡛ᱟ㾱߉ࠪ༴
⨶Ҷ⇿ањ䇝㓳ṧֻਾਈරオ䰤Ⲵ⢩↺઼а㡜䗩⭼Ǆ
˄b˅䇑㇇㔉ᇊⲴ‫ٷ‬䇮オ䰤ѝᴹཊቁ‫ٷ‬䇮ол䶒Ⲵ↓ֻа㠤˖
+ < <male black short Portuguese> <female blonde tall Indian> >
˄c˅ྲ᷌ᆖҐಘਚᴹањ䇝㓳ṧֻྲ˄b˅ѝᡰ⽪ˈ⧠൘⭡ᆖҐಘᨀࠪḕ䈒ˈᒦ⭡ᯭᮉ
㘵㔉ࠪަ࠶㊫Ǆ≲ࠪањ⢩ᇊⲴḕ䈒ᒿࡇˈԕ‫؍‬䇱ᆖҐಘ᭦ᮋࡠঅњ↓⺞Ⲵ‫ٷ‬䇮ˈ㘼н䇪䈕
‫ٷ‬䇮ᱟଚањ˄‫ٷ‬ᇊⴞḷᾲᘥਟԕ֯⭘㔉ᇊⲴ‫ٷ‬䇮㺘⽪䈝䀰ᶕ᧿䘠˅Ǆ≲ࠪᴰ⸝Ⲵḕ䈒ᒿࡇǄ
䘉аᒿࡇⲴ䮯ᓖо䰞仈˄b˅ⲴㆄṸᴹӰѸ‫ޣ‬㚄˛
˄d˅⌘᜿ࡠ䘉䟼Ⲵ‫ٷ‬䇮㺘⽪䈝䀰н㜭ཏ㺘⽪䘉ӋᇎֻкⲴᡰᴹᾲᘥ˄ྲᡁԜਟᇊѹࠪ
а㌫ࡇⲴ↓઼ֻ৽ֻˈᆳԜᒦ⋑ᴹ⴨ᓄⲴਟ᧿䘠‫ٷ‬䇮˅Ǆྲ᷌㾱ᢙኅ䘉а䈝䀰ˈ֯ަ㜭ཏ㺘
䗮䈕ᇎֻ䈝䀰кⲴᡰᴹᾲᘥˈ䛓Ѹ˄c˅ⲴㆄṸᓄ䈕ྲօᴤ᭩Ǆ
2.6 ᆼᡀਈරオ䰤㺘⽪ᇊ⨶Ⲵ䇱᰾˄ᇊ⨶ 2.1˅Ǆ
2.7 㘳㲁ањᾲᘥᆖҐ䰞仈ˈަѝ⇿њᇎֻѪаᇎᮠˈ㘼⇿њ‫ٷ‬䇮ѪᇎᮠѝⲴ४䰤Ǆ㋮
⺞ൠᇊѹѪ˖‫ٷ‬䇮オ䰤 H ѝⲴ⇿њ‫ٷ‬䇮ᖒᔿѪ a<x<bˈަѝ aǃb Ѫԫ᜿ᇎᑨᮠˈx ԓ㺘䈕
ᇎֻǄֻྲ 4.5<x<6.1 䘉њ‫ٷ‬䇮ሶ 4.5 ઼ 6.1 ѻ䰤Ⲵᇎֻࡂ࠶Ѫ↓ֻˈަԆѪ৽ֻǄㆰ㾱䀓
䟺ѪӰѸнᆈ൘ањሩԫ᜿↓ֻ䳶ਸ䜭а㠤Ⲵᴰ⢩↺‫ٷ‬䇮Ǆ䈅‫؞‬᭩‫ٷ‬䇮Ⲵ㺘⽪ᯩ⌅ԕ䚯‫ݽ‬䘉
а㕪⛩Ǆ
2.8 ᵜㄐѝᤷࠪྲ᷌㔉ᇊањᰐ‫ٷⲴٿ‬䇮オ䰤˄ণᇎֻⲴᑲ䳶˅ˈᆖҐಘሶਁ⧠⇿аᵚ㿲
ሏⲴᇎֻሶࡊྭоਈරオ䰤ѝॺᮠⲴᡀઈ३䝽ˈ㘼н䇪ᐢ㓿䗷ҶᘾṧⲴ䇝㓳ṧֻǄ䇱᰾䘉а
㔃䇪Ǆ⺞࠷ൠ䇢ˈ䇱᰾ሩҾԫ᜿ᇎֻオ䰤Xˈԫ᜿䇝㓳ṧֻ䳶Dˈ઼ԫ᜿нवਜ਼൘DѝⲴᇎ
ֻxęXˈྲ᷌HᱟXⲴᑲ䳶ˈ䛓Ѹ൘VSH,DѝᴹᚠྭॺᮠⲴ‫ٷ‬䇮ሶxࡂ࠶Ѫ↓ֻˈਖཆॺᮠࡂ
࠶Ѫ৽ֻǄ
2.9 ᴹаᆖҐ䰞仈ˈަѝ⇿њᇎֻ䜭⭡nњᐳቄ٬኎ᙗa1, a2, … ,anⲴਸਆᶕ᧿䘠Ǆഐ↔ˈ

ањިරⲴᇎֻྲл˖
(a1=T)ġ(a2=F)ġ…ġ(an=T)
⧠㘳㲁ањ‫ٷ‬䇮オ䰤 H ѝˈ⇿њ‫ٷ‬䇮ᱟ䘉Ӌ኎ᙗ㓖ᶏⲴ᷀ਆˈֻྲ˖
(a1=T)Ģ(a5=F)Ģ(a7=T)
䇮䇑а㇇⌅ˈᆳ㓿䗷а㌫ࡇⲴṧֻ䇝㓳ਾ䗃ࠪања㠤Ⲵ‫ٷ‬䇮˄ྲᆈ൘Ⲵ䈍˅Ǆ㇇⌅Ⲵ
ᰦ䰤㾱≲Ѫ n ઼䇝㓳ṧֻᮠⴞⲴཊ亩ᔿ࠭ᮠǄ
2.10 ᇎ⧠ Find-S ㇇⌅Ǆ俆‫ˈݸ‬傼䇱ᆳਟᡀ࣏ൠӗ⭏ 2.4 㢲ѝ EnjoySport ֻᆀѝ਴↕僔㔃

᷌Ǆ❦ਾ֯⭘䘉а〻ᒿˈ⹄ウѪҶᆖҐࡠ⺞࠷Ⲵⴞḷᾲᘥᡰ䴰Ⲵ䲿ᵪ䇝㓳ṧֻᮠⴞǄᇎ⧠а
䇝㓳ṧֻ⭏ᡀಘᶕ⭏ᡀ䘉Ӌ䲿ᵪⲴᇎֻˈ޽⭘л䶒Ⲵⴞḷᾲᘥӗ⭏࠶㊫㔃᷌˖
<SunnyˈWarmˈ˛ˈ˛ˈ˛ˈ˛>
䈅⭘䲿ᵪӗ⭏Ⲵṧֻ䇝㓳֐Ⲵ Find-S ㇇⌅ˈᒦ⍻䟿䴰㾱ཊቁṧֻ᡽㜭֯〻ᒿⲴ‫ٷ‬䇮о

ⴞḷᾲᘥ⴨ㅹǄ㜭੖亴⍻ᡰ䴰Ⲵᒣ൷ṧֻᮠⴞ˛䘀㹼䈕ᇎ傼 20 ⅑ᒦᣕ੺ᡰ䴰ṧֻⲴᒣ൷ᮠǄ
䘉аᮠⴞՊᘾṧ䲿⵰ⴞḷᾲᘥѝⲴĀ?āᮠⴞ㘼ਈࣘ˛ԕ৺ᆳՊᘾṧ䲿⵰ᇎֻᡆ‫ٷ‬䇮ѝ኎ᙗ
Ⲵᮠⴞ㘼ਈࣘ˛
ㅢ3ㄖ ߩㆌṇᆜҖ
决策树学习是应用最广的归纳推理算法之一。它是一种逼近离散函数的方法，
且对噪声数据有很好的鲁棒性，能够学习析取表达式。本章描述了一系列决策树学
习算法，包括象 ID3、ASSISTANT 和 C4.5 这样广为应用的算法。这些决策树学
习方法搜索完整表示的假设空间，从而避免了受限假设空间的不足。决策树学习的
归纳偏置是优先选择较小的树。
3.1 ㆶԁ
‫އ‬ㄪᷥᄺдᰃϔ⾡䘐䖥⾏ᬷؐⳂᷛߑ᭄ⱘᮍ⊩ˈ೼䖭⾡ᮍ⊩Ёᄺдࠄⱘߑ᭄㹿㸼⼎
Ўϔỉ‫އ‬ㄪᷥǄᄺдᕫࠄⱘ‫އ‬ㄪᷥг㛑‫ݡ‬㹿㸼⼎Ў໮Ͼ if-then ⱘ㾘߭ˈҹᦤ催ৃ䇏ᗻǄ
䖭⾡ᄺдㅫ⊩ᰃ᳔⌕㸠ⱘᔦ㒇᥼⧚ㅫ⊩ПϔˈᏆ㒣㹿៤ࡳഄᑨ⫼ࠄҢᄺдए⭫䆞ᮁࠄᄺ
д䆘Ԅ䌋ℒ⬇䇋ⱘֵ⫼亢䰽ⱘᑓ䯨乚ඳǄ
3.2 ߩㆌṇ㺞⽰⌋
‫އ‬ㄪᷥ䗮䖛ᡞᅲ՟Ңḍ㒧⚍ᥦ߫˄sort˅ࠄᶤϾ৊ᄤ㒧⚍ᴹߚ㉏ᅲ՟ˈ৊ᄤ㒧⚍े
Ўᅲ՟᠔ሲⱘߚ㉏ǄᷥϞⱘ↣ϔϾ㒧⚍ᣛᅮњᇍᅲ՟ⱘᶤϾሲᗻ˄attribute˅ⱘ⌟䆩ˈ
ᑊϨ䆹㒧⚍ⱘ↣ϔϾৢ㒻ߚᬃᇍᑨѢ䆹ሲᗻⱘϔϾৃ㛑ؐǄߚ㉏ᅲ՟ⱘᮍ⊩ᰃҢ䖭ỉᷥ
ⱘḍ㒧⚍ᓔྟˈ⌟䆩䖭Ͼ㒧⚍ᣛᅮⱘሲᗻˈ✊ৢᣝ✻㒭ᅮᅲ՟ⱘ䆹ሲᗻؐᇍᑨⱘᷥᵱ৥
ϟ⿏ࡼǄ䖭Ͼ䖛⿟‫ݡ‬೼ҹᮄ㒧⚍ЎḍⱘᄤᷥϞ䞡໡Ǆ
೒ 3-1 ὖᗉ PlayTennis ⱘ‫އ‬ㄪᷥ
ߚ㉏ϔϾḋ՟ⱘᮍ⊩ᰃˈᇚ݊⊓ḍ㒧⚍ᥦ߫ࠄড়䗖ⱘ৊ᄤ㒧⚍ˈ✊ৢ䖨ಲϢ䖭Ͼ৊ᄤ㒧⚍݇㘨ⱘߚ㉏
˄ᴀ՟ЁЎ Yes ៪ No˅Ǆ䖭ỉ‫އ‬ㄪᷥḍ᥂໽⇨ߚ㉏Ā᯳ᳳ݁Ϟजᰃ৺䗖ড়ᠧ㔥⧗āǄ
೒ 3-1 ⬏ߎњϔỉ‫݌‬ൟⱘᄺдࠄⱘ‫އ‬ㄪᷥǄ䖭ỉ‫އ‬ㄪᷥḍ᥂໽⇨ᚙ‫㉏ߚމ‬Ā᯳ᳳ݁
Ϟजᰃ৺䗖ড়ᠧ㔥⧗āǄ՟བˈϟ䴶ⱘᅲ՟˖
< Outlook=SunnyˈTemperature=HotˈHumidity=HighˈWind=Strong >
ᇚ㹿⊓ⴔ䖭ỉ‫އ‬ㄪᷥⱘ᳔Ꮊߚᬃ৥ϟᥦ߫ˈ಴㗠㹿䆘ᅮЎড՟˄гህᰃ䖭ỉᷥ乘⌟䖭Ͼ
ᅲ՟ PlayTennis=No˅Ǆ䖭ỉᷥҹঞ㸼 3-2 Ё⫼ᴹⓨ⼎ ID3 ᄺдㅫ⊩ⱘ՟ᄤᨬ㞾˄Quinlan
1986˅Ǆ
䗮ᐌ‫އ‬ㄪᷥҷ㸼ᅲ՟ሲᗻؐ㑺ᴳⱘড়প˄conjunction˅ⱘᵤপᓣ˄disjunction˅ǄҢ
ᷥḍࠄᷥ৊ⱘ↣ϔᴵ䏃ᕘᇍᑨϔ㒘ሲᗻ⌟䆩ⱘড়পˈᷥᴀ䑿ᇍᑨ䖭ѯড়পⱘᵤপǄ՟བˈ
೒ 3-1 㸼⼎ⱘ‫އ‬ㄪᷥᇍᑨѢҹϟ㸼䖒ᓣ˖
˄Outlook=Sunny Humidity=Normal˅
˄Outlook=Overcast˅
˄Outlook=Rain Wind=Weak˅
3.3 ߩㆌṇᆜҖⲺ䘸⭞䰤从
ሑㅵᏆ㒣ᓔথⱘ⾡⾡‫އ‬ㄪᷥᄺдㅫ⊩᳝䖭ḋ៪䙷ḋϡ໾ϔ㟈ⱘ㛑࡯੠㽕∖ˈ䗮ᐌ‫އ‬
ㄪᷥᄺд᳔䗖ড়‫݋‬᳝ҹϟ⡍ᕕⱘ䯂乬˖
z ᅲ՟ᰃ⬅Āሲᗻ-ؐāᇍ˄pair˅㸼⼎ⱘǄᅲ՟ᰃ⫼ϔ㋏߫೎ᅮⱘሲᗻ˄՟བˈ
Temperature˅੠ᅗӀⱘؐ˄՟བˈHot˅ᴹᦣ䗄ⱘǄ᳔ㅔऩⱘ‫އ‬ㄪᷥᄺдЁˈ↣
ϔϾሲᗻপᇥ᭄ⱘߚ⾏ⱘؐ˄՟བˈHotǃMildǃCold˅Ǆ✊㗠ˈᠽሩⱘㅫ⊩˄೼
3.7.2 㡖Ё䅼䆎˅г‫ܕ‬䆌໘⧚ؐඳЎᅲ᭄ⱘሲᗻ˄՟བˈ᭄ᄫ㸼⼎ⱘ⏽ᑺ˅Ǆ
z Ⳃᷛߑ᭄‫݋‬᳝⾏ᬷⱘ䕧ߎؐǄ೒ 3-1 ⱘ‫އ‬ㄪᷥ㒭↣Ͼᅲ՟䌟ќϔϾᏗᇨൟⱘߚ㉏
˄՟བˈyes ៪ no˅Ǆ‫އ‬ㄪᷥᮍ⊩ᕜᆍᯧᠽሩࠄᄺд᳝ϸϾҹϞ䕧ߎؐⱘߑ᭄Ǆ
ϔ⾡᳈ᔎ᳝࡯ⱘᠽሩㅫ⊩‫ܕ‬䆌ᄺд‫݋‬᳝ᅲ᭄ؐ䕧ߎⱘߑ᭄ˈሑㅵ‫އ‬ㄪᷥ೼䖭⾡
ᚙ‫މ‬ϟⱘᑨ⫼ϡ໾ᐌ㾕Ǆ
z ৃ㛑䳔㽕ᵤপⱘᦣ䗄˄disjunctive description˅ǄབϞ䴶ᣛߎⱘˈ‫އ‬ㄪᷥᕜ㞾✊
ഄҷ㸼њᵤপ㸼䖒ᓣǄ
z 䆁㒗᭄᥂ৃҹࣙ৿䫭䇃Ǆ‫އ‬ㄪᷥᄺдᇍ䫭䇃᳝ᕜདⱘ剕Ầᗻˈ᮴䆎ᰃ䆁㒗ḋ՟
᠔ሲⱘߚ㉏䫭䇃䖬ᰃᦣ䗄䖭ѯḋ՟ⱘሲᗻؐ䫭䇃Ǆ
z 䆁㒗᭄᥂ৃҹࣙ৿㔎ᇥሲᗻؐⱘᅲ՟Ǆ‫އ‬ㄪᷥᄺд⫮㟇ৃҹ೼᳝᳾ⶹሲᗻؐⱘ
䆁㒗ḋ՟ЁՓ⫼˄՟བˈҙ᳝ϔ䚼ߚ䆁㒗ḋ՟ⶹ䘧ᔧ໽ⱘ⑓ᑺ˅Ǆ䖭Ͼ䯂乬ᇚ
೼㄀ 3.7.4 ᇣ㡖Ё䅼䆎Ǆ
Ꮖ㒣থ⦄ᕜ໮ᅲ䰙ⱘ䯂乬ヺড়䖭ѯ⡍ᕕˈ᠔ҹ‫އ‬ㄪᷥᄺдᏆ㒣㹿ᑨ⫼ࠄᕜ໮䯂乬
ЁǄ՟བḍ᥂⮒⮙ߚ㉏ᙷ㗙˗ḍ᥂䍋಴ߚ㉏䆒໛ᬙ䱰˗ḍ᥂ᢪ⃴ᬃҬⱘৃ㛑ᗻߚ㉏䌋ℒ
⬇䇋ǄᇍѢ䖭ѯ䯂乬ˈḌᖗӏࡵ䛑ᰃ㽕ᡞḋ՟ߚ㉏ࠄ৘ৃ㛑ⱘ⾏ᬷؐᇍᑨⱘ㉏߿
˄category˅Ёˈ಴ℸ㒣ᐌ㹿⿄Ўߚ㉏䯂乬˄Classification Problem˅Ǆ
䖭ϔゴⱘ݊ԭ䚼ߚᰃ䖭ḋᅝᥦⱘǄ3.4 㡖㒭ߎᄺд‫އ‬ㄪᷥⱘ෎ᴀ ID3 ㅫ⊩ᑊⓨ⼎ᅗ

ⱘ‫݋‬ԧ᪡԰Ǆ3.5 㡖ߚᵤՓ⫼䖭⾡ᄺдㅫ⊩䖯㸠ⱘ‫؛‬䆒ぎ䯈᧰㋶ˈᑊϢ㄀ 2 ゴⱘㅫ⊩䖯
㸠њ↨䕗Ǆ3.6 㡖ࠏ⬏њ‫އ‬ㄪᷥᄺдㅫ⊩ⱘᔦ㒇‫أ‬㕂ˈᑊ᳈ϔ㠀࣪ⱘ᥶㋶њϔ⾡㹿⿄Ў
༹ഢྚࠗߔⱘᔦ㒇‫أ‬㕂ˈ䆹‫أ‬㕂Ӭ‫ܜ‬䗝ᢽ᳔ㅔऩⱘ‫؛‬䆒Ǆ3.7 㡖䅼䆎њ䆁㒗᭄᥂ⱘ䖛ᑺ
ᢳড়˄overfitting˅ˈҹঞ㾷‫އ‬䖭⾡䯂乬ⱘㄪ⬹ˈ↨བ㾘߭ৢׂ࠾˄post-pruning˅Ǆ䖭ϔ
㡖䖬䅼䆎њϔѯ᳈⏅ܹⱘ䆱乬ˈ↨བᇚㅫ⊩ᠽሩҹ䗖ᑨᅲ᭄ؐሲᗻǃᏺ᳝᳾㾖⌟ࠄሲᗻ
ⱘ䆁㒗᭄᥂ǃҹঞ᳝ϡৠҷӋⱘሲᗻǄ
3.4 ะᵢⲺߩㆌṇᆜҖ㇍⌋
໻໮᭄Ꮖᓔথⱘ‫އ‬ㄪᷥᄺдㅫ⊩ᰃϔ⾡Ḍᖗㅫ⊩ⱘবԧǄ䆹ㅫ⊩䞛⫼㞾乊৥ϟⱘ䋾
ှ᧰㋶䘡ग़ৃ㛑ⱘ‫އ‬ㄪᷥぎ䯈Ǆ䖭⾡ᮍ⊩ᰃ ID3 ㅫ⊩˄Quinlan 1986˅੠ৢ㒻ⱘ C4.5
ㅫ⊩˄Quinlan 1993˅ⱘ෎⸔ˈгᰃ䖭䞠䅼䆎ⱘ䞡⚍Ǆ䖭ϔ㡖ᇚ㒭ߎ‫އ‬ㄪᷥᄺдⱘ෎ᴀ
ㅫ⊩ˈ໻㟈ⳌᔧѢ ID3 ㅫ⊩Ǆ೼ 3.7 㡖៥Ӏ㗗㰥䆹෎ᴀㅫ⊩ⱘϔѯᠽሩˈࣙᣀ㹿ড়ᑊࠄ
C4.5 ⱘᠽሩ੠݊Ҫϔѯ䕗ᮄⱘ‫އ‬ㄪᷥᄺдㅫ⊩Ǆ
෎ᴀⱘ ID3 ㅫ⊩䗮䖛㞾乊৥ϟᵘ䗴‫އ‬ㄪᷥᴹ䖯㸠ᄺдǄᵘ䗴䖛⿟ᰃҢĀાϔϾሲ

ᗻᇚ೼ᷥⱘḍ㒧⚍㹿⌟䆩˛ā䖭Ͼ䯂乬ᓔྟⱘǄЎњಲㄨ䖭Ͼ䯂乬ˈՓ⫼㒳䅵⌟䆩ᴹ⹂
ᅮ↣ϔϾᅲ՟ሲᗻऩ⣀ߚ㉏䆁㒗ḋ՟ⱘ㛑࡯Ǆߚ㉏㛑࡯᳔དⱘሲᗻ㹿䗝԰ᷥⱘḍ㒧⚍ⱘ
⌟䆩Ǆ✊ৢЎḍ㒧⚍ሲᗻⱘ↣Ͼৃ㛑ؐѻ⫳ϔϾߚᬃˈᑊᡞ䆁㒗ḋ՟ᥦ߫ࠄ䗖ᔧⱘߚᬃ
˄гህᰃˈḋ՟ⱘ䆹ሲᗻؐᇍᑨⱘߚᬃ˅ПϟǄ✊ৢ䞡໡ᭈϾ䖛⿟ˈ⫼↣Ͼߚᬃ㒧⚍݇
㘨ⱘ䆁㒗ḋ՟ᴹ䗝প೼䆹⚍㹿⌟䆩ⱘ᳔ՇሲᗻǄ䖭ᔶ៤њᇍড়Ḑ‫އ‬ㄪᷥⱘ䋾ှ᧰㋶
˄greedy search˅ˈгህᰃㅫ⊩Ңϡಲ⒃䞡ᮄ㗗㰥ҹࠡⱘ䗝ᢽǄ㸼 3-1 ᦣ䗄њ䆹ㅫ⊩ⱘϔ
Ͼㅔ࣪⠜ᴀüüϧ䮼⫼ᴹᄺдᏗᇨؐߑ᭄˄ेὖᗉᄺд˅Ǆ
㸼 3-1 ϧ⫼ѢᄺдᏗᇨߑ᭄ⱘ ID3 ㅫ⊩ὖ㽕
ID3 ᰃϔ⾡㞾乊৥ϟ๲䭓ᷥⱘ䋾ှㅫ⊩ˈ೼↣Ͼ㒧⚍䗝প㛑᳔དഄߚ㉏ḋ՟ⱘሲᗻǄ㒻㓁䖭Ͼ䖛⿟Ⳉ
ࠄ䖭ỉᷥ㛑ᅠ㕢ߚ㉏䆁㒗ḋ՟ˈ៪᠔᳝ⱘሲᗻ䛑Փ⫼䖛њǄ
ID3(ExamplesˈTarget_attributeˈAttributes)
Examples े䆁㒗ḋ՟䲚ǄTarget_attribute ᰃ䖭ỉᷥ㽕乘⌟ⱘⳂᷛሲᗻǄAttributes
ᰃ䰸Ⳃᷛሲᗻ໪կᄺдࠄⱘ‫އ‬ㄪᷥ⌟䆩ⱘሲᗻ߫㸼Ǆ䖨ಲ㛑ℷ⹂ߚ㉏㒭ᅮ
Examples ⱘ‫އ‬ㄪᷥǄ
z ߯ᓎᷥⱘ Root 㒧⚍
z བᵰ Examples 䛑Ўℷˈ䙷М䖨ಲ label =+ ⱘऩ㒧⚍ᷥ Root
z བᵰ Examples 䛑Ўডˈ䙷М䖨ಲ label =- ⱘऩ㒧⚍ᷥ Root
z བᵰ Attributes Ўぎˈ䙷М䖨ಲऩ㒧⚍ᷥ Rootˈlabel=Examples Ё᳔᱂䘡ⱘ
Target_attribute ؐ
z ৺߭
z AĕAttributes Ёߚ㉏ Examples 㛑࡯᳔ད*ⱘሲᗻ
z Root ⱘ‫އ‬ㄪሲᗻĕA
z ᇍѢAⱘ↣Ͼৃ㛑ؐvi
z ೼RootϟࡴϔϾᮄⱘߚᬃᇍᑨ⌟䆩A= vi
z Ҹ Examplesvi ЎExamplesЁ⒵䎇AሲᗻؐЎviⱘᄤ䲚
z བᵰ Examplesvi Ўぎ
z ೼䖭ϾᮄߚᬃϟࡴϔϾ৊ᄤ㒧⚍ˈ㒧⚍ⱘ label=Examples
Ё᳔᱂䘡ⱘ Target_attribute ؐ
z ৺߭೼䖭ϾᮄߚᬃϟࡴϔϾᄤᷥ ID3˄ Examplesvi ,
Target_attribute, Attributes-{A}˅
z 㒧ᴳ
z 䖨ಲ Root
*ḍ᥂݀ᓣ 3.4 ⱘᅮНˈ‫݋‬᳔᳝催ֵᙃ๲Ⲟ˄information gain˅ⱘሲᗻᰃ᳔དⱘሲᗻǄ
3.4.1 ଠѠኔᙝᱥᴶ֩Ⲻ࠼㊱ኔᙝϋ
ID3 ㅫ⊩ⱘḌᖗ䯂乬ᰃ䗝প೼ᷥⱘ↣Ͼ㒧⚍㽕⌟䆩ⱘሲᗻǄ៥ӀᏠᳯ䗝ᢽⱘᰃ᳔᳝
ࡽѢߚ㉏ᅲ՟ⱘሲᗻǄ䙷М㸵䞣ሲᗻӋؐⱘϔϾདⱘᅮ䞣ᷛ‫ޚ‬ᰃҔМਸ਼˛䖭䞠ᇚᅮНϔ
Ͼ㒳䅵ሲᗻˈ⿄ЎĀֵᙃ๲Ⲟ˄information gain˅
āˈ⫼ᴹ㸵䞣㒭ᅮⱘሲᗻऎߚ䆁㒗ḋ՟
ⱘ㛑࡯ǄID3 ㅫ⊩೼๲䭓ᷥⱘ↣ϔℹՓ⫼䖭Ͼֵᙃ๲Ⲟᷛ‫ޚ‬Ң‫׭‬䗝ሲᗻЁ䗝ᢽሲᗻǄ
3.4.1.1 ⭞⟫ᓜ䠅ṭׁⲺൽжᙝ
Ўњ㊒⹂ഄᅮНֵᙃ๲Ⲟˈ៥Ӏ‫ܜ‬ᅮНֵᙃ䆎Ёᑓ⊯Փ⫼ⱘϔϾᑺ䞣ᷛ‫⿄ˈޚ‬Ў➉
˄entropy˅
ˈᅗࠏ⬏њӏᛣḋ՟䲚ⱘ㒃ᑺ˄purity˅Ǆ㒭ᅮࣙ৿݇ѢᶤϾⳂᷛὖᗉⱘℷড
ḋ՟ⱘḋ՟䲚 Sˈ䙷М S Ⳍᇍ䖭ϾᏗᇨൟߚ㉏ⱘ➉Ў˖
Entropy(S) {-plog2p-plog2p ˄3.1˅
݊Ёpᰃ೼SЁℷ՟ⱘ↨՟ˈpᰃ೼SЁ䋳՟ⱘ↨՟Ǆ೼᳝݇➉ⱘ᠔᳝䅵ㅫЁ៥Ӏᅮ
Н 0log0 Ў 0Ǆ
В՟䇈ᯢˈ‫؛‬䆒 S ᰃϔϾ݇ѢᶤᏗᇨὖᗉⱘ᳝ 14 Ͼḋ՟ⱘ䲚ড়ˈᅗࣙᣀ 9 Ͼℷ՟

੠ 5 Ͼড՟˄៥Ӏ䞛⫼䆄ো[9+ˈ5-]ᴹὖᣀ䖭ḋⱘ᭄᥂ḋ՟˅Ǆ䙷М S ⳌᇍѢ䖭ϾᏗᇨ
ߚ㉏ⱘ➉˄Entropy˅Ў˖
Entropy ([9,5]) (9 / 14) log 2 (9 / 14) (5 / 14) log 2 (5 / 14)
=0.940 ˄3.2˅
⊼ᛣˈབᵰSⱘ᠔᳝៤ਬሲѢৠϔ㉏ˈ䙷МSⱘ➉Ў 0Ǆ՟བˈབᵰ᠔᳝ⱘ៤ਬᰃℷ
ⱘ ˄ p=1 ˅ ˈ 䙷 М p ህ ᰃ 0 ˈ Ѣ ᰃ Entropy(S) =
1 log 2 (1) (0) log 2 (0) 1 0 0 log 2 0 0 Ǆ঺໪ˈᔧ䲚ড়Ёℷডḋ՟ⱘ᭄䞣
Ⳍㄝᯊ➉Ў 1Ǆབᵰ䲚ড়Ёℷড՟ⱘ᭄䞣ϡㄝᯊˈ➉ҟѢ 0 ੠ 1 П䯈Ǆ೒ 3-2 ᰒ⼎њ݇
ѢᶤᏗᇨߚ㉏ⱘ➉ߑ᭄䱣ⴔpҢ 0 ࠄ 1 ব࣪ⱘ᳆㒓Ǆ
೒ 3-2 ݇ѢᶤᏗᇨߚ㉏ⱘⱘ➉ߑ᭄
೒Ё⬏ߎњ䱣ⴔℷ՟᠔ऴ↨՟pҢ 0 ࠄ 1ˈ➉ߑ᭄ব࣪ⱘ᳆㒓Ǆ
ֵᙃ䆎Ё➉ⱘϔ⾡㾷䞞ᰃˈ➉⹂ᅮњ㽕㓪ⷕ䲚ড় S Ёӏᛣ៤ਬ˄ेҹഛࣔⱘὖ⥛䱣
ᴎᢑߎⱘϔϾ៤ਬ˅ⱘߚ㉏᠔䳔㽕ⱘ᳔ᇥѠ䖯ࠊԡ᭄ǄВ՟ᴹ䇈ˈབᵰ p ᰃ 1ˈ᥹ᬊ
㗙ⶹ䘧ᢑߎⱘḋ՟ᖙЎℷˈ᠔ҹϡᖙথӏԩ⍜ᙃˈℸᯊⱘ➉Ў 0Ǆ঺ϔᮍ䴶ˈབᵰ p ᰃ
0.5ˈᖙ乏⫼ϔϾѠ䖯ࠊԡᴹ䇈ᯢᢑߎⱘḋ՟ᰃℷ䖬ᰃ䋳Ǆབᵰ p ᰃ 0.8ˈ䙷Мᇍ᠔䳔
ⱘ⍜ᙃ㓪ⷕᮍ⊩ᰃ䌟㒭ℷ՟䲚ড়䕗ⷁⱘ㓪ⷕˈৃ㛑ᗻ䕗ᇣⱘড՟䲚ড়䕗䭓ⱘ㓪ⷕˈᑇഛ
↣ᴵ⍜ᙃⱘ㓪ⷕᇥѢ 1 ϾѠ䖯ࠊԡǄ
㟇ℸ៥Ӏ䅼䆎њⳂᷛߚ㉏ᰃᏗᇨൟⱘᚙ‫މ‬ϟⱘ➉Ǆ᳈ϔ㠀ⱘˈབᵰⳂᷛሲᗻ‫݋‬᳝ c
Ͼϡৠⱘؐˈ䙷М S ⳌᇍѢ c Ͼ⢊ᗕ˄c-wise˅ⱘߚ㉏ⱘ➉ᅮНЎ˖
c
Entropy ( S ) { ¦ pi log 2 pi (3.3)
i 1
݊ЁpiᰃSЁሲѢ㉏߿iⱘ↨՟Ǆ䇋⊼ᛣᇍ᭄ⱘᑩ᭄ҡ✊Ў 2ˈॳ಴ᰃ➉ᰃҹӂ䘑࡬ԡ
ⱘϾ᭄ᴹᑺ䞣㓪ⷕ䭓ᑺⱘǄৠᯊ⊼ᛣབᵰⳂᷛሲᗻ‫݋‬᳝cϾৃ㛑ؐˈ䙷М➉᳔໻ৃ㛑ࠄ
log2cǄ
3.4.1.2 ⭞‫ؗ‬ᚥ໔ⴀᓜ䠅ᵕᵑⲺ⟫䲃ք
Ꮖ㒣᳝њ➉԰Ў㸵䞣䆁㒗ḋ՟䲚ড়㒃ᑺⱘᷛ‫⦄ˈޚ‬೼ৃҹᅮНሲᗻߚ㉏䆁㒗᭄᥂ⱘ
ᬜ࡯ⱘᑺ䞣ᷛ‫ޚ‬Ǆ䖭Ͼᷛ‫ޚ‬㹿⿄ЎĀֵᙃ๲Ⲟ˄information gain˅
āǄㅔऩⱘ䇈ˈϔϾሲ
ᗻⱘֵᙃ๲Ⲟህᰃ⬅ѢՓ⫼䖭Ͼሲᗻߚࡆḋ՟㗠ᇐ㟈ⱘᳳᳯ➉䰡ԢǄ᳈㊒⹂ഄ䆆ˈϔϾ
ሲᗻ A Ⳍᇍḋ՟䲚ড় S ⱘֵᙃ๲Ⲟ Gain(S,A)㹿ᅮНЎ
| Sv |
Gain( S , A) { Entropy ( S ) ¦
vValues ( A ) | S |
Entropy ( S v ) ˄3.4˅
݊Ё Values(A)ᰃሲᗻA᠔᳝ৃ㛑ؐⱘ䲚ড়ˈS v ᰃSЁሲᗻAⱘؐЎvⱘᄤ䲚˄гህᰃˈ
S v ={sS|A(s)=v}˅Ǆ䇋⊼ᛣˈㄝᓣ˄3.4˅ⱘ㄀ϔ乍ህᰃॳᴹ䲚ড়Sⱘ➉ˈ㄀Ѡ乍ᰃ⫼A
ߚ㉏Sৢ➉ⱘᳳᳯؐǄ䖭Ͼ㄀Ѡ乍ᦣ䗄ⱘᳳᳯ➉ህᰃ↣Ͼᄤ䲚ⱘ➉ⱘࡴᴗ੠ˈᴗؐЎሲ
| Sv |
ѢSvⱘḋ՟ऴॳྟḋ՟Sⱘ↨՟ Ǆ᠔ҹGain(S,A)ᰃ⬅Ѣⶹ䘧ሲᗻAⱘؐ㗠ᇐ㟈ⱘᳳ
|S|
ᳯ➉‫ޣ‬ᇥǄᤶহ䆱ᴹ䆆ˈGain(S,A)ᰃ⬅Ѣ㒭ᅮሲᗻAⱘؐ㗠ᕫࠄⱘ݇ѢⳂᷛߑ᭄ؐⱘֵ
ᙃǄᔧᇍSⱘϔϾӏᛣ៤ਬⱘⳂᷛؐ㓪ⷕᯊˈGain(S,A)ⱘؐᰃ೼ⶹ䘧ሲᗻAⱘؐৢৃҹ
㡖ⳕⱘѠ䖯ࠊԡ᭄Ǆ
՟བˈ‫؛‬ᅮ S ᰃϔ༫᳝݇໽⇨ⱘ䆁㒗ḋ՟ˈᦣ䗄ᅗⱘሲᗻࣙᣀৃ㛑ᰃ‫݋‬᳝ Weak ੠

Strong ϸϾؐⱘ WindǄ‫ࠡڣ‬䴶ϔḋˈ‫؛‬ᅮ S ࣙ৿ 14 Ͼḋ՟ˈ[9+ˈ5-]Ǆ೼䖭 14 Ͼḋ՟
Ёˈ‫؛‬ᅮℷ՟Ёⱘ 6 Ͼ੠ড՟Ёⱘ 2 Ͼ᳝ Wind =Weakˈ݊Ҫⱘ᳝ Wind=StrongǄ⬅Ѣᣝ
✻ሲᗻ Wind ߚ㉏ 14 Ͼḋ՟ᕫࠄⱘֵᙃ๲Ⲟৃҹ䅵ㅫབϟǄ
Values(Wind) Weak,Strong
S [9,5-]
SWeak m [6,2]
S Strong m [3,3]
| Sv |
Gain( S ,Wind ) Entropy ( S ) ¦
v{Weak , Strong } | S |
Entropy ( S v )
Entropy ( S ) (8 / 14) Entropy ( SWeak ) (6 / 14) Entropy ( S Strong )

0.940 (8 / 14)0.811 (6 / 14)1.00
0.048
ֵᙃ๲Ⲟℷᰃ ID3 ㅫ⊩๲䭓ᷥⱘ↣ϔℹЁ䗝প᳔Շሲᗻⱘᑺ䞣ᷛ‫ޚ‬Ǆ೒ 3-3 ὖ䗄њ

བԩՓ⫼ֵᙃ๲Ⲟᴹ䆘Ԅሲᗻⱘߚ㉏㛑࡯Ǆ೼䖭Ͼ೒Ёˈ䅵ㅫњϸϾϡৠሲᗻ˖⑓ᑺ
˄Humidity˅੠亢࡯˄Wind˅ⱘֵᙃ๲Ⲟˈҹ֓‫އ‬ᅮᇍѢߚ㉏㸼 3-2 ⱘ䆁㒗ḋ՟ાϔϾ
ሲᗻ᳈དǄ
Which attribute is the best classifier? ાϔϾሲᗻᰃ᳔Շⱘߚ㉏ሲᗻ˛
೒ 3-3 䅵ㅫሲᗻⱘֵᙃ๲Ⲟ
ⳌᇍѢⳂᷛߚ㉏˄े᯳ᳳ݁Ϟजᰃ৺䗖ড়ᠧ㔥⧗˅ˈHumidity ↨ Wind ᳝᳈໻ⱘֵᙃ๲ⲞǄ䖭䞠ˈE ҷ

㸼➉ˈS ҷ㸼ॳྟḋ՟䲚ড়ǄᏆⶹ߱ྟ䲚ড় S ᳝ 9 Ͼℷ՟੠ 5 Ͼ䋳՟ˈे[9+ˈ5-]Ǆ⫼ Humidity ߚ㉏䖭
ѯḋ՟ѻ⫳њᄤ䲚[3+ˈ4-]˄Humidity=High˅੠[6+ˈ1-]˄Humidity=Normal˅Ǆ䖭⾡ߚ㉏ⱘֵᙃ๲Ⲟ
Ў 0.151ˈ㗠ᇍѢሲᗻ Wind ๲ⲞҙЎ 0.048Ǆ
3.4.2 ⽰ׁ
Ўњⓨ⼎ ID3 ㅫ⊩ⱘ‫݋‬ԧ᪡԰ˈ㗗㰥㸼 3-2 ⱘ䆁㒗᭄᥂᠔ҷ㸼ⱘᄺдӏࡵǄ䖭䞠ˈ

Ⳃᷛሲᗻ PlayTennis ᇍѢϡৠⱘ᯳ᳳ݁Ϟज‫݋‬᳝ yes ੠ no ϸϾؐˈ៥Ӏᇚḍ᥂݊Ҫሲ
ᗻᴹ乘⌟䖭ϾⳂᷛሲᗻؐǄ‫ܜ‬㗗㰥䖭Ͼㅫ⊩ⱘ㄀ϔℹˈ߯ᓎ‫އ‬ㄪᷥⱘ᳔乊ッ㒧⚍Ǆાϔ
Ͼሲᗻ䆹೼ᷥϞ㄀ϔϾ㹿⌟䆩ਸ਼˛ID3 ㅫ⊩䅵ㅫ↣ϔϾ‫׭‬䗝ሲᗻ˄гህᰃ Outlookˈ
TemperatureˈHumidityˈ੠ Wind˅ⱘֵᙃ๲Ⲟˈ✊ৢ䗝ᢽֵᙃ๲Ⲟ᳔催ⱘϔϾǄ݊Ё
ϸϾሲᗻⱘֵᙃ๲Ⲟⱘ䅵ㅫᰒ⼎೼೒ 3-3 ЁǄ᠔᳝ಯϾሲᗻⱘֵᙃ๲ⲞЎ
Gain(S,Outlook)=0.246
Gain(S,Humidity)=0.151
Gain(S,Wind)=0.048
Gain(S,Temperature)=0.029
݊Ё S 㸼⼎ᴹ㞾㸼 3-2 ⱘ䆁㒗ḋ՟ⱘ䲚ড়Ǆ
㸼 3-2 Ⳃᷛὖᗉ PlayTennis ⱘ䆁㒗ḋ՟

Day Outlook Temperature Humidity Wind PlayTennis
D1 Sunny Hot High Weak No
D2 Sunny Hot High Strong No
D3 Overcast Hot High Weak Yes
D4 Rain Mild High Weak Yes
D5 Rain Cool Normal Weak Yes
D6 Rain Cool Normal Strong No
D7 Overcast Cool Normal Strong Yes
D8 Sunny Mild High Weak No
D9 Sunny Cool Normal Weak Yes
D10 Rain Mild Normal Weak Yes
D11 Sunny Mild Normal Strong Yes
D12 Overcast Mild High Strong Yes
D13 Overcast Hot Normal Weak Yes
D14 Rain Mild High Strong No
ḍ᥂ֵᙃ๲Ⲟᷛ‫ˈޚ‬ሲᗻ Outlook ೼䆁㒗ḋ՟ϞᦤկњᇍⳂᷛሲᗻ PlayTennis ⱘ᳔

ད乘⌟Ǆ᠔ҹˈOutlook 㹿䗝԰ḍ㒧⚍ⱘ‫އ‬ㄪሲᗻˈ ᑊЎᅗⱘ↣ϔϾৃ㛑ؐ˄гህᰃ Sunnyˈ
Overcast ੠ Rain˅೼ḍ㒧⚍ϟ߯ᓎߚᬃǄ䚼ߚ‫އ‬ㄪᷥⱘ㒧ᵰᰒ⼎೼೒ 3-4 Ёˈৠᯊ⬏ߎ
ⱘ䖬᳝㹿ᥦ߫ࠄ↣Ͼᮄⱘৢ㒻㒧⚍ⱘ䆁㒗ḋ՟Ǆ⊼ᛣࠄ↣ϔϾ Outlook=Overcast ⱘḋ՟
г䛑ᰃ PlayTennis ⱘℷ՟Ǆ᠔ҹˈᷥⱘ䖭Ͼ㒧⚍៤ЎϔϾ৊ᄤ㒧⚍ˈᅗᇍⳂᷛሲᗻⱘߚ
㉏ᰃ PlayTennis=YesǄⳌডˈᇍᑨ Outlook=Sunny ੠ Outlook=Rain ⱘৢ㒻㒧⚍䖬᳝䴲 0
ⱘ➉ˈ᠔ҹ‫އ‬ㄪᷥӮ೼䖭ѯ㒧⚍ϟ䖯ϔℹሩᓔǄ
ᇍѢ䴲㒜ッⱘৢ㒻㒧⚍ˈ‫ݡ‬䞡໡ࠡ䴶ⱘ䖛⿟䗝ᢽϔϾᮄⱘሲᗻᴹߚࡆ䆁㒗ḋ՟ˈ䖭
ϔ⃵ҙՓ⫼Ϣ䖭Ͼ㒧⚍݇㘨ⱘ䆁㒗ḋ՟ǄᏆ㒣㹿ᬊ㓪ܹᷥⱘ䕗催㒧⚍ⱘሲᗻ㹿ᥦ䰸೼
໪ˈҹ֓ӏԩ㒭ᅮⱘሲᗻ೼ᷥⱘӏᛣ䏃ᕘϞ᳔໮ҙߎ⦄ϔ⃵ǄᇍѢ↣ϔϾᮄⱘ৊ᄤ㒧⚍
㒻㓁䖭Ͼ䖛⿟ˈⳈࠄ⒵䎇ҹϟϸϾᴵӊЁⱘӏϔϾ˖˄1˅᠔᳝ⱘሲᗻᏆ㒣㹿䖭ᴵ䏃ᕘࣙ
ᣀˈ៪˄2˅Ϣ䖭Ͼ㒧⚍݇㘨ⱘ᠔᳝䆁㒗ḋ՟䛑‫݋‬᳝ৠḋⱘⳂᷛሲᗻؐ˄гህᰃᅗӀⱘ
➉Ў 0˅Ǆ೒ 3-4 ߫ߎњϟϔℹ๲䭓ᷥ㽕䅵ㅫⱘֵᙃ๲ⲞǄҢ㸼 3-2 ⱘ 14 Ͼ䆁㒗ḋ՟䗮
䖛 ID3 ㅫ⊩ᕫࠄⱘ᳔㒜‫އ‬ㄪᷥ㹿⬏೼೒ 3-1 ЁǄ
Which attribute should be test here? ાϔϾሲᗻᑨ೼䖭䞠㹿⌟䆩˛
೒ 3-4 ID3 ㅫ⊩㄀ϔℹৢᔶ៤ⱘ䚼ߚ‫އ‬ㄪᷥ
䆁㒗ḋ՟㹿ᥦ߫ࠄᇍᑨⱘߚᬃ㒧⚍Ǆߚᬃ Overcast ⱘ᠔᳝ḋ՟䛑ᰃℷ՟ˈ᠔ҹ៤ЎⳂᷛߚ㉏Ў Yes ⱘ

৊㒧⚍Ǆ঺ϸϾ㒧⚍ᇚ㹿䖯ϔℹሩᓔˈᮍ⊩ᰃᣝ✻ᮄⱘḋ՟ᄤ䲚䗝পֵᙃ๲Ⲟ᳔催ⱘሲᗻǄ
3.5 ߩㆌṇᆜҖѣⲺ‫ٽ‬䇴グ䰪ᩒ㍘
Ϣ݊Ҫⱘᔦ㒇ᄺдㅫ⊩ϔḋˈID3 ㅫ⊩ৃҹ㹿ᦣ䗄ЎҢϔϾ‫؛‬䆒ぎ䯈Ё᧰㋶ϔϾᢳ
ড়䆁㒗ḋ՟ⱘ‫؛‬䆒Ǆ㹿 ID3 ㅫ⊩᧰㋶ⱘ‫؛‬䆒ぎ䯈ህᰃৃ㛑ⱘ‫އ‬ㄪᷥⱘ䲚ড়ǄID3 ㅫ⊩ҹ
ϔ⾡Ңㅔऩࠄ໡ᴖⱘ⠀ቅㅫ⊩䘡ग़䖭Ͼ‫؛‬䆒ぎ䯈ˈҢぎⱘᷥᓔྟˈ✊ৢ䗤ℹ㗗㰥᳈ࡴ໡
ᴖⱘ‫؛‬䆒ˈⳂⱘᰃ᧰㋶ࠄϔϾℷ⹂ߚ㉏䆁㒗᭄᥂ⱘ‫އ‬ㄪᷥǄᓩᇐ䖭⾡⠀ቅ᧰㋶ⱘ䆘Ԅߑ
᭄ᰃֵᙃ๲Ⲟᑺ䞣Ǆ೒ 3-5 ᦣ䗄њ䖭⾡᧰㋶Ǆ
೒ 3-5 ID3 ᧰㋶ⱘ‫؛‬䆒ぎ䯈
ID3 䘡ग़ৃ㛑‫އ‬ㄪᷥⱘぎ䯈ˈҢ᳔ㅔऩⱘᷥࠄ䗤⏤໡ᴖⱘᷥǄ݊᧰㋶⬅ֵᙃ๲Ⲟਃথᓣ㾘߭ᓩᇐǄ
䗮䖛㾖ᆳ ID3 ㅫ⊩ⱘ᧰㋶ぎ䯈੠᧰㋶ㄪ⬹ˈ៥Ӏৃҹ⏅ܹ䅸䆚䖭Ͼㅫ⊩ⱘӬ࢓੠

ϡ䎇Ǆ
z ID3 ㅫ⊩Ёⱘ‫؛‬䆒ぎ䯈ࣙ৿᠔᳝ⱘ‫އ‬ㄪᷥˈᅗᰃⳌᇍѢ⦄᳝ሲᗻⱘ᳝䰤⾏ᬷؐߑ
᭄ⱘϔϾᅠ ᅠᭈぎ䯈Ǆ಴Ў↣Ͼ᳝䰤⾏ᬷؐߑ᭄ৃ㹿㸼⼎ЎᶤϾ‫އ‬ㄪᷥˈ᠔ҹ ID3
ㅫ⊩䙓‫ܡ‬њ᧰㋶ϡᅠᭈ‫؛‬䆒ぎ䯈˄՟བ䙷ѯҙ㗗㰥ড়প‫؛‬䆒ⱘᮍ⊩˅ⱘϔϾЏ
㽕亢䰽˖‫؛‬䆒ぎ䯈ৃ㛑ϡࣙ৿Ⳃᷛߑ᭄Ǆ
z ᔧ䘡ग़‫އ‬ㄪᷥぎ䯈ᯊˈID3 ҙ㓈ᡸऩϔⱘᔧࠡ‫؛‬䆒Ǆ䖭Ϣ㄀ 2 ゴ䅼䆎ⱘবൟぎ䯈
‫׭‬䗝⍜䰸ᮍ⊩ϡৠˈৢ㗙㓈ᡸњϢᔧࠡⱘ䆁㒗ḋ՟ϔ㟈ⱘ᠔᳝‫؛‬䆒ⱘ䲚ড়Ǆ಴
Ўҙ㗗㰥ऩϔⱘ‫؛‬䆒ˈID3 ㅫ⊩༅এњ㸼⼎᠔᳝ϔ㟈‫؛‬䆒᠔ᏺᴹⱘӬ࢓Ǆ՟བˈ
ᅗϡ㛑߸ᮁ᳝໮ᇥϾ݊Ҫⱘ‫އ‬ㄪᷥгᰃϢ⦄᳝ⱘ䆁㒗᭄᥂ϔ㟈ⱘˈ៪㗙Փ⫼ᮄ
ⱘᅲ՟ᶹ䆶ᴹ᳔Ӭഄऎߚ䖭ѯゲѝ‫؛‬䆒Ǆ
z ෎ᴀⱘ ID3 ㅫ⊩೼᧰㋶Ёϡ䖯㸠ಲ⒃Ǆ↣ᔧ೼ᷥⱘᶤϔሖ⃵䗝ᢽњϔϾሲᗻ䖯
㸠⌟䆩ˈᅗϡӮ‫ݡ‬ಲ⒃䞡ᮄ㗗㰥䖭Ͼ䗝ᢽǄ᠔ҹˈᅗᯧফ᮴ಲ⒃ⱘ⠀ቅ᧰㋶Ё
ᐌ㾕亢䰽ᕅડ˖ᬊᬯࠄሔ䚼᳔ӬⱘㄨḜˈԚϡᰃܼሔ᳔ӬⱘǄᇍѢ ID3 ㅫ⊩ˈ
ϔϾሔ䚼᳔ӬⱘㄨḜᇍᑨⴔᅗ೼ϔᴵ᧰㋶䏃ᕘϞ᥶㋶ᯊ䗝ᢽⱘ‫އ‬ㄪᷥǄ✊㗠ˈ
䖭Ͼሔ䚼᳔ӬⱘㄨḜৃ㛑ϡབ⊓ⴔ঺ϔᴵߚᬃ᧰㋶ࠄⱘ᳈ҸҎ⒵ᛣǄৢ䴶៥Ӏ
䅼䆎ϔϾᠽሩˈ๲ࡴϔ⾡ᔶᓣⱘಲ⒃˄ৢׂ࠾‫އ‬ㄪᷥ˅Ǆ
z ID3 ㅫ⊩೼᧰㋶ⱘ↣ϔℹ䛑Փ⫼ᔧࠡⱘ᠔᳝䆁㒗ḋ՟ˈҹ㒳䅵Ў෎⸔‫އ‬ᅮᗢḋ㊒
࣪ᔧࠡⱘ‫؛‬䆒Ǆ䖭Ϣ䙷ѯ෎Ѣऩ⣀ⱘ䆁㒗ḋ՟䗦๲԰ߎ‫އ‬ᅮⱘᮍ⊩˄՟བˈFind-S
៪‫׭‬䗝⍜䰸⊩˅ϡৠǄՓ⫼᠔᳝ḋ՟ⱘ㒳䅵ሲᗻ˄՟བˈֵᙃ๲Ⲟ˅ⱘϔϾӬ
⚍ᰃ໻໻‫ޣ‬ᇣњᇍϾ߿䆁㒗ḋ՟䫭䇃ⱘᬣᛳᗻǄ಴ℸˈ䗮䖛ׂᬍ ID3 ㅫ⊩ⱘ㒜
ℶ‫߭ޚ‬ҹ᥹ফϡᅠܼᢳড়䆁㒗᭄᥂ⱘ‫؛‬䆒ˈᅗৃҹ㹿ᕜᆍᯧഄᠽሩࠄ໘⧚৿᳝
ాໄⱘ䆁㒗᭄᥂Ǆ
3.6 ߩㆌṇᆜҖⲺᖈ㓩‫څ‬㖤
ID3 ㅫ⊩Ң㾖⌟ࠄⱘ䆁㒗᭄᥂⊯࣪ҹߚ㉏᳾㾕ᅲ՟ⱘㄪ⬹ᰃҔМਸ਼˛ᤶহ䆱䇈ˈᅗ
ⱘᔦ㒇‫أ‬㕂ᰃҔМ˛ಲᖚ㄀ 2 ゴЁˈᔦ㒇‫أ‬㕂ᰃϔ㋏߫ࠡᦤˈ䖭ѯࠡᦤϢ䆁㒗᭄᥂ϔ䍋
ⓨ㒢䆎䆕᳾ᴹᅲ՟ⱘߚ㉏Ǆ
བᵰ㒭ᅮϔϾ䆁㒗ḋ՟ⱘ䲚ড়ˈ䙷М䗮ᐌ᳝ᕜ໮‫އ‬ㄪᷥϢ䖭ѯḋ՟ϔ㟈Ǆ᠔ҹˈ㽕
ᦣ䗄 ID3 ㅫ⊩ⱘᔦ㒇‫أ‬㕂ˈᑨᡒࠄᅗҢ᠔᳝ϔ㟈ⱘ‫؛‬䆒Ё䗝ᢽϔϾⱘḍ᥂ǄID3 Ң䖭ѯ
‫އ‬ㄪᷥЁ䗝ᢽાϔϾਸ਼˛ᅗ䗝ᢽ೼Փ⫼ㅔऩࠄ໡ᴖⱘ⠀ቅㅫ⊩䘡ग़ৃ㛑ⱘᷥぎ䯈ᯊ䘛
ࠄⱘ㄀ϔϾৃ᥹ফⱘᷥǄὖ⬹ഄ䆆ˈID3 ⱘ᧰㋶ㄪ⬹Ў˄a˅Ӭ‫ܜ‬䗝ᢽ䕗ⷁⱘᷥ㗠ϡᰃ
䕗䭓ⱘˈ੠˄b˅䗝ᢽ䙷ѯֵᙃ๲Ⲟ催ⱘሲᗻ⾏ḍ㒧⚍䕗䖥ⱘᷥǄ೼ ID3 ЁՓ⫼ⱘ䗝ᢽ
ሲᗻⱘਃথᓣ㾘߭੠ᅗ䘛ࠄⱘ⡍ᅮ䆁㒗ḋ՟П䯈ᄬ೼ⴔᖂ཭ⱘⳌѦ԰⫼ˈ⬅Ѣ䖭ϔ⚍ˈ
ᕜ䲒‫⹂ޚ‬ഄࠏߦߎ ID3 ⱘᔦ㒇‫أ‬㕂Ǆ✊㗠៥Ӏৃҹ䖥Ԑഄᡞᅗⱘᔦ㒇‫أ‬㕂ᦣ䗄Ўϔ⾡
ᇍⷁⱘ‫އ‬ㄪᷥⱘ‫أ‬དǄ
䖥Ԑⱘ ID3 ㅫ⊩ᔦ㒇‫أ‬㕂˖䕗ⷁⱘᷥ↨䕗䭓ⱘӬ‫ܜ‬
џᅲϞˈ៥Ӏৃҹᛇ䈵ϔϾ㉏ԐѢ ID3 ⱘㅫ⊩ˈᅗ㊒⹂ഄ‫݋‬᳝䖭⾡ᔦ㒇‫أ‬㕂Ǆ㗗

㰥ϔ⾡ㅫ⊩ˈᅗҢϔϾぎⱘᷥᓔྟᑓᑺӬ‫˄ܜ‬breadth first˅᧰㋶䗤⏤໡ᴖⱘᷥˈ‫ܜ‬㗗㰥
᠔᳝⏅ᑺЎ 1 ⱘᷥˈ✊ৢ᠔᳝⏅ᑺЎ 2 ⱘˈĂĂǄϔᮺᅗᡒࠄњϔϾϢ䆁㒗᭄᥂ϔ㟈ⱘ
‫އ‬ㄪᷥˈᅗ䖨ಲ᧰㋶⏅ᑺⱘ᳔ᇣⱘϔ㟈ᷥ˄՟བˈ‫݋‬᳔᳝ᇥ㒧⚍ⱘᷥ˅Ǆ䅽៥Ӏ⿄䖭⾡
ᑓᑺӬ‫᧰ܜ‬㋶˄breadth first search˅ㅫ⊩Ў BFS-ID3ǄBFS-ID3 ᇏᡒ᳔ⷁⱘ‫އ‬ㄪᷥˈ಴
ℸ㊒⹂ഄ‫݋‬᳝Ā䕗ⷁⱘᷥ↨䕗䭓ⱘᕫࠄӬ‫ܜ‬āⱘ‫أ‬㕂ǄID3 ৃ㹿ⳟ԰ BFS-ID3 ⱘϔϾ᳝
ᬜ䖥ԐˈᅗՓ⫼ϔ⾡䋾ှⱘਃথᓣ᧰㋶ӕ೒থ⦄᳔ⷁⱘᷥˈ㗠ϡ⫼䖯㸠ᅠᭈⱘᑓᑺӬ‫ܜ‬
᧰㋶ᴹ䘡ग़‫؛‬䆒ぎ䯈Ǆ
಴Ў ID3 Փ⫼ֵᙃ๲Ⲟਃথᓣ㾘߭੠Ā⠀ቅāㄪ⬹ˈᅗࣙ৿↨ BFS-ID3 ᳈໡ᴖⱘ

‫أ‬㕂Ǆᇸ݊ᰃˈᅗᑊ䴲ᘏᰃᡒ᳔ⷁⱘϔ㟈ᷥˈ㗠ᰃؒ৥Ѣ䙷ѯֵᙃ๲Ⲟ催ⱘሲᗻ᳈䴴䖥
ḍ㒧⚍ⱘᷥǄ
ID3 ᔦ㒇‫أ‬㕂ⱘ᳈䌈ߛ䖥Ԑ˖䕗ⷁⱘᷥ↨䕗䭓ⱘᕫࠄӬ‫ܜ‬Ǆ䙷ѯֵᙃ๲Ⲟ催ⱘሲᗻ
᳈䴴䖥ḍ㒧⚍ⱘᷥᕫࠄӬ‫ܜ‬Ǆ
3.6.1 䲆ᇐ‫څ‬㖤ૂՎ䘿‫څ‬㖤
೼ ID3 ㅫ⊩੠㄀ 2 ゴ䅼䆎ⱘ‫׭‬䗝⍜䰸ㅫ⊩ᰒ⼎ߎⱘᔦ㒇‫أ‬㕂П䯈᳝ϔϾ᳝䍷ⱘϡ

ৠǄϟ䴶㗗㰥ϔϟ䖭ϸ⾡ᮍ⊩Ёᇍ‫؛‬䆒ぎ䯈᧰㋶ⱘᏂᓖ˖
z ID3 ⱘ᧰㋶㣗ೈᰃϔϾᇂ᮪Ⲻ‫؛‬䆒ぎ䯈˄՟བˈ㛑㸼⼎ӏԩ᳝䰤ⱘ⾏ᬷؐߑ᭄ⱘ
ぎ䯈˅ǄԚᅗуᖱᓋ൦᧰㋶䖭Ͼぎ䯈ˈҢㅔऩⱘ‫؛‬䆒ࠄ໡ᴖⱘ‫؛‬䆒ˈⳈࠄ䘛ࠄ
㒜ℶᴵӊ˄՟བˈᅗথ⦄њϔϾϢ᭄᥂ϔ㟈ⱘ‫؛‬䆒˅Ǆᅗⱘᔦ㒇‫أ‬㕂ᅠܼᰃ᧰
㋶ㄪ⬹ᥦᑣ‫؛‬䆒ⱘ㒧ᵰǄᅗⱘ‫؛‬䆒ぎ䯈≵᳝ᓩܹ乱໪ⱘ‫أ‬㕂Ǆ
z বൟぎ䯈‫׭‬䗝⍜䰸ㅫ⊩ⱘ᧰㋶㣗ೈᰃуᇂ᮪Ⲻ‫؛‬䆒ぎ䯈˄ेϔϾҙ㛑㸼⼎┰೼
ৃᬭᥜὖᗉᄤ䲚ⱘぎ䯈˅ǄԚᅗᖱᓋ൦᧰㋶䖭Ͼぎ䯈ˈᶹᡒ᠔᳝Ϣ䆁㒗᭄᥂ϔ
㟈ⱘ‫؛‬䆒Ǆᅗⱘᔦ㒇‫أ‬㕂ᅠܼᰃ‫؛‬䆒㸼⼎ⱘ㸼䖒㛑࡯ⱘ㒧ᵰǄᅗⱘ᧰㋶ㄪ⬹≵
᳝ᓩܹ乱໪ⱘ‫أ‬㕂Ǆ
ㅔऩഄ䆆ˈID3 ⱘᔦ㒇‫أ‬㕂ᴹ㞾ᅗⱘᩒ㍘ㆌ⮛ˈ㗠‫׭‬䗝⍜䰸ㅫ⊩ⱘᔦ㒇‫أ‬㕂ᴹ㞾ᅗ
ᇍᩒ㍘グ䰪ⱘᅮНǄ
ID3 ⱘᔦ㒇‫أ‬㕂ᰃᇍᶤ⾡‫؛‬䆒˄՟བˈᇍѢ䕗ⷁⱘ‫؛‬䆒˅㚰䖛݊Ҫ‫؛‬䆒ⱘϔ⾡Վ䘿
δpreferenceε
ˈᅗᇍ᳔㒜ৃ߫Вⱘ‫؛‬䆒≵᳝⹀ᗻ䰤ࠊȾ䖭⾡㉏ൟⱘ‫أ‬㕂䗮ᐌ㹿⿄ЎՎ䘿
‫څ‬㖤δpreference biasε
˄៪িᩒ㍘‫څ‬㖤δsearch biasε˅ǄⳌডˈ‫׭‬䗝⍜䰸ㅫ⊩ⱘ‫أ‬㕂ᰃᇍ
ᕙ㗗㰥‫؛‬䆒ⱘϔ⾡䲆ᇐ˄restriction˅ Ǆ䖭⾡ᔶᓣⱘ‫أ‬㕂䗮ᐌ㹿⿄Ў䲆ᇐ‫څ‬㖤˄៪㗙ি䈣
䀶‫څ‬㖤˄language bias˅ ˅Ǆ
བᵰ䳔㽕ᶤ⾡ᔶᓣⱘᔦ㒇‫أ‬㕂ᴹҢ䆁㒗᭄᥂Ё⊯࣪˄㾕㄀ 2 ゴ˅ˈ䙷М៥Ӏ䆹Ӭ‫ܜ‬
㗗㰥ા⾡ᔶᓣⱘᔦ㒇‫أ‬㕂ਸ਼˖ᰃӬ䗝‫أ‬㕂䖬ᰃ䰤ᅮ‫أ‬㕂˛
䗮ᐌˈӬ䗝‫أ‬㕂↨䰤ᅮ‫أ‬㕂᳈ড়Т䳔㽕ˈ಴Ўᅗ‫ܕ‬䆌ᄺд఼Ꮉ԰೼ᅠᭈⱘ‫؛‬䆒ぎ䯈
Ϟˈ䖭ֱ䆕њ᳾ⶹⱘⳂᷛߑ᭄㹿ࣙ৿೼‫ݙ‬ǄⳌডⱘˈ䰤ᅮ‫أ‬㕂ϹḐഄ䰤ࠊњ‫؛‬䆒䲚ড়ⱘ
┰೼ぎ䯈ˈ䗮ᐌϡᰃ៥ӀᏠᳯⱘˈ಴Ўᅗৠᯊᓩܹњᡞ᳾ⶹⱘⳂᷛߑ᭄ᥦ䰸೼໪ⱘৃ㛑
ᗻǄ
䡈Ѣ ID3 䞛⫼㒃㊍ⱘӬ䗝‫أ‬㕂㗠‫׭‬䗝⍜䰸ㅫ⊩䞛⫼㒃㊍ⱘ䰤ᅮ‫أ‬㕂ˈϔѯᄺд㋏
㒳㓐ড়њ䖭ϸ㗙Ǆ՟བˈ㗗㰥㄀ 1 ゴᦣ䗄ⱘϟẟ⿟ᑣⱘ՟ᄤǄ݊Ёˈᄺдࠄⱘ䆘Ԅߑ᭄
㹿㸼⼎Ўϔѯ೎ᅮⱘẟⲬ⡍ᕕⱘ㒓ᗻ㒘ড়ˈᄺдㅫ⊩䇗ᭈ䖭Ͼ㒓ᗻ㒘ড়ⱘখ᭄ᴹ᳔དഄ
ᢳড়⦄᳝ⱘ䆁㒗᭄᥂Ǆ䖭䞠ˈՓ⫼㒓ᗻߑ᭄ᴹ㸼⼎䆘Ԅߑ᭄ⱘ‫އ‬ᅮህᓩܹњ䰤ᅮ‫أ‬㕂˄䴲
㒓ᗻⱘ䆘Ԅߑ᭄ϡৃ㛑㹿㸼⼎៤䖭⾡ᔶᓣ˅Ǆৠᯊˈ䗝ᢽ⡍ᅮখ᭄ⱘ䇗ᭈᮍ⊩˄LMS ㅫ
⊩˅ᓩܹњϔϾӬ䗝‫أ‬㕂ˈᅗ⑤㞾᠔᳝ৃ㛑খ᭄ؐぎ䯈Ϟⱘ乎ᑣ᧰㋶Ǆ
3.6.2 ѰӶѾՎ‫ٽⲺ⸣ݾ‬䇴ϋ
ID3 ㅫ⊩ЁӬ䗝䕗ⷁ‫އ‬ㄪᷥⱘᔦ㒇‫أ‬㕂ˈᰃϡᰃҢ䆁㒗᭄᥂Ё⊯࣪ⱘৃ䴴෎⸔˛૆
ᄺᆊӀҹঞ݊Ҫᄺ㗙Ꮖ㒣ᇍ䖭ḋⱘ䯂乬ѝ䆎޴ϾϪ㑾њˈ㗠Ϩ䖭Ͼѝ䆎㟇Ҟ䖬᳾㾷‫އ‬Ǆ
࿕ᒝg༹ഢྚ໻㑺೼ 1320 ᑈᦤߎ㉏Ԑⱘ䆎⚍ķˈᰃ᳔ᮽ䅼䆎䖭Ͼ䯂乬ⱘҎПϔˈ᠔ҹ䖭
Ͼ‫أ‬㕂㒣ᐌ㹿⿄ЎĀ༹ഢྚࠗߔā˄Occam’s razor˅Ǆ
༹ഢྚࠗߔ˖Ӭ‫ܜ‬䗝ᢽᢳড়᭄᥂ⱘ᳔ㅔऩ‫؛‬䆒Ǆ
ᔧ✊㒭ߎϔϾᔦ㒇‫أ‬㕂ⱘৡᄫϡㄝѢ䆕ᯢњᅗǄЎҔМᑨ䆹Ӭ‫ܜ‬䗝ᢽ䕗ㅔऩⱘ‫؛‬䆒
ਸ਼˛䇋⊼ᛣ⾥ᄺᆊӀ᳝ᯊԐТг䙉ᕾ䖭Ͼᔦ㒇‫أ‬㕂Ǆ՟བ⠽⧚ᄺᆊӬ‫ܜ‬䗝ᢽ㸠᯳䖤ࡼㅔ
ऩⱘ㾷䞞ˈ㗠ϡ⫼໡ᴖⱘ㾷䞞ǄЎҔМ˛ϔ⾡㾷䞞ᰃⷁ‫؛‬䆒ⱘ᭄䞣ᇥѢ䭓‫؛‬䆒˄෎Ѣㅔ
ऩⱘখ᭄㒘ড়˅ˈ᠔ҹᡒࠄϔϾⷁⱘ‫؛‬䆒ԚৠᯊᅗϢ䆁㒗᭄᥂ᢳড়ⱘৃ㛑ᗻ䕗ᇣǄⳌডˈ
ᐌᐌ᳝ᕜ໮䴲ᐌ໡ᴖⱘ‫؛‬䆒ᢳড়ᔧࠡⱘ䆁㒗᭄᥂ˈԚै᮴⊩ℷ⹂ഄ⊯࣪ࠄৢᴹⱘ᭄᥂Ǆ
՟བ㗗㰥‫އ‬ㄪᷥ‫؛‬䆒Ǆ500 Ͼ㒧⚍ⱘ‫އ‬ㄪᷥ↨ 5 Ͼ㒧⚍ⱘ‫އ‬ㄪᷥ໮ᕫ໮Ǆབᵰ㒭ᅮϔϾ
20 Ͼ䆁㒗ḋ՟ⱘ䲚ড়ˈৃҹ乘ᳳ㛑໳ᡒࠄᕜ໮ 500 Ͼ㒧⚍ⱘ‫އ‬ㄪᷥϢ䆁㒗᭄᥂ϔ㟈ˈ
㗠བᵰϔϾ 5 㒧⚍ⱘ‫އ‬ㄪᷥৃҹᅠ㕢ഄᢳড়䖭ѯ᭄᥂߭ᰃߎТᛣ໪ⱘǄ᠔ҹ៥ӀӮⳌֵ
5 Ͼ㒧⚍ⱘᷥϡ໾ৃ㛑ᰃ㒳䅵Ꮋড়ˈ಴㗠Ӭ‫ܜ‬䗝ᢽ䖭Ͼ‫؛‬䆒ˈ㗠ϡ䗝ᢽ 500 Ͼ㒧⚍ⱘǄ
ḍ᥂᳈⏅ܹⱘߚᵤˈৃҹথ⦄Ϟ䴶ⱘ㾷䞞᳝ϔϾЏ㽕ⱘೄ䲒ǄЎҔМ៥Ӏϡড䯂˖
Փ⫼ৠḋⱘ᥼⧚ˈᑨ䆹Ӭ‫ܜ‬䗝ᢽࣙ৿ᙄད᳝ 17 Ͼ৊ᄤ㒧⚍੠ 11 Ͼ䴲৊ᄤ㒧⚍ⱘ‫އ‬ㄪ
ᷥ˛䖭ỉᷥ೼ḍ㒧⚍Փ⫼‫އ‬ㄪሲᗻA1ˈ✊ৢҹ᭄ᄫ乎ᑣ⌟䆩ሲᗻA2ⳈࠄA11Ǆ䖭ḋⱘ‫އ‬
ㄪᷥⳌᔧᇥˈ಴ℸ˄⫼੠Ϟ䴶ৠḋⱘ᥼⧚˅ˈᡒࠄ݊ЁПϔϢӏᛣ᭄᥂䲚ϔ㟈ⱘ‫ܜ‬偠ৃ
㛑ᗻгᕜᇣǄ䖭䞠ⱘೄ䲒೼ѢৃҹᅮНᕜ໮ᇣⱘ‫؛‬䆒䲚ড়üü݊Ёⱘ໻໮᭄Ⳍᔧ᰺⍽䲒
㾷Ǆ䙷Мˈ៥Ӏḍ᥂ҔМⳌֵ᳝ⷁᦣ䗄˄short description˅ⱘ‫އ‬ㄪᷥ㒘៤ⱘᇣ‫؛‬䆒䲚ড়
ህ↨݊Ҫӫ໮ৃᅮНⱘᇣ‫؛‬䆒䲚ড়᳈䗖ᔧਸ਼˛
Ϟ䴶ⱘ༹ഢྚࠗߔॳ߭ⱘ㾷䞞ⱘ㄀ѠϾ䲒乬ᰃˈ‫؛‬䆒ⱘ໻ᇣᰃ⬅ᄺд㗙‫ݙ‬ ‫ݙ‬䚼Փ⫼ⱘ
⡍ᅮ㸼⼎‫އ‬ᅮⱘǄ᠔ҹϸϾᄺд఼Փ⫼ϡৠⱘ‫ݙ‬䚼㸼⼎Ӯᕫࠄϡৠⱘ‫؛‬䆒ˈϸ㗙জ䛑⫼
༹ഢྚࠗߔॳ߭ᕫࠄⳌѦ⶯Ⳓⱘ㒧䆎ʽ՟བˈབᵰ៥ӀᅮНሲᗻ XYZˈᅗᇍѢ㹿೒ 3-1
ⱘ‫އ‬ㄪᷥߚ㉏Ўℷ՟ⱘᅲ՟ㄝѢⳳˈⳌডЎ‫ˈ؛‬䙷МϔϾᄺд఼ህৃҹᡞ೒ 3-1 Ё‫އ‬ㄪ
ᷥ㸼⼎ⱘߑ᭄㸼⼎Ўা᳝ϔϾ‫އ‬ㄪ㒧⚍ⱘᷥǄѢᰃˈϸϾᄺд఼བᵰϔϾՓ⫼њ XYZ
ሲᗻᦣ䗄ᅗⱘᅲ՟ˈ㗠঺ϔϾাՓ⫼ OutlookǃTemperatureǃHumidity ੠ Wind ሲᗻˈԚ
䛑ᑨ⫼༹ഢྚࠗߔॳ߭ˈ䙷М㒧ᵰᅗӀӮҹϡৠⱘᮍᓣ⊯࣪Ǆ
ҹϞ䇈ᯢˈᇍѢৠϔ༫䆁㒗ḋ՟ˈᔧϸϾᄺд఼ҹϡৠ‫ݙ‬䚼㸼⼎ᮍᓣ⧚㾷੠Փ⫼䖭
ѯḋ՟ᯊˈӮѻ⫳ϸϾϡৠⱘ‫؛‬䆒Ǆ෎Ѣ䖭ϔ⚍ˈԐТ៥Ӏᑨᅠܼᢉࠊ༹ഢྚࠗߔॳ߭Ǆ
ϡ䖛ˈ䅽៥Ӏⳟϔⳟϟ䴶䖭Ͼഎ᱃ˈᑊߚᵤાϔϾ‫ݙ‬䚼㸼⼎ӮҢ㞾✊䗝ᢽ੠䖯࣪Ё㜅乪
㗠ߎǄᛇ䈵ϔϾ⬅Ҏ䗴ⱘᄺд agent 㒘៤ⱘ㕸ԧˈ䖭Ͼ㕸ԧᰃ⬅῵ᢳⱘ䖯࣪䖛⿟ѻ⫳ⱘˈ
䖯࣪䖛⿟ࣙᣀ agent ⱘ㐕Ⅺǃবᓖ੠㞾✊䗝ᢽǄ‫؛‬ᅮ䖭Ͼ䖯࣪䖛⿟㛑໳ϔҷ᥹ϔҷഄᬍ
ব䖭ѯ agent ⱘᛳⶹ㋏㒳ˈ⬅ℸᬍবᅗӀ⫼ᴹᛳⶹϪ⬠ⱘ఼ᅬⱘ‫ݙ‬䚼ሲᗻǄߎѢ䆎䆕ⱘ
㗗㰥ˈ៥Ӏг‫؛‬ᅮ䖭ѯᄺд agent 䞛⫼ϔϾϡӮ㹿䖯࣪᠔ᬍবⱘ೎ᅮⱘㅫ⊩˄↨བ ID3˅
Ǆ
᳝⧚⬅᥼ᮁˈ䱣ⴔᯊ䯈ⱘ⌕䗱ˈ䖯࣪Ӯѻ⫳᳈དⱘ‫ݙ‬䚼㸼⼎ˈՓ agent 㛑ᛜࡴ៤ࡳഄ⫳
ᄬ೼ᅗӀⱘ⦃๗ЁǄ‫؛‬ᅮ agent ⱘ៤ࡳձ䌪Ѣᅗ㊒⹂⊯࣪ⱘ㛑࡯ˈ᠔ҹ៥Ӏৃҹᳳᳯˈ
䖯࣪ѻ⫳ⱘ‫ݙ‬䚼㸼⼎ᇍӏԩᄺдㅫ⊩੠ᔦ㒇‫أ‬㕂䛑᳝ᕜདⱘᗻ㛑ǄབᵰᶤϾ agent ⾡㕸
䞛⫼њᏺ༹᳝ഢྚࠗߔᔦ㒇‫أ‬㕂ⱘᄺдㅫ⊩ˈ䙷М៥Ӏᳳᳯ䖯࣪Ӯѻ⫳䗖ড়༹ഢྚࠗߔ
ķ
ᰒ✊ᰃ೼ࠂ㚵乏ᯊᛇࠄⱘǄ
ㄪ⬹ⱘ‫ݙ‬䚼㸼⼎Ǆ䖭Ͼ䆎⚍ⱘ㊒傧೼Ѣˈ䖯࣪ѻ⫳ⱘ‫ݙ‬䚼㸼⼎Փᕫᄺдㅫ⊩ⱘᔦ㒇‫أ‬㕂
៤Ў㞾៥ᅲ⦄ⱘ乘㿔˄self-fulfilling prophecy˅ˈা಴Ўᅗᬍব‫ݙ‬䚼㸼⼎↨ᬍবᄺдㅫ⊩
᳈ᆍᯧǄ
᱖ᯊˈ៥Ӏᬒϟ݇Ѣ༹ഢྚࠗߔⱘѝ䆎Ǆ㄀ 6 ゴ៥ӀӮ‫ᦤ⃵ݡ‬䍋䖭Ͼ䆱乬ˈ䙷䞠ᇚ
䅼䆎᳔ᇣᦣ䗄䭓ᑺ˄Minimum Description Length˅ॳ߭ˈᅗᰃ঺ϔ⠜ᴀⱘ༹ഢྚࠗߔˈ
ᅗৃ⫼䋱৊ᮃḚᶊᴹ㾷䞞Ǆ
3.7 ߩㆌṇᆜҖⲺᑮ㿷䰤从
‫އ‬ㄪᷥᄺдⱘᅲ䰙䯂乬ࣙᣀ⹂ᅮ‫އ‬ㄪᷥ๲䭓ⱘ⏅ᑺ˗໘⧚䖲㓁ؐⱘሲᗻ˗䗝ᢽϔϾ
䗖ᔧⱘሲᗻㄯ䗝ᑺ䞣ᷛ‫˗ޚ‬໘⧚ሲᗻؐϡᅠᭈⱘ䆁㒗᭄᥂˗໘⧚ϡৠҷӋⱘሲᗻ˗ҹঞ
ᦤ催䅵ㅫᬜ⥛Ǆϟ䴶៥Ӏ䅼䆎↣ϔϾ䯂乬ˈᑊ䩜ᇍ䖭ѯ䯂乬ᠽሩ෎ᴀⱘ ID3 ㅫ⊩Ǆџ
ᅲϞˈЎњ㾷‫݊އ‬Ё໮᭄ⱘ䯂乬ˈ ID3 ㅫ⊩Ꮖ㒣㹿ᠽሩњˈᠽሩৢⱘ㋏㒳㹿ᬍৡЎ C4.5
˄Quinlan 1993˅Ǆ
3.7.1 䚵‫ރ‬䗽ᓜᤕਾδOverfittingεᮦᦤ
㸼 3-1 ᦣ䗄ⱘㅫ⊩๲䭓ᷥⱘ↣ϔϾߚᬃⱘ⏅ᑺˈⳈࠄᙄད㛑ᇍ䆁㒗ḋ՟ᅠ㕢ഄߚ
㉏Ǆ✊㗠䖭Ͼㄪ⬹ᑊ䴲ᘏᰃ㸠ᕫ䗮ⱘˈџᅲϞˈᔧ᭄᥂Ё᳝ాໄˈ៪䆁㒗ḋ՟ⱘ᭄䞣໾
ᇥҹ㟇Ѣϡ㛑ѻ⫳Ⳃᷛߑ᭄ⱘ᳝ҷ㸼ᗻⱘ䞛ḋᯊˈ䖭Ͼㄪ⬹֓Ӯ䘛ࠄೄ䲒Ǆ೼ҹϞӏϔ
⾡ᚙ‫މ‬থ⫳ᯊˈ䖭Ͼㅔऩⱘㅫ⊩ѻ⫳ⱘᷥӮ䖛ᑺᢳড়䆁㒗ḋ՟Ǆ
ᇍѢϔϾ‫؛‬䆒ˈᔧᄬ೼݊Ҫⱘ‫؛‬䆒ᇍ䆁㒗ḋ՟ⱘᢳড়↨ᅗᏂˈԚџᅲϞ೼ᅲ՟ⱘᭈ
ϾߚᏗ˄гህᰃࣙ৿䆁㒗䲚ড়ҹ໪ⱘᅲ՟˅Ϟ㸼⦄ⱘै᳈དᯊˈ៥Ӏ䇈䖭Ͼ‫؛‬䆒䖛ᑺᢳ
ড়˄overfit˅䆁㒗ḋ՟Ǆ
定义：㒭ᅮϔϾ‫؛‬䆒ぎ䯈 HˈϔϾ‫؛‬䆒 hHˈབᵰᄬ೼݊Ҫⱘ‫؛‬䆒 h´HˈՓ

ᕫ೼䆁㒗ḋ՟Ϟ h ⱘ䫭䇃⥛↨ h´ᇣˈԚ೼ᭈϾᅲ՟ߚᏗϞ h´ⱘ䫭䇃⥛↨ h ᇣˈ䙷М
ህ䇈‫؛‬䆒 h 䖛ᑺᢳড়˄overfit˅䆁㒗᭄᥂Ǆ
೒ 3-6 ⬏ߎњ೼‫އ‬ㄪᷥᄺдⱘϔϾ‫݌‬ൟᑨ⫼Ё䖛ᑺᢳড়ⱘᕅડǄ೼䖭Ͼ՟ᄤЁˈID3
ㅫ⊩⫼ᴹᄺдાϔϾ⮙Ҏᙷ᳝ᶤ⾡㊪ሓ⮙Ǆ䖭ᐙ೒ⱘ῾䕈㸼⼎೼‫އ‬ㄪᷥ߯ᓎ䖛⿟Ёᷥⱘ
㒧⚍ᘏ᭄ˈ㒉䕈㸼⼎‫އ‬ㄪᷥ԰ߎⱘ乘⌟ⱘ㊒ᑺǄᅲ㒓ᰒ⼎‫އ‬ㄪᷥ೼䆁㒗ḋ՟Ϟⱘ㊒ᑺˈ
㰮㒓ᰒ⼎೼ϔ༫⣀ゟⱘ⌟䆩ḋ՟˄≵᳝㹿ࣙᣀ೼䆁㒗ḋ՟Ё˅Ϟ⌟䞣ߎⱘ㊒ᑺǄৃҹⳟ
ߎˈ䱣ⴔᷥⱘ๲䭓ˈ೼䆁㒗ḋ՟Ϟⱘ㊒ᑺᰃऩ䇗ϞछⱘǄ✊㗠ˈ೼⣀ゟⱘ⌟䆩ḋ՟Ϟ⌟
ߎⱘ㊒ᑺ‫ܜ‬Ϟछৢϟ䰡Ǆབ೒᠔⼎ˈᔧᷥ䍙䖛໻㑺 25 Ͼ㒧⚍ᯊˈᇍᷥ䖯ϔℹ㊒㒚࣪ሑ
ㅵৃҹᦤ催ᅗ೼䆁㒗᭄᥂Ϟⱘ㊒ᑺˈै䰡Ԣњᅗ೼⌟䆩ḋ՟Ϟⱘ㊒ᑺǄ
Accuracy – ㊒ᑺ
Size of tree(number of nodes) – ᷥⱘ㾘῵˄㒧⚍᭄˅
On training data –೼䆁㒗᭄᥂Ϟ

On test data – ೼⌟䆩᭄᥂Ϟ
೒ 3-6 ‫އ‬ㄪᷥᄺдЁⱘ䖛ᑺᢳড়Ǆ
䱣ⴔ ID3 ㅫ⊩๲ࡴᮄⱘ㒧⚍๲䭓‫އ‬ㄪᷥˈ೼䆁㒗ḋ՟Ϟⱘ㊒ᑺᰃऩ䇗ϞछⱘǄ✊㗠ˈ೼⣀ゟѢ䆁㒗ḋ
՟ⱘ⌟䆩ḋ՟Ϟˈ㊒ᑺ‫ܜ‬Ϟछˈ✊ৢϟ䰡Ǆᅲ偠䖭Ͼ೒᠔䳔ⱘ䕃ӊ੠᭄᥂ৃҹ䗮䖛㔥ഔ
http://www.cs.cmu.edu/~tom/mlbook.html ᕫࠄǄ
ᰃҔМॳ಴ᇐ㟈 h ↨ hc᳈དഄᢳড়䆁㒗ḋ՟ˈԚᇍѢৢᴹⱘᅲ՟ै㸼⦄᳈Ꮒਸ਼˛䖭
⾡ᚙ‫މ‬থ⫳ⱘϔ⾡ৃ㛑ॳ಴ᰃ䆁㒗ḋ՟৿᳝䱣ᴎ䫭䇃៪ాໄǄВ՟䇈ᯢˈ㗗㰥೼㸼 3-2
ⱘᴀᴹℷ⹂ⱘḋ՟Ёࡴܹϔᴵ䆁㒗ℷ՟ˈԚै㹿䇃ᷛ⼎Ўড՟ˈབϟ˖
<Outlook=SunnyˈTemperature=HotˈHumidity=NormalˈWind=StrongˈPlayTennis=No>
ᇍѢᴀᴹ≵᳝䫭䇃ⱘ᭄᥂ˈID3 ⫳៤೒ 3-1 㸼⼎ⱘ‫އ‬ㄪᷥǄ✊㗠ˈ๲ࡴ䖭Ͼϡℷ⹂

ⱘḋ՟ᇐ㟈 ID3 ᓎゟϔϾ᳈໡ᴖⱘᷥǄ⹂ߛഄ䆆ˈᮄⱘḋ՟Ӯ㹿ᥦ߫ࠄ೒ 3-1 㸼⼎ⱘᷥ
ⱘᎺ䍋㄀ѠϾ৊ᄤ㒧⚍ˈϢҹࠡⱘℷ՟ D9 ੠ D11 ᥦ೼ϔ䍋Ǆ಴Ўᮄⱘḋ՟㹿ᷛ䆄Ўড
՟ˈ᠔ҹ ID3 Ӯ೼䖭Ͼ㒧⚍ϟ䴶䖯ϔℹ᧰㋶᳈໮ⱘ㒚㡖Ǆᔧ✊া㽕ᮄⱘ䫭䇃ḋ՟Ϣॳ
ᴹ䖭Ͼ㒧⚍ⱘϸϾḋ՟᳝ӏԩᏂᓖˈID3 Ӯ៤ࡳᡒࠄϔϾᮄⱘ‫އ‬ㄪሲᗻᴹᡞᮄⱘḋ՟Ң
ҹࠡⱘϸϾℷ՟ЁߚᓔǄ䖭ḋⱘ㒧ᵰᰃ ID3 Ӯ䕧ߎϔϾ‫އ‬ㄪᷥ(h)ˈᅗ↨೒ 3-1 Ёॳᴹ
ⱘᷥ˄h嘔˅᳈໡ᴖǄᔧ✊ˈh Ӯᅠ㕢ഄᢳড়䆁㒗ḋ՟䲚ˈ㗠䕗ㅔऩⱘ h嘔ϡӮǄ✊㗠ˈ⬅
Ѣᮄⱘ‫އ‬ㄪ㒧⚍াᰃᢳড়䆁㒗ḋ՟Ёాໄⱘ㒧ᵰˈ៥Ӏৃҹᮁᅮ೼প㞾ৠϔᅲ՟ߚᏗⱘ
ৢ㓁᭄᥂Ϟˈh嘔Ӯ㚰䖛 hǄ
Ϟ䴶ⱘ՟ᄤⓨ⼎њ䆁㒗ḋ՟Ёⱘ䱣ᴎాໄབԩᇐ㟈䖛ᑺᢳড়ǄџᅲϞˈᔧ䆁㒗᭄᥂
≵᳝ాໄᯊˈ䖛ᑺᢳড়г᳝ৃ㛑থ⫳ˈ⡍߿ᰃᔧᇥ䞣ⱘḋ՟㹿݇㘨ࠄ৊ᄤ㒧⚍ᯊǄ䖭⾡
ᚙ‫މ‬ϟˈᕜৃ㛑ߎ⦄Ꮋড়ⱘ㾘ᕟᗻˈՓᕫϔѯሲᗻᙄᎻৃҹᕜདഄߚࡆḋ՟ˈԚैϢᅲ
䰙ⱘⳂᷛߑ᭄ᑊ᮴݇㋏Ǆϔᮺ䖭ḋⱘᎻড়ⱘ㾘ᕟᗻᄬ೼ˈህ᳝䖛ᑺᢳড়ⱘ亢䰽Ǆ
䖛ᑺᢳড়ᇍѢ‫އ‬ㄪᷥᄺд੠݊Ҫᕜ໮ᄺдㅫ⊩ᰃϔϾ䞡㽕ⱘᅲ䏉ೄ䲒Ǆ՟བˈ೼ϔ
⃵݇Ѣ ID3 ㅫ⊩ⱘᅲ偠ⷨおЁ˄Mingers 1989b˅
ˈᇍѢ 5 ⾡ᏺ᳝ాໄ੠ϡ⹂ᅮ᭄᥂ⱘϡ
ৠᄺдӏࡵˈҎӀথ⦄೼໮᭄䯂乬Ё䖛ᑺᢳড়Փ‫އ‬ㄪᷥⱘ㊒ᑺ䰡Ԣњ 10-25%Ǆ
᳝޴⾡䗨ᕘ⫼ᴹ䙓‫އܡ‬ㄪᷥᄺдЁⱘ䖛ᑺᢳড়ǄᅗӀৃ㹿ߚЎϸ㉏˖
z ঞᮽ‫ذ‬ℶ๲䭓ᷥ⊩ˈ೼ ID3 ㅫ⊩ᅠ㕢ߚ㉏䆁㒗᭄᥂Пࠡ‫ذ‬ℶ๲䭓ᷥ˗
z ৢׂ࠾⊩˄post-prune˅ˈे‫ܕ‬䆌ᷥ䖛ᑺᢳড়᭄᥂ˈ✊ৢᇍ䖭Ͼᷥৢׂ࠾Ǆ
ሑㅵ㄀ϔ⾡ᮍ⊩ৃ㛑ⳟ䍋ᴹ᳈Ⳉ᥹ˈԚᰃᇍ䖛ᑺᢳড়ⱘᷥ䖯㸠ৢׂ࠾ⱘ㄀Ѡ⾡ᮍ⊩
㹿䆕ᯢ೼ᅲ䏉Ё᳈៤ࡳǄ䖭ᰃ಴Ў೼㄀ϔ⾡ᮍ⊩Ё㊒⹂ഄԄ䅵ԩᯊ‫ذ‬ℶ๲䭓ᷥᕜೄ䲒Ǆ
᮴䆎ᰃ䗮䖛ঞᮽ‫ذ‬ℶ䖬ᰃৢׂ࠾ᴹᕫࠄℷ⹂໻ᇣⱘᷥˈϔϾ݇䬂ⱘ䯂乬ᰃՓ⫼ҔМ
ḋⱘ‫߭ޚ‬ᴹ⹂ᅮ᳔㒜ℷ⹂ᷥⱘ໻ᇣǄ㾷‫އ‬䖭Ͼ䯂乬ⱘᮍ⊩ࣙᣀ˖
z Փ⫼Ϣ䆁㒗ḋ՟៾✊ϡৠⱘϔ༫ߚ⾏ⱘḋ՟ˈᴹ䆘Ԅ䗮䖛ৢׂ࠾ᮍ⊩ҢᷥϞׂ
࠾㒧⚍ⱘᬜ⫼Ǆ
z Փ⫼᠔᳝ৃ⫼᭄᥂䖯㸠䆁㒗ˈԚ䖯㸠㒳䅵⌟䆩ᴹԄ䅵ᠽሩ˄៪ׂ࠾˅ϔϾ⡍ᅮ
ⱘ㒧⚍ᰃ৺᳝ৃ㛑ᬍ୘೼䆁㒗䲚ড়໪ⱘᅲ՟Ϟⱘᗻ㛑Ǆ՟བˈQuinlan ˄1986˅
Փ⫼ϔ⾡वᮍ˄chi-square˅⌟䆩ᴹԄ䅵䖯ϔℹᠽሩ㒧⚍ᰃ৺㛑ᬍ୘೼ᭈϾᅲ՟
ߚᏗϞⱘᗻ㛑ˈ䖬ᰃҙҙᬍ୘њ೼ᔧࠡⱘ䆁㒗᭄᥂Ϟⱘᗻ㛑Ǆ
z Փ⫼ϔϾᯢ⹂ⱘᷛ‫ޚ‬ᴹ㸵䞣䆁㒗ḋ՟੠‫އ‬ㄪᷥ㓪ⷕⱘ໡ᴖᑺˈᔧ䖭Ͼ㓪ⷕⱘ䭓
ᑺ᳔ᇣᯊ‫ذ‬ℶ๲䭓ᷥǄ䖭Ͼᮍ⊩෎Ѣϔ⾡ਃথᓣ㾘߭ˈ㹿⿄Ў᳔ᇣᦣ䗄䭓ᑺ
˄Minimum Description Length˅ⱘ‫ˈ߭ޚ‬៥Ӏᇚ೼㄀ 6 ゴЁ䅼䆎䖭⾡ᮍ⊩Ǆ
Quinlan & Rivest˄1989˅੠ Mehta et al.˄1995˅г䅼䆎њ䖭⾡ᮍ⊩Ǆ
Ϟ䴶ⱘ㄀ϔ⾡ᮍ⊩ᰃ᳔᱂䗮ⱘˈᅗᐌ㹿⿄Ў䇣㓹ૂ僂䇷䳼˄training and validation

set˅⊩Ǆϟ䴶៥Ӏ䅼䆎䖭⾡ᮍ⊩ⱘϸϾЏ㽕ব⾡Ǆ䖭⾡ᮍ⊩Ёˈৃ⫼ⱘ᭄᥂㹿ߚ៤ϸ
Ͼḋ՟䲚ড়˖ϔϾ䆁㒗䲚ড়⫼ᴹᔶ៤ᄺдࠄⱘ‫؛‬䆒ˈϔϾߚ⾏ⱘ偠䆕䲚ড়⫼ᴹ䆘Ԅ䖭Ͼ
‫؛‬䆒೼ৢ㓁᭄᥂Ϟⱘ㊒ᑺˈ⹂ߛഄ䇈ᰃ⫼ᴹ䆘Ԅׂ࠾䖭Ͼ‫؛‬䆒ⱘᕅડǄ䖭Ͼᮍ⊩ⱘࡼᴎ
ᰃ˖ेՓᄺд఼ৃ㛑Ӯ㹿䆁㒗䲚ড়Ёⱘ䱣ᴎ䫭䇃੠Ꮋড়㾘ᕟᗻ᠔䇃ᇐˈԚ偠䆕䲚ড়ϡ໻
ৃ㛑㸼⦄ߎৠḋⱘ䱣ᴎ⊶ࡼǄ᠔ҹˈ偠䆕䲚ড়ৃҹ⫼ᴹᇍ䖛ᑺᢳড়䆁㒗䲚Ёⱘ㰮‫؛‬⡍ᕕ
ᦤկϔϾ䰆ᡸẔ偠Ǆᔧ✊ˈᕜ䞡㽕ⱘϔ⚍ˈ偠䆕䲚ড়ᑨ䆹䎇໳໻ˈҹ֓ᅗᴀ䑿ৃᦤկ‫݋‬
᳝㒳䅵ᛣНⱘᅲ՟ḋᴀǄϔ⾡ᐌ㾕ⱘ‫⊩خ‬ᰃপߎৃ⫼ḋ՟ⱘϝߚПϔ⫼԰偠䆕䲚ড়ˈՓ
⫼঺໪ϝߚПѠ⫼԰䆁㒗䲚ড়Ǆ
3.7.1.1 䭏䈥⦽䲃ք‫ࢠؤ‬
Փ⫼偠䆕䲚ড়ᴹ䰆ℶ䖛ᑺᢳড়ⱘ⹂ߛᮍ⊩ᰃҔМ˛ϔ⾡⿄ЎĀ䫭䇃⥛䰡Ԣׂ࠾
˄error-reduced pruning˅āⱘᮍ⊩˄Quinlan 1987˅ᰃ㗗㰥ᇚᷥϞⱘ↣ϔϾ㒧⚍԰Ўׂ࠾
ⱘ‫׭‬䗝ᇍ䈵Ǆׂ࠾ϔϾ㒧⚍⬅ҹϟℹ偸㒘៤˖ߴ䰸ҹℸ㒧⚍Ўḍⱘᄤᷥ˗Փᅗ៤Ў৊ᄤ
㒧⚍˗ᡞ੠䆹㒧⚍݇㘨ⱘ䆁㒗ḋ՟ⱘ᳔ᐌ㾕ߚ㉏䌟㒭ᅗǄҙᔧׂ࠾ৢⱘᷥᇍѢ偠䆕䲚ড়
ⱘᗻ㛑ϡᏂѢॳᴹⱘᷥᯊᠡߴ䰸䆹㒧⚍Ǆ䖭ḋ֓Փ಴Ў䆁㒗䲚ড়ⱘᎻড়㾘ᕟᗻ㗠ࡴܹⱘ
㒧⚍ᕜৃ㛑㹿ߴ䰸ˈ಴ЎৠḋⱘᎻড়ϡ໻Ӯথ⫳೼偠䆕䲚ড়ЁǄড໡ഄׂ࠾㒧⚍ˈ↣⃵
ᘏᰃ䗝পᅗⱘߴ䰸ৃҹ᳔໻ᦤ催‫އ‬ㄪᷥ೼偠䆕䲚ড়Ϟⱘ㊒ᑺⱘ㒧⚍Ǆ㒻㓁ׂ࠾㒧⚍Ⳉࠄ
䖯ϔℹⱘׂ࠾ᰃ᳝ᆇⱘ˄гህᰃ䰡Ԣњ೼偠䆕䲚ড়Ϟⱘ㊒ᑺ˅Ǆ
Ā䫭䇃⥛䰡Ԣׂ࠾āᇍ‫އ‬ㄪᷥ㊒ᑺⱘᕅડ㹿⬏೼೒ 3-7 ЁǄ੠೒ 3-6 ϔḋˈ೒ 3-7

ᰒ⼎њ೼䆁㒗ḋ՟੠⌟䆩ḋ՟Ϟⱘ‫އ‬ㄪᷥ㊒ᑺǄ೒ 3-7 Ё঺໪ϔᴵ㒓ᰒ⼎ⱘᰃ䱣ⴔᷥⱘ
ׂ࠾ˈᅗ೼⌟䆩ḋ՟Ϟⱘ㊒ᑺব࣪Ǆᔧׂ࠾ᓔྟᯊˈᷥⱘ㾘῵᳔໻ˈᑊϨᅗ೼⌟䆩ḋ՟
Ϟⱘ㊒ᑺ᳔ᇣǄ䱣ⴔׂ࠾ⱘ䖯㸠ˈ㒧⚍ⱘ᭄䞣ϟ䰡ˈԚ೼⌟䆩䲚ড়Ϟⱘ㊒ᑺϞछǄ䖭䞠ˈ
ৃկՓ⫼ⱘ᭄᥂Ꮖ㒣㹿ߚ៤ 3 Ͼᄤ䲚˖䆁㒗ḋ՟ǃկׂ࠾ᷥ⫼ⱘ偠䆕ḋ՟੠ϔϾ⌟䆩ḋ
՟䲚ড়Ǆ⌟䆩ḋ՟⫼ᴹᦤկ೼᳾ᴹⱘ᳾㾕ᅲ՟Ϟⱘ㊒ᑺⱘ᮴‫أ‬Ԅ䅵Ǆ೒Ёᰒ⼎њ೼䆁㒗
䲚੠⌟䆩䲚Ϟⱘ㊒ᑺǄ೼⫼԰ׂ࠾ⱘ偠䆕䲚ড়Ϟⱘ㊒ᑺ≵᳝⬏ߎᴹǄ
བᵰ᳝໻䞣ⱘ᭄᥂ৃկՓ⫼ˈ䙷МՓ⫼ߚ⾏ⱘ᭄᥂䲚ড়ᴹᓩᇐׂ࠾ᰃϔϾ᳝ᬜⱘᮍ
⊩Ǆ䖭Ͼᮍ⊩ⱘЏ㽕㔎⚍ᰃᔧ᭄᥂᳝䰤ᯊˈҢЁֱ⬭ϔ䚼ߚ⫼԰偠䆕䲚ড়䖯ϔℹ‫ޣ‬ᇥњ
䆁㒗ৃҹՓ⫼ⱘḋ՟Ǆϟϔ㡖㒭ߎњ঺ϔ⾡ׂ࠾ᮍ⊩ˈ೼᭄᥂᳝䰤ⱘ䆌໮ᅲ䰙ᚙᔶϟˈ
䖭⾡ᮍ⊩ᕜ᳝ᬜǄҎӀ䖬ᦤߎњ䆌໮݊ҪⱘᡔᴃǄ՟བˈҹϡৠⱘᮍᓣ໮⃵ߚࡆৃկՓ
⫼ⱘ᭄᥂ˈ✊ৢᑇഛᕫࠄⱘ㒧ᵰǄMingers˄1989b˅੠ Malerba et al.˄1995˅Ё᡹ਞњ
ᇍϡৠׂᷥ࠾ᮍ⊩ⱘ㒣偠䆘ԄǄ
插图——原书页码：Page 70
Accuracy – ㊒ᑺ
Size of tree (number of nodes)- ᷥⱘ㾘῵˄㒧⚍᭄䞣˅
On training data-೼䆁㒗᭄᥂Ϟ
On test data-೼⌟䆩᭄᥂Ϟ
On test data(During pruning)- ೼⌟䆩᭄᥂Ϟ˄ׂ࠾ᳳ䯈˅
೒ 3-7 ‫އ‬ㄪᷥᄺдЁ䫭䇃⥛䰡Ԣׂ࠾ⱘᬜᵰ
䖭ᐙ೒ᰒ⼎њϢ೒ 3-6 ৠḋⱘ೼䆁㒗䲚੠⌟䆩䲚Ϟⱘ㊒ᑺ᳆㒓Ǆ঺໪ˈᅗᰒ⼎њĀ䫭䇃⥛䰡Ԣׂ࠾ā

ᇍ ID3 ㅫ⊩ѻ⫳ⱘᷥⱘᕅડǄ⊼ᛣ䱣ⴔᷥ㒧⚍ⱘ࠾䰸ˈ‫އ‬ㄪᷥ೼⌟䆩䲚ড়Ϟⱘ㊒ᑺϞछǄ䖭䞠ˈկׂ
࠾⫼ⱘ偠䆕䲚ড়Ϣ䆁㒗੠⌟䆩䲚ড়䛑ᰃᅠܼϡৠⱘǄ
3.7.1.2 㿺ࡏ੄‫ࢠؤ‬
ᅲ䏉Ёˈϔ⾡⫼ᴹথ⦄催㊒ᑺ‫؛‬䆒ⱘ䴲ᐌ៤ࡳⱘᮍ⊩ЎĀ㾘߭ৢׂ࠾˄rule
āǄ䖭⾡ׂ࠾ᮍ⊩ⱘϔϾবԧ㹿⫼೼ C4.5 Ё˄Quinlan 1993˅ˈC4.5 ᰃҢॳ
post-pruning˅
ྟⱘ ID3 ㅫ⊩ⱘ⌒⫳ߎᴹⱘǄ㾘߭ৢׂ࠾ࣙᣀϟ䴶ⱘℹ偸˖
1. Ң䆁㒗䲚ড়᥼ᇐߎ‫އ‬ㄪᷥˈ๲䭓‫އ‬ㄪᷥⳈࠄሑৃ㛑དഄᢳড়䆁㒗᭄᥂ˈ‫ܕ‬䆌䖛ᑺᢳ
ড়থ⫳Ǆ
2. ᇚ‫އ‬ㄪᷥ䕀࣪ЎㄝӋⱘ㾘߭䲚ড়ˈᮍ⊩ᰃЎҢḍ㒧⚍ࠄ৊ᄤ㒧⚍ⱘ↣ϔᴵ䏃ᕘ߯ᓎ
ϔᴵ㾘߭Ǆ
3. 䗮䖛ߴ䰸ӏԩ㛑ᇐ㟈Ԅ䅵㊒ᑺᦤ催ⱘࠡӊ˄preconditions˅ᴹׂ࠾˄⊯࣪˅↣ϔᴵ㾘
߭Ǆ
4. ᣝ✻ׂ࠾䖛ⱘ㾘߭ⱘԄ䅵㊒ᑺᇍᅗӀ䖯㸠ᥦᑣ˗ᑊᣝ䖭ḋⱘ乎ᑣᑨ⫼䖭ѯ㾘߭ᴹߚ
㉏ৢᴹⱘᅲ՟Ǆ
Ўњⓨ⼎ҹϞ䖛⿟ˈ‫⃵ݡ‬㗗㰥೒ 3-1 Ёⱘ‫އ‬ㄪᷥǄ೼Ā㾘߭ৢׂ࠾āㅫ⊩ЁˈЎᷥЁⱘ

↣Ͼ৊ᄤ㒧⚍ѻ⫳ϔᴵ㾘߭ǄҢḍ㒧⚍ࠄ৊ᄤ㒧⚍䏃ᕘϞⱘ↣ϔϾሲᗻ⌟䆩៤ЎϔϾ㾘߭‫ܜ‬
㸠䆡˄ेࠡӊ˅ˈ৊ᄤ㒧⚍ⱘߚ㉏⿄Ў㾘߭ⱘ㒧䆎˄ेৢӊ˅ Ǆ՟བˈ೒ 3-1 Ёᷥⱘ᳔Ꮊϔᴵ䏃
ᕘ㹿䕀ᤶ៤㾘߭˖
IF ˄Outlook=Sunny˅˄Humidity=High˅
THEN PlayTennis=No
᥹ϟᴹˈ䗮䖛ߴ䰸ϡӮ䰡ԢԄ䅵㊒ᑺⱘ‫ܜ‬㸠䆡ᴹׂ࠾↣ϔϾ㾘߭Ǆ՟བᇍѢϞ䴶ⱘ
㾘߭ˈ㾘߭ৢׂ࠾ㅫ⊩Ӯ㗗㰥ߴ䰸‫ܜ‬㸠䆡˄Outlook=Sunny˅੠˄Humidity=High˅ǄᅗӮ
䗝ᢽ䖭ѯׂ࠾ℹ偸ЁՓԄ䅵㊒ᑺ᳔᳝໻ᦤछⱘℹ偸ˈ✊ৢ㗗㰥ׂ࠾㄀ѠϾࠡӊ԰Ў䖯ϔ
ℹⱘׂ࠾ℹ偸ǄབᵰᶤϾׂ࠾ℹ偸䰡ԢњԄ䅵㊒ᑺˈ䙷М䖭Ͼℹ偸ϡӮ㹿ᠻ㸠Ǆ
བৠࠡ䴶ᦤߎⱘˈԄ䅵㾘߭㊒ᑺⱘϔ⾡ᮍ⊩ᰃՓ⫼Ϣ䆁㒗䲚੠ϡⳌѸⱘ偠䆕䲚ড়Ǆ
঺ϔ⾡㹿 C4.5 Փ⫼ⱘᮍ⊩ᰃ෎Ѣ䆁㒗䲚ড়ᴀ䑿䆘Ԅᗻ㛑ˈԚՓ⫼ϔ⾡ֱᅜԄ䅵
˄pessimistic estimate˅ᴹᓹ㸹䆁㒗᭄᥂᳝߽Ѣᔧࠡ㾘߭ⱘԄ䅵‫أ‬㕂Ǆ᳈‫⹂ޚ‬ഄ䆆ˈC4.5
䗮䖛ҹϟᮍ⊩䅵ㅫֱᅜԄ䅵ˈ‫ܜ‬䅵ㅫ㾘߭೼ᅗᑨ⫼ⱘ䆁㒗ḋ՟Ϟⱘ㊒ᑺˈ✊ৢ‫؛‬ᅮℸԄ
䅵㊒ᑺЎѠ乍ߚᏗˈᑊ䅵ㅫᅗⱘᷛ‫ޚ‬Ꮒ˄standard deviation˅ǄᇍѢϔϾ㒭ᅮⱘ㕂ֵऎ䯈ˈ
䞛⫼ϟ⬠Ԅ䅵԰Ў㾘߭ᗻ㛑ⱘᑺ䞣˄՟བˈᇍѢϔϾ 95%ⱘ㕂ֵऎ䯈ˈ㾘߭㊒ᑺ㹿ֱ
ᅜԄ䅵Ў˖೼䆁㒗䲚ড়Ϟⱘ㾖ᆳ㊒ᑺ‫ޣ‬এ 1.96 ЬԄ䅵ⱘᷛ‫ޚ‬Ꮒ˅Ǆ䖭ḋ‫ⱘخ‬ᬜᵰᰃˈᇍ
Ѣ໻ⱘ᭄᥂䲚ˈֱᅜ乘⌟䴲ᐌ᥹䖥㾖ᆳ㊒ᑺ˄гህᰃᷛ‫ޚ‬Ꮒ䴲ᐌᇣ˅ˈ✊㗠䱣ⴔ᭄᥂䲚
ড়ⱘ‫ޣ‬ᇣˈᅗᓔྟ⾏㾖ᆳ㊒ᑺ䍞ᴹ䍞䖰Ǆ㱑✊䖭⾡ਃথᓣᮍ⊩ϡᰃ㒳䅵᳝ᬜ˄statistically
valid˅ⱘˈԚᰃᏆ㒣থ⦄ᅗ೼ᅲ䏉Ёᰃ᳝⫼ⱘǄ㄀ 5 ゴ䅼䆎њ㒳䅵᳝ᬜⱘ乘⌟ഛؐ੠㕂
ֵऎ䯈ⱘᮍ⊩Ǆ
ЎҔМׂ࠾Пࠡ㽕ᡞ‫އ‬ㄪᷥ䕀࣪៤㾘߭䲚ਸ਼˛䖭ḋ‫خ‬Џ㽕᳝ϝϾད໘˖
z 䕀࣪Ў㾘߭䲚ৃҹऎߚ‫އ‬ㄪ㒧⚍Փ⫼ⱘϡৠϞϟ᭛Ǆ಴Ў䌃こ‫އ‬ㄪ㒧⚍ⱘ↣ᴵ
ϡৠ䏃ᕘѻ⫳ϔᴵϡৠⱘ㾘߭ˈ᠔ҹᇍѢϡৠ䏃ᕘˈ݇ѢϔϾሲᗻ⌟䆩ⱘׂ࠾
‫އ‬ㄪৃҹϡৠǄⳌডˈབᵰⳈ᥹ׂ࠾ᷥᴀ䑿ˈা᳝ϸϾ䗝ᢽˈ㽕Мᅠܼߴ䰸‫އ‬
ㄪ㒧⚍ˈ㽕Мֱ⬭ᅗⱘᴀᴹ⢊ᗕǄ
z 䕀࣪Ў㾘߭䲚⍜䰸њḍ㒧⚍䰘䖥ⱘሲᗻ⌟䆩੠৊㒧⚍䰘䖥ⱘሲᗻ⌟䆩ⱘऎ߿Ǆ
Ѣᰃ䙓‫ܡ‬њ䳊хⱘ䆄ᔩ䯂乬ˈ↨བ㢹ᰃḍ㒧⚍㹿ׂ࠾њԚֱ⬭ᅗϟ䴶ⱘ䚼ߚᄤ
ᷥᯊབԩ䞡ᮄ㒘㒛䖭ỉᷥǄ
z 䕀࣪Ў㾘߭ᦤ催њৃ䇏ᗻǄᇍѢҎᴹ䇈㾘߭ᘏᰃ᳈ᆍᯧ⧚㾷ⱘǄ
3.7.2 ਾᒬ䘔㔣‫ٲ‬ኔᙝ
៥Ӏ᳔߱ⱘID3 ᅮН䰤ࠊЎপ⾏ᬷؐⱘሲᗻǄ佪‫ˈܜ‬ᄺдࠄⱘ‫އ‬ㄪᷥ㽕乘⌟ⱘⳂᷛ
ሲᗻᖙ乏ᰃ⾏ᬷⱘǄ݊⃵ˈᷥⱘ‫އ‬ㄪ㒧⚍ⱘሲᗻгᖙ乏ᰃ⾏ᬷⱘǄৃҹㅔऩഄߴ䰸㄀Ѡ
Ͼ䰤ࠊˈҹ֓ᡞ䖲㓁ؐⱘ‫އ‬ㄪሲᗻࡴܹࠄ‫އ‬ㄪᷥЁǄ䖭ৃҹ䗮䖛ࡼᗕഄᅮНᮄⱘ⾏ᬷؐ
ሲᗻᴹᅲ⦄ˈे‫ܜ‬ᡞ䖲㓁ؐሲᗻⱘؐඳߚࡆЎ⾏ᬷⱘऎ䯈䲚ড়Ǆ՟བˈᇍѢ䖲㓁ؐⱘሲ
ᗻAˈㅫ⊩ৃࡼᗕഄ߯ᓎϔϾᮄⱘᏗᇨሲᗻAcˈབᵰA<cˈ䙷МЎAcⳳˈ৺߭Ў‫؛‬Ǆଃ
ϔⱘ䯂乬ᰃབԩ䗝প᳔Շⱘ䯜ؐcǄ
В՟ᴹ䇈ˈ‫؛‬ᅮ៥ӀᏠᳯ೼㸼 3-2 ⱘᄺдӏࡵЁࣙ৿䖲㓁ؐⱘሲᗻ Temperature ᴹ

ᦣ䗄䆁㒗ḋ՟ǄᇍѢϢ‫އ‬ㄪᷥⱘ⡍ᅮ㒧⚍݇㘨ⱘ䆁㒗ḋ՟ˈ䖯ϔℹ‫؛‬ᅮ݊ሲᗻ
Temperature ੠Ⳃᷛሲᗻ PlayTennis ⱘؐབϟ˖
Temperature: 40 48 60 72 80 90
PlayTennis: No No Yes Yes Yes No
ᇍሲᗻTempratureˈᑨ䆹ᅮНҔМḋⱘ෎Ѣ䯜ؐⱘᏗᇨሲᗻਸ਼˛᮴⭥ˈ៥ӀӮ䗝ᢽ
ѻ⫳᳔໻ֵᙃ๲Ⲟⱘ䯜ؐcǄ佪‫ܜ‬ᣝ✻䖲㓁ሲᗻAᥦᑣḋ՟ˈ✊ৢ⹂ᅮⳂᷛߚ㉏ϡৠⱘ
Ⳍ䚏ᅲ՟ˈѢᰃ៥Ӏৃҹѻ⫳ϔ㒘‫׭‬䗝䯜ؐˈᅗӀⱘؐᰃⳌᑨⱘAؐП䯈ⱘЁ䯈ؐǄৃ
ҹ䆕ᯢѻ⫳᳔໻ֵᙃ๲ⲞⱘcؐᖙᅮԡѢ䖭ḋⱘ䖍⬠Ё˄Fayyad 1991˅Ǆ✊ৢৃҹ䗮䖛
䅵ㅫϢ↣Ͼ‫׭‬䗝䯜ؐ݇㘨ⱘֵᙃ๲Ⲟ䆘Ԅ䖭ѯ‫׭‬䗝ؐǄ೼ᔧࠡⱘ՟ᄤЁˈ᳝ϸϾ‫׭‬䗝䯜
ؐˈᅗӀᇍᑨѢⳂᷛሲᗻPlayTennisব࣪ᯊሲᗻTemperatureⱘؐ˖ ˄48+60˅/2 ੠˄80+90˅
/2Ǆ✊ৢ䅵ㅫ↣ϔϾ‫׭‬䗝ሲᗻüüTemperature>54 ੠Temperature>85ⱘֵᙃ๲Ⲟˈᑊ䗝ᢽ
᳔དⱘ˄Temperature>54˅Ǆ⦄೼䖭Ͼࡼᗕ߯ᓎⱘᏗᇨሲᗻ֓ৃҹ੠݊Ҫ‫׭‬䗝ⱘ⾏ᬷؐሲ
ᗻϔৠĀゲѝāˈҹ⫼Ѣ๲䭓‫އ‬ㄪᷥǄFayyad & Irani˄1993˅䅼䆎њ䖭⾡ᮍ⊩ⱘϔϾᠽ
ሩˈेᡞ䖲㓁ⱘሲᗻߚࡆ៤໮Ͼऎ䯈ˈ㗠ϡᰃ෎Ѣऩϔ䯜ؐⱘϸϾऎ䯈ǄUtgoff &
Brodley˄1991˅੠Murthy et al.˄1994˅䅼䆎њ䗮䖛ᇍ޴Ͼ䖲㓁ؐሲᗻⱘ㒓ᗻ㒘ড়ᅮН
䯜ؐখ᭄ⱘᮍ⊩Ǆ
3.7.3 ኔᙝ䘿᤟ⲺެԌᓜ䠅ḽ߼
ֵᙃ๲Ⲟᑺ䞣ᄬ೼ϔϾ‫ݙ‬೼‫أ‬㕂ˈᅗ‫أ‬㹦‫݋‬᳝䕗໮ؐⱘሲᗻǄВϔϾᵕッⱘ՟ᄤˈ
㗗㰥ሲᗻ Dateˈᅗ᳝໻䞣ⱘৃ㛑ؐ˄՟བ March 4,1979˅Ǆ㽕ᰃ៥Ӏᡞ䖭Ͼሲᗻࡴࠄ㸼
3-2 ⱘ᭄᥂ЁˈᅗӮ೼᠔᳝ሲᗻЁ᳔᳝໻ⱘֵᙃ๲ⲞǄ䖭ᰃ಴Ўऩ⣀ Date ህৃҹᅠܼ乘
⌟䆁㒗᭄᥂ⱘⳂᷛሲᗻǄѢᰃ䖭ϾሲᗻӮ㹿䗝԰ᷥⱘḍ㒧⚍ⱘ‫އ‬ㄪሲᗻᑊᔶ៤ϔỉ⏅ᑺ
Ўϔ㑻Ԛै䴲ᐌᆑⱘᷥˈ䖭ỉᷥৃҹ⧚ᛇഄߚ㉏䆁㒗᭄᥂Ǆᔧ✊ˈ䖭Ͼ‫އ‬ㄪᷥᇍѢৢᴹ
᭄᥂ⱘᗻ㛑ӮⳌᔧᏂˈ಴Ўሑㅵᅗᅠ㕢ഄߚࡆњ䆁㒗᭄᥂ˈԚᅗϡᰃϔϾདⱘ乘⌟఼
˄predicator˅
Ǆ
ሲᗻ Date ߎњҔМ䯂乬њਸ਼˛ㅔऩഄ䆆ˈᰃ಴Ўᅗ໾໮ⱘৃ㛑ؐᖙ✊ᡞ䆁㒗ḋ՟
ߚࡆ៤䴲ᐌᇣⱘぎ䯈Ǆ಴ℸˈⳌᇍ䆁㒗ḋ՟ˈᅗӮ᳝䴲ᐌ催ⱘֵᙃ๲ⲞˈሑㅵᇍѢ᳾㾕
ᅲ՟ᅗᰃϔϾ䴲ᐌᏂⱘⳂᷛߑ᭄乘⌟఼Ǆ
䙓‫ܡ‬䖭Ͼϡ䎇ⱘϔ⾡ᮍ⊩ᰃ⫼݊Ҫᑺ䞣ˈ㗠ϡᰃֵᙃ๲Ⲟˈᴹ䗝ᢽ‫އ‬ㄪሲᗻǄϔϾ
ৃҹ䗝ᢽⱘᑺ䞣ᷛ‫ޚ‬ᰃ๲Ⲟ↨⥛˄gain ratio˅˄Quinlan 1986˅Ǆ๲Ⲟ↨⥛䗮䖛ࡴܹϔϾ
⿄԰ߚ㺖ֵᙃ˄split information˅ⱘ乍ᴹᚽ㔮㉏Ԑ Date ⱘሲᗻˈߚ㺖ֵᙃ⫼ᴹ㸵䞣ሲ
ᗻߚ㺖᭄᥂ⱘᑓᑺ੠ഛࣔᗻ˖
c
| Si | |S |
SplitInformation( S , A) { ¦ log 2 i ˄3.5˅
i 1 | S | |S|
݊ЁS1ࠄScᰃcϾؐⱘሲᗻAߚࡆS㗠ᔶ៤ⱘcϾḋ՟ᄤ䲚Ǆ⊼ᛣߚ㺖ֵᙃᅲ䰙Ϟህᰃ
S݇ѢሲᗻAⱘ৘ؐⱘ➉Ǆ䖭Ϣ៥Ӏࠡ䴶ᇍ➉ⱘՓ⫼ϡৠˈ೼䙷䞠៥Ӏা㗗㰥S݇Ѣᄺд
ࠄⱘᷥ㽕乘⌟ⱘⳂᷛሲᗻⱘؐⱘ➉Ǆ
๲Ⲟ↨⥛ᑺ䞣ᰃ⫼ࠡ䴶ⱘ๲Ⲟᑺ䞣੠䖭䞠ⱘߚ㺖ֵᙃᑺ䞣ᴹ݅ৠᅮНⱘˈे˖
Gain( S , A)
GainRatio( S , A) { ˄3.6˅
SplitInformation( S , A)
䇋⊼ᛣˈߚ㺖ֵᙃ乍䰏⹡䗝ᢽؐЎഛࣔߚᏗⱘሲᗻǄ՟བˈ㗗㰥ϔϾ৿᳝nϾḋ՟
ⱘ䲚ড়㹿ሲᗻAᕏᑩߚࡆ˄䆥⊼˖ߚ៤n㒘ˈेϔϾḋ՟ϔ㒘˅Ǆ䖭ᯊߚ㺖ֵᙃⱘؐЎlog2nǄ
ⳌডˈϔϾᏗᇨሲᗻBߚࡆৠḋⱘnϾᅲ՟ˈབᵰᙄདᑇߚϸञˈ䙷Мߚ㺖ֵᙃᰃ 1Ǆབ
ᵰሲᗻA੠Bѻ⫳ৠḋⱘֵᙃ๲Ⲟˈ䙷Мḍ᥂๲Ⲟ↨⥛ᑺ䞣ˈᯢᰒBӮᕫߚ᳈催Ǆ
Փ⫼๲Ⲟ↨⥛ҷ᳓๲Ⲟᴹ䗝ᢽሲᗻѻ⫳ⱘϔϾᅲ䰙䯂乬ᰃˈᔧᶤϾSi᥹䖥S˄|Si|||S|˅
ᯊߚ↡ৃ㛑Ў 0 ៪䴲ᐌᇣǄབᵰᶤϾሲᗻᇍѢSⱘ᠔᳝ḋ՟᳝޴Тৠḋⱘؐˈ䖭ᯊ㽕М
ᇐ㟈๲Ⲟ↨⥛᳾ᅮНˈ㽕Мᰃ๲Ⲟ↨⥛䴲ᐌ໻ǄЎњ䙓‫ܡ‬䗝ᢽ䖭⾡ሲᗻˈ៥Ӏৃҹ䞛⫼
䖭ḋϔѯਃথᓣ㾘߭ˈ↨བ‫ܜ‬䅵ㅫ↣Ͼሲᗻⱘ๲Ⲟˈ✊ৢҙᇍ䙷ѯ๲Ⲟ催䖛ᑇഛؐⱘሲ
ᗻᑨ⫼๲Ⲟ↨⥛⌟䆩˄Quinlan 1986˅
Ǆ
䰸њֵᙃ๲ⲞˈLopez de Mantaras˄1991˅ҟ㒡њ঺ϔ⾡Ⳉ᥹䩜ᇍϞ䗄䯂乬㗠䆒䅵
ⱘᑺ䞣ˈᅗᰃ෎Ѣ䎱⾏ⱘ˄distance-based˅Ǆ䖭Ͼᑺ䞣ᷛ‫ޚ‬෎Ѣ᠔ᅮНⱘϔϾ᭄᥂ߦߚ
䯈ⱘ䎱⾏ሎᑺǄ↣Ͼሲᗻⱘ䆘Ԅḍ᥂ᅗѻ⫳ⱘߦߚϢ⧚ᛇߦߚ˄гህᰃᅠ㕢ߚ㉏䆁㒗᭄
᥂ⱘߦߚ˅䯈ⱘ䎱⾏Ǆ✊ৢ䗝ᢽߦߚ᳔᥹䖥ᅠ㕢ߦߚⱘሲᗻǄLopez de Mantaras ˄1991˅
ᅮНњ䖭Ͼ䎱⾏ᑺ䞣ˈ䆕ᯢњᅗϡ‫أ‬৥᳝໻䞣ؐⱘሲᗻˈᑊ᡹ਞњ݊ᅲ偠ⷨおˈ䇈ᯢ䖭
⾡ᮍ⊩ѻ⫳ⱘ‫އ‬ㄪᷥⱘ乘⌟㊒ᑺϢ๲Ⲟ⊩੠๲Ⲟ↨⥛⊩ᕫࠄⱘ≵᳝ᯢᰒⱘᏂ߿Ǆ㗠Ϩ䖭
⾡䎱⾏ᑺ䞣䙓‫ܡ‬њ๲Ⲟ↨⥛ᑺ䞣ⱘᅲ䰙ೄ䲒ˈ೼Ҫⱘᅲ偠ЁˈᇍѢሲᗻؐϾ᭄Ꮒᓖ䴲ᐌ
໻ⱘ᭄᥂䲚ˈ䖭⾡ᮍ⊩ѻ⫳њᬜᵰᕜདⱘ䕗ᇣⱘᷥǄ
ℸ໪ˈᄺ㗙Ӏ䖬ᦤߎњ໮⾡ሲᗻ䗝ᢽᑺ䞣˄՟བˈBreiman et al. 1984˗Mingers 1989a˗

Kearns & Mansour 1996˗Dietterich et al. 1996˅ǄMingers˄1989a˅ᦤկњᅲ偠ߚᵤˈ↨
䕗њ䩜ᇍϡৠ䯂乬ⱘ޴⾡䗝ᢽᑺ䞣ⱘ᳝ᬜᑺǄҪ᡹ਞњՓ⫼ϡৠሲᗻ䗝ᢽᑺ䞣ѻ⫳ⱘ᳾
ׂ࠾‫އ‬ㄪᷥⱘ໻ᇣⱘᰒ㨫ᏂᓖǄ✊㗠೼Ҫⱘᅲ偠Ёˈϡৠⱘሲᗻ䗝ᢽᑺ䞣ᇍ᳔㒜㊒ᑺⱘ
ᕅડᇣѢৢׂ࠾ⱘ⿟ᑺ੠ᮍ⊩ᇍ᳔㒜㊒ᑺⱘᕅડǄ
3.7.4 ༺⨼㕰ቇኔᙝ‫Ⲻٲ‬䇣㓹ṭׁ
೼ᶤѯᚙ‫މ‬ϟˈৃկՓ⫼ⱘ᭄᥂ৃ㛑㔎ᇥᶤѯሲᗻⱘؐǄ՟བˈ೼एᄺ乚ඳ៥ӀᏠ
ᳯḍ᥂໮乍࣪偠ᣛᷛ乘⌟ᙷ㗙ⱘ㒧ᵰˈ✊㗠ৃ㛑ҙ᳝䚼ߚᙷ㗙‫݋‬᳝偠㸔㒧ᵰǄ೼䖭⾡ᚙ
‫މ‬ϟˈ㒣ᐌ䳔㽕ḍ᥂ℸሲᗻؐᏆⶹⱘ݊Ҫᅲ՟ˈᴹԄ䅵䖭Ͼ㔎ᇥⱘሲᗻؐǄ
㗗㰥ҹϟᚙ‫ˈމ‬Ўњ䆘Ԅሲᗻ A ᰃ৺ᰃ‫އ‬ㄪ㒧⚍ n ⱘ᳔Շ⌟䆩ሲᗻˈ㽕䅵ㅫ‫އ‬ㄪ

ᷥ೼䆹㒧⚍ⱘֵᙃ๲Ⲟ Gain(S, A)Ǆ‫؛‬ᅮ<x, c(x)>ᰃ S ЁⱘϔϾ䆁㒗ḋ՟ˈᑊϨሲᗻ A
ⱘؐ A(x)᳾ⶹǄ
໘⧚㔎ᇥሲᗻؐⱘϔ⾡ㄪ⬹ᰃ䌟㒭ᅗ㒧⚍ n ⱘ䆁㒗ḋ՟Ё䆹ሲᗻⱘ᳔ᐌ㾕ؐǄ঺ϔ
⾡ㄪ⬹ᰃৃҹ䌟㒭ᅗ㒧⚍ n ⱘ㹿ߚ㉏Ў c(x)ⱘ䆁㒗ḋ՟Ё䆹ሲᗻⱘ᳔ᐌ㾕ؐǄ✊ৢՓ⫼
䖭ϾԄ䅵ؐⱘ䆁㒗ḋ՟ህৃҹ㹿⦄᳝ⱘ‫އ‬ㄪᷥᄺдㅫ⊩Փ⫼њǄMingers˄1989a˅Ёߚ
ᵤњ䖭Ͼㄪ⬹Ǆ
㄀Ѡ⾡⿡ᖂ໡ᴖⱘㄪ⬹ᰃЎ A ⱘ↣Ͼৃ㛑ؐ䌟ϢϔϾὖ⥛ˈ㗠ϡᰃㅔऩഄᇚ᳔ᐌ
㾕ⱘؐ䌟㒭 A(x)Ǆḍ᥂㒧⚍ n ⱘḋ՟Ϟ A ⱘϡৠؐⱘߎ⦄乥⥛ˈ䖭ѯὖ⥛ৃҹ㹿‫⃵ݡ‬
Ԅ䅵Ǆ՟བˈ㒭ᅮϔϾᏗᇨሲᗻ Aˈབᵰ㒧⚍ n ࣙ৿ 6 ϾᏆⶹ A=1 ੠ 6 Ͼ A=0 ⱘḋ՟ˈ
䙷М A(x)=1 ⱘὖ⥛ᰃ 0.6ˈA(x)=0 ⱘὖ⥛ᰃ 0.4ǄѢᰃˈᅲ՟ x ⱘ 60%㹿ߚ䜡ࠄ A=1 ⱘ
ߚᬃˈ40%㹿ߚ䜡ࠄ঺ϔϾߚᬃǄ䖭ѯ⠛↉ḋ՟˄fractional examples˅ⱘⳂⱘᰃ䅵ㅫֵ
ᙃ๲Ⲟˈ঺໪ˈབᵰ᳝㄀ѠϾ㔎ᇥؐⱘሲᗻᖙ乏㹿⌟䆩ˈ䖭ѯḋ՟ৃҹ೼ৢ㒻ⱘᷥߚᬃ
㹿䖯ϔℹ㒚ߚǄϞ䗄ⱘḋ՟ⱘ⠛↉гৃҹ೼ᄺдПৢՓ⫼ˈ⫼ᴹߚ㉏㔎ᇥሲᗻⱘᮄᅲ՟Ǆ
೼䖭⾡ᚙ‫މ‬ϟˈᮄᅲ՟ⱘߚ㉏ህᰃ᳔ৃ㛑ⱘߚ㉏ˈ䅵ㅫⱘᮍ⊩ᰃ䗮䖛೼ᷥⱘ৊㒧⚍ᇍᣝ
ϡৠᮍᓣߚ㉏ⱘᅲ՟⠛↉ⱘࡴᴗ∖੠ǄC4.5˄Quinlan 1993˅Փ⫼䖭⾡ᮍ⊩໘⧚㔎ᇥⱘ
ሲᗻؐǄ
3.7.5 ༺⨼ԙԭуੂⲺኔᙝ
೼ᶤѯᄺдӏࡵЁˈᅲ՟ⱘሲᗻৃ㛑ϢҷӋⳌ݇Ǆ՟བˈ೼ᄺдߚ㉏⮒⮙ᯊ៥Ӏৃ
㛑ҹ䖭ѯሲᗻᴹᦣ䗄ᙷ㗙˖ԧ⏽ǃ⌏㒘㒛ߛ⠛Ẕᶹǃ㛝᧣ǃ㸔⎆࣪偠㒧ᵰㄝǄ䖭ѯሲᗻ
೼ҷӋᮍ䴶Ꮒ߿䴲ᐌ໻ˈϡ䆎ᰃ᠔䳔ⱘ䌍⫼䖬ᰃᙷ㗙㽕ᡓফⱘϡ䗖ǄᇍѢ䖭ḋⱘӏࡵˈ
៥ӀᇚӬ‫ܜ‬䗝ᢽሑৃ㛑Փ⫼ԢҷӋሲᗻⱘ‫އ‬ㄪᷥˈҙᔧ䳔㽕ѻ⫳ৃ䴴ⱘߚ㉏ᯊᠡձ䌪催
ҷӋሲᗻǄ
䗮䖛ᓩܹϔϾҷӋ乍ࠄሲᗻ䗝ᢽᑺ䞣ˈৃҹՓ ID3 ㅫ⊩㗗㰥ሲᗻҷӋǄ՟བˈ៥

Ӏৃҹ⫼ֵᙃ๲Ⲟ䰸ҹሲᗻⱘҷӋˈҹՓԢҷӋⱘሲᗻӮ㹿Ӭ‫ܜ‬䗝ᢽǄ㱑✊䖭⾡ҷӋᬣ
ᛳᑺ䞣ϡֱ䆕ᡒࠄ᳔ӬⱘҷӋᬣᛳ‫އ‬ㄪᷥˈᅗӀ⹂ᅲՓ᧰㋶‫أ‬㕂ࠄ᳝߽ѢԢҷӋሲᗻǄ
Tan & Schlimmer˄1990˅੠ Tan˄1993˅ᦣ䗄њϔ⾡䖭ḋⱘᮍ⊩ˈᑊᡞᅗᑨ⫼ࠄᴎ

఼ҎᛳⶹӏࡵЁǄ೼䖭ϾӏࡵЁᴎ఼Ҏᖙ乏ḍ᥂䖭ѯ⠽ԧབԩ㛑㹿ᅗⱘᴎẄ᠟ᡧԣˈҢ
㗠ᄺӮߚ䕼ϡৠⱘ⠽ԧǄ䖭⾡ᚙ‫މ‬ϟˈሲᗻᇍᑨѢᴎ఼Ҏ䑿Ϟⱘ⿏ࡼໄ㒇㦋ᕫⱘϡৠӴ
ᛳ఼䇏᭄ǄሲᗻⱘҷӋ䗮䖛ᅮԡ៪᪡԰ໄ㒇ᴹ㦋পሲᗻؐ᠔䳔ⱘ⾦᭄ᴹ㸵䞣ǄҪӀ䆕ᯢˈ
䗮䖛⫼ϟ䴶ⱘᑺ䞣ҷ᳓ֵᙃ๲Ⲟሲᗻ䗝ᢽᑺ䞣ˈᄺࠄњ᳈ࡴ᳝ᬜⱘ䆚߿ㄪ⬹ˈৠᯊ≵᳝
ᤳ༅ߚ㉏ⱘ㊒ᑺǄ
Gain 2 ( S , A)
Cost ( A)
Nunez˄1988˅Ёᦣ䗄њϔ⾡᳝݇ⱘᮍ⊩ˈᑊᡞᅗᑨ⫼ࠄᄺдए⭫䆞ᮁ㾘߭ϞǄ䖭
䞠ሲᗻᰃ‫݋‬᳝ϡৠҷӋⱘϡৠ⮛⢊੠࣪偠⌟䆩Ǆᅗⱘ㋏㒳Փ⫼њ⿡ᖂ᳝⚍ϡৠⱘሲᗻ䗝
ᢽᑺ䞣˖
2 Gain ( S , A) 1
(Cost ( A) 1) w
݊Ё w[0ˈ1]ᰃϔϾᐌ᭄ˈ‫އ‬ᅮҷӋᇍֵᙃ๲ⲞⱘⳌᇍ䞡㽕ᗻǄNunez˄1991˅䩜
ᇍϔ㋏߫ӏࡵ㒭ߎњ䖭ϸ⾡ᮍ⊩ⱘ䆩偠ᇍ↨Ǆ
䖭ϔゴⱘ㽕⚍ࣙᣀ˖
z ‫އ‬ㄪᷥᄺдЎὖᗉᄺд੠ᄺд݊Ҫ⾏ᬷؐⱘߑ᭄ᦤկњϔϾᅲ⫼ⱘᮍ⊩ǄID3
㋏߫ㅫ⊩Փ⫼Ңḍ৥ϟ๲䭓⊩᥼ᮁ‫އ‬ㄪᷥˈЎ↣Ͼ㽕ࡴܹᷥⱘᮄ‫އ‬ㄪߚᬃ䋾ှ
ഄ䗝ᢽ᳔དⱘሲᗻǄ
z ID3 ㅫ⊩᧰㋶ᅠᭈⱘ‫؛‬䆒ぎ䯈˄гህᰃ䇈ˈ‫އ‬ㄪᷥぎ䯈㛑໳㸼⼎ӏԩᅮН೼⾏ᬷ
ؐᅲ՟Ϟⱘӏԩ⾏ᬷؐߑ᭄˅Ǆ᠔ҹᅗ䙓‫ܡ‬њҙ㗗㰥᳝䰤ⱘ‫؛‬䆒䲚ড়ⱘᮍ⊩ⱘ
Џ㽕䯂乬˖Ⳃᷛߑ᭄ৃ㛑ϡ೼‫؛‬䆒ぎ䯈ЁǄ
z 䱤৿೼ ID3 ㅫ⊩Ёⱘᔦ㒇‫أ‬㕂ࣙᣀӬ‫ܜ‬䗝ᢽ䕗ᇣⱘᷥˈгህᰃ䇈ˈᅗ䗮䖛ᇍ‫؛‬
䆒ぎ䯈ⱘ᧰㋶๲䭓ᷥˈՓᷥⱘ໻ᇣЎℷད㛑ߚ㉏Ꮖ᳝ⱘ䆁㒗ḋ՟Ǆ
z 䖛ᑺᢳড়䆁㒗᭄᥂ᰃ‫އ‬ㄪᷥᄺдЁⱘ䞡㽕䯂乬Ǆ಴Ў䆁㒗ḋ՟ҙҙᰃ᠔᳝ৃ㛑
ᅲ՟ⱘϔϾḋᴀˈ৥ᷥ๲ࡴߚᬃৃ㛑ᦤ催೼䆁㒗ḋ՟Ϟⱘᗻ㛑ˈԚै䰡Ԣ೼䆁
㒗ᅲ՟໪ⱘ݊Ҫᅲ՟Ϟⱘᗻ㛑Ǆ಴ℸˈৢׂ࠾‫އ‬ㄪᷥⱘᮍ⊩ᇍѢ䙓‫އܡ‬ㄪᷥᄺ
дЁ˄੠݊ҪՓ⫼Ӭ䗝‫أ‬㕂ⱘᔦ㒇᥼⧚ᮍ⊩˅ⱘ䖛ᑺᢳড়ᰃᕜ䞡㽕ⱘǄ
z ᇍѢ෎ᴀ ID3 ㅫ⊩ˈⷨお㗙Ꮖ㒣ᓔথњ໻䞣ⱘᠽሩǄ݊Ёࣙᣀৢׂ࠾ⱘᮍ⊩˗
໘⧚ᅲ᭄ؐⱘሲᗻ˗ᆍ㒇㔎ᇥሲᗻؐⱘ䆁㒗ḋ՟˗ᔧ᳝њᮄⱘ䆁㒗ᅲ՟ᯊ䗦๲
㊒࣪‫އ‬ㄪᷥ˗Փ⫼ֵᙃ๲ⲞП໪ⱘ݊Ҫሲᗻ䗝ᢽᑺ䞣˗㗗㰥Ϣᅲ՟ሲᗻ݇㘨ⱘ
ҷӋǄ
݇Ѣ‫އ‬ㄪᷥᄺдⱘ᳔ᮽⱘ㨫԰᳝ Hunt ⱘὖᗉᄺд㋏㒳˄Concept Learning Systemˈ

CLS˅˄Hunt et al. 1966˅ҹঞ Friedman ੠ Breiman ⱘ CART ㋏㒳˄Friedman 1977;Breiman
et al. 1984˅ǄQuinlan ⱘ ID3 ㋏㒳˄Quinlan 1979ˈ1983˅ᵘ៤њᴀゴ䅼䆎ⱘ෎⸔Ǆ݇Ѣ
‫އ‬ㄪᷥᄺдⱘ݊Ҫᮽᳳ㨫԰ࣙᣀ ASSISTANT˄Kononenko et al. 1984˗Cestnik et al.
1987˅Ǆ‫އ‬ㄪᷥᔦ㒇ㅫ⊩೼໮᭄䅵ㅫᴎᑇৄϞⱘᅲ⦄ৃҹଚϮᮍᓣᕫࠄǄ
݇Ѣ‫އ‬ㄪᷥᔦ㒇ⱘ䖯ϔℹ㒚㡖ˈQuinlan˄1993˅ᰃϔᴀ㊒ᔽⱘ㨫԰ˈ݊Ё䅼䆎њ
ᕜ໮ᅲ䏉䯂乬ˈᑊᦤկњ C4.5 ㅫ⊩ⱘৃᠻ㸠ҷⷕǄMingers˄1989a˅੠ Buntine & Niblett
˄1992˅ᦤկњ↨䕗ϡৠሲᗻ䗝ᢽᑺ䞣ⱘᅲ偠ⷨおǄMingers˄1989b˅ᦤկњᇍϡৠׂ
࠾ㄪ⬹ⱘⷨおǄ↨䕗‫އ‬ㄪᷥᄺд੠݊Ҫᄺдᮍ⊩ⱘ䆩偠ৃ೼ӫ໮ⱘ䆎᭛Ёᡒࠄˈࣙᣀ
˄Dietterich et al. 1995; Fisher & McKusick 1989; Quinlan 1988a; Shavlik et al. 1991;
Thrun et al. 1991; Weiss and Kapouleas 1989˅Ǆ
д乬
3.1 ⬏ߎ㸼⼎ϟ䴶Ꮧᇨߑ᭄ⱘ‫އ‬ㄪᷥ˖
˄a˅A¬B
˄b˅A[BC]
˄c˅A XOR B
˄d˅[AB] [CD]
3.2 㗗㰥ϟ䴶ⱘ䆁㒗ḋ՟䲚ড়˖
ᅲ՟ ߚ㉏ a1 a2
1 + T T
2 + T T
3 - T F
4 + F F
5 - F T
6 - F T
˄a˅䇋䅵ㅫ䖭Ͼ䆁㒗ḋ՟䲚ড়ᇍѢⳂᷛߑ᭄ߚ㉏ⱘ➉Ǆ
˄b˅䇋䅵ㅫሲᗻa2Ⳍᇍ䖭ѯ䆁㒗ḋ՟ⱘֵᙃ๲ⲞǄ
3.3 ߸ᮁҹϟੑ乬ⱘℷ䇃˖བᵰᷥ D2 ᰃҢᷥ D1 ࡴᎹⱘˈ䙷М D1 more-general-than D2Ǆ

‫؛‬ᅮ D1 ੠ D2 ᰃ㸼⼎ӏᛣᏗᇨߑ᭄ⱘ‫އ‬ㄪᷥˈ㗠Ϩᔧ ID3 㛑ᡞ D1 ᠽሩ៤ D2 ᯊˈ䙷М
D2 ᰃ D1 ⱘࡴᎹǄབᵰℷ⹂ˈ㒭ߎ䆕ᯢ˗བᵰ䫭䇃ˈВߎϔϾড՟Ǆ˄more-general-than
㹿ᅮН೼㄀ 2 ゴЁǄ˅
3.4 ID3 ҙᇏᡒϔϾϔ㟈ⱘ‫؛‬䆒ˈ㗠‫׭‬䗝⍜䰸ㅫ⊩ᇏᡒ᠔᳝ϔ㟈ⱘ‫؛‬䆒Ǆ㗗㰥䖭ϸ⾡ᄺ

дㅫ⊩䯈ⱘᇍᑨ݇㋏Ǆ
˄a˅‫؛‬ᅮ㒭ᅮ EnjoySport ⱘಯϾ䆁㒗ḋ՟ˈ⬏ߎ ID3 ᄺдⱘ‫އ‬ㄪᷥǄ݊Ё EnjoySport

Ⳃᷛὖᗉ߫೼㄀ 2 ゴⱘ㸼 2-1 ЁǄ
˄b˅ᄺдࠄⱘ‫އ‬ㄪᷥ੠Ңৠḋⱘḋ՟Փ⫼বൟぎ䯈ㅫ⊩ᕫࠄⱘবൟぎ䯈˄㾕㄀ 2
ゴ೒ 2-3˅䯈᳝ҔМ݇㋏˛ᷥㄝӋѢবൟぎ䯈ⱘϔϾ៤ਬ৫˛
˄c˅๲ࡴϟ䴶ⱘ䆁㒗ḋ՟ˈ䅵ㅫᮄⱘ‫އ‬ㄪᷥǄ䖭ϔ⃵ˈᰒ⼎ߎ๲䭓ᷥⱘ↣ϔℹЁ
↣Ͼ‫׭‬䗝ሲᗻⱘֵᙃ๲ⲞǄ
Sky Air-Temp Humidity Wind Water Forecast Enjoy-Sport?
Sunny Warm Normal Weak Warm Same No
˄d˅‫؛‬ᅮ៥ӀᏠᳯ䆒䅵ϔϾᄺд఼ˈᅗ᧰㋶‫އ‬ㄪᷥ‫؛‬䆒ぎ䯈˄㉏Ԑ ID3˅ᑊᇏᡒ
Ϣ᭄᥂ϔ㟈ⱘ᠔᳝‫؛‬䆒˄㉏Ԑ‫׭‬䗝⍜䰸˅Ǆㅔऩഄ䇈ˈ៥ӀᏠᳯᑨ⫼‫׭‬䗝⍜䰸ㅫ⊩᧰㋶
‫އ‬ㄪᷥ‫؛‬䆒ぎ䯈Ǆ‫ߎݭ‬㒣䖛㸼 2-1 ⱘ㄀ϔϾ䆁㒗ḋ՟ৢⱘ S ੠ G 䲚ড়Ǆ⊼ᛣ S ᖙ乏ࣙ
৿Ϣ᭄᥂ϔ㟈ⱘ᳔⡍⅞ⱘ‫އ‬ㄪᷥˈ㗠 G ᖙ乏ࣙ৿᳔ϔ㠀ⱘǄ䇈ᯢ䘛ࠄ㄀ѠϾ䆁㒗ḋ՟
ᯊ S ੠ G 䲚ড়ᰃབԩ㹿ᬍ䖯ⱘ˄ৃҹএᥝᦣ䗄ৠϔϾὖᗉⱘ䇁⊩ϡৠⱘᷥ˅Ǆ೼ᡞ‫׭‬䗝
⍜䰸ㅫ⊩ᑨ⫼ࠄ‫އ‬ㄪᷥ‫؛‬䆒ぎ䯈ᯊˈ乘䅵Ӯ⺄ࠄҔМḋⱘೄ䲒˛
ㅢ4ㄖ Ӱᐛ⾔㔅㖇㔒
人工神经网络（Artificial Neural Networks——ANNs）提供了一种普遍而且实

用的方法，来从样例中学习值为实数、离散或向量的函数。像反向传播 BP
（BackPropagation）这样的算法使用梯度下降来调节网络参数以最佳拟合由输入
-输出对组成的训练集合。ANN 学习对于训练数据中的错误鲁棒性很好，且已经成
功地应用到很多领域，例如视觉场景分析（interpreting visual scenes）、语音识别、
以及机器人控制等。
4.1 ㆶԁ
⾎㓿㖁㔌ᆖҐᯩ⌅ሩҾ䙬䘁ᇎᮠ٬ǃ⿫ᮓ٬ᡆੁ䟿٬Ⲵⴞḷ࠭ᮠᨀ‫׋‬Ҷа⿽励ἂᙗᖸᕪ
Ⲵᯩ⌅ǄሩҾḀӋ㊫රⲴ䰞仈ˈྲᆖҐ䀓䟺༽ᵲⲴ⧠ᇎц⭼ѝⲴՐᝏಘᮠᦞˈӪᐕ⾎㓿㖁㔌
ᱟⴞࡽ⸕䚃Ⲵᴰᴹ᭸ᆖҐᯩ⌅Ǆֻྲˈᵜㄐ㾱᧿䘠Ⲵ৽ੁՐ᫝㇇⌅ᐢ൘ᖸཊᇎ䱵Ⲵ䰞仈ѝਆ
ᗇҶ᛺ӪⲴᡀ࣏ˈ∄ྲᆖҐ䇶࡛᡻߉ᆇㅖ˄LeCun et al. 1989˅ˈᆖҐ䇶࡛ਓ䈝˄Lang et al.
1990˅઼ᆖҐ䇶࡛Ӫ㝨˄Cottrell 1990˅ǄRumelhart et al.˄1994˅ѝᾲ㿸Ҷަᇎ䱵Ⲵᓄ⭘Ǆ
4.1.1 ⭕⢟ᆜࣞᵰ
Ӫᐕ⾎㓿㖁㔌Ⲵ⹄ウ൘аᇊ〻ᓖкਇࡠҶ⭏⢙ᆖⲴ੟ਁˈഐѪ⭏⢙ⲴᆖҐ㌫㔏ᱟ⭡⴨ӂ
䘎᧕Ⲵ⾎㓿‫˄ݳ‬neuron˅㓴ᡀⲴᔲᑨ༽ᵲⲴ㖁㔌Ǆ㘼Ӫᐕ⾎㓿㖁㔌о↔བྷփ⴨լˈᆳᱟ⭡а
㌫ࡇㆰঅঅ‫⴨ݳ‬ӂᇶ䳶䘎᧕ᶴᡀˈަѝ⇿ањঅ‫ݳ‬ᴹаᇊᮠ䟿Ⲵᇎ٬䗃‫˄ޕ‬ਟ㜭ᱟަԆঅ‫ݳ‬
Ⲵ䗃ࠪ˅ˈᒦӗ⭏অаⲴᇎᮠ٬䗃ࠪ˄ਟ㜭ᡀѪަԆᖸཊঅ‫Ⲵݳ‬䗃‫˅ޕ‬Ǆ
ѪҶ࣐␡ሩ䘉⿽㊫∄Ⲵ䇔䇶ˈ䇙ᡁԜ㘳㲁аӋᶕ㠚⭏⢙ᆖⲴһᇎǄֻྲˈᦞՠ䇑Ӫ㊫Ⲵ
བྷ㝁ᱟ⭡བྷ㓖 1011њ⾎㓿‫⴨ݳ‬ӂ䘎᧕㓴ᡀⲴᇶ䳶㖁㔌ˈᒣ൷⇿ањ⾎㓿‫ݳ‬оަԆ 104њ⾎㓿
‫⴨ݳ‬䘎Ǆ⾎㓿‫Ⲵݳ‬⍫ᙗ䙊ᑨ㻛䙊ੁަԆ⾎㓿‫Ⲵݳ‬䘎᧕◰⍫ᡆᣁࡦǄⴞࡽ⸕䚃ⲴᴰᘛⲴ⾎㓿‫ݳ‬
䖜ᦒᰦ䰤ᱟ൘ 10-3。㓗࡛——о䇑㇇ᵪⲴ䖜ᦒᰦ䰤 10-10。⴨∄ធᖸཊǄ❦㘼Ӫ㊫㜭ཏԕ᛺Ӫ
Ⲵ䙏ᓖ‫༽ࠪڊ‬ᵲᓖ᛺ӪⲴߣㆆǄֻྲˈ֐㾱䙊䗷㿶㿹䇔ࠪ㠚ᐡⲴ⇽Ӣབྷ㓖䴰㾱 10-1。Ǆ⌘᜿
൘䘉 10-1。Ⲵ䰤䳄޵ˈ㻛◰ਁⲴ⾎㓿‫ݳ‬ᒿࡇн䮯ҾᮠⲮ↕ˈഐѪঅњ⾎㓿‫Ⲵݳ‬䖜ᦒ䙏ᓖᐢ⸕Ǆ
䘉њһᇎ֯ᖸཊӪ᧘⍻ˈ⭏⢙⾎㓿㌫㔏Ⲵؑ᚟༴⨶㜭࣋аᇊᗇ⳺Ҿሩ࠶ᐳ൘བྷ䟿⾎㓿‫ݳ‬кⲴ
ؑ᚟㺘⽪Ⲵ儈ᓖᒦ㹼༴⨶ǄANN㌫㔏Ⲵањࣘᵪቡᱟ㧧ᗇ䘉⿽สҾ࠶ᐳ㺘⽪Ⲵ儈ᓖᒦ㹼㇇
⌅ǄབྷཊᮠⲴANN䖟Ԧ൘Ѣ㹼ᵪಘкԯⵏ࠶ᐳ༴⨶ˈ❦㘼ᴤᘛ⡸ᵜⲴ㇇⌅ҏᐢ㓿൘儈ᓖᒦ
㹼ᵪ઼⢩࡛ѪANNᓄ⭘䇮䇑Ⲵу⭘⺜Ԧкᇎ⧠Ǆ
⭡Ҿ ANN ਚᱟаᇊ〻ᓖൠਇ⭏⢙⾎㓿㌫㔏Ⲵ੟ਁˈᡰԕ ANN ᒦᵚ⁑ᤏ⭏⢙⾎㓿㌫㔏

ѝⲴᖸཊ༽ᵲ⢩ᖱˈ㘼фᐢ㓿⸕䚃 ANN Ⲵᖸཊ⢩ᖱо⭏⢙㌫㔏ҏᱟна㠤ⲴǄֻྲˈሩҾ
ᡁԜ㘳㲁Ⲵ ANNˈ⇿њঅ‫ݳ‬䗃ࠪঅаⲴнਈ٬ˈ❦㘼⭏⢙⾎㓿‫ݳ‬䗃ࠪⲴᱟ༽ᵲⲴᰦᒿ㜹ߢǄ
䮯ᵏԕᶕˈӪᐕ⾎㓿㖁㔌亶ฏⲴ⹄ウ㘵࠶ѪєњഒփǄањഒփⲴⴞḷᱟ֯⭘ ANN ⹄
ウ઼⁑ᤏ⭏⢙ᆖҐ䗷〻ǄਖањഒփⲴⴞḷᱟ㧧ᗇ儈᭸ⲴᵪಘᆖҐ㇇⌅ˈн㇑䘉⿽㇇⌅ᱟ੖
৽᱐Ҷ⭏⢙䗷〻Ǆ൘ᵜҖѝᡁԜⲴ‫ޤ‬䏓ㅖਸਾаഒփˈᡰԕᡁԜнՊ޽ᢺ⌘᜿࣋⭘൘⭏⢙⁑
රкǄ㤕㾱㧧ᗇ‫ޣ‬Ҿ֯⭘ ANN ⁑ᤏ⭏⢙㌫㔏Ⲵᴤཊؑ᚟䈧৲㘳 Churchland & Sejnowski
˄1992˅ˈZornetzer et al.˄1994˅ˈGabriel & Moore˄1990˅Ǆ
4.2 ⾔㔅㖇㔒㺞⽰
Pomerleau˄1993˅Ⲵ ALVINN ㌫㔏ᱟ ANN ᆖҐⲴањިරᇎֻˈ䘉њ㌫㔏֯⭘ањ

ᆖҐࡠⲴ ANN ԕ↓ᑨⲴ䙏ᓖ൘儈䙏‫ޜ‬䐟к傮傦⊭䖖ǄANN Ⲵ䗃‫ޕ‬ᱟањ 30u32 ‫ۿ‬㍐Ⲵ㖁
Ṭˈ‫ۿ‬㍐ⲴӞᓖᶕ㠚ањᆹ㻵൘䖖䖶кⲴࡽੁᩴ‫ۿ‬ᵪǄANN Ⲵ䗃ࠪᱟ䖖䖶㹼䘋ⲴᯩੁǄ䘉
њ ANN 䙊䗷㿲ሏӪ㊫傮傦ᰦⲴ᫽㓥ભԔ䘋㹼䇝㓳ˈ䇝㓳䗷〻བྷ㓖 5 ࠶䫏ǄALVINN ⭘ᆖҐ
ࡠⲴ㖁㔌൘儈䙏‫ޜ‬䐟кԕ 70 㤡䟼ᰦ䙏ᡀ࣏ൠ傮傦Ҷ 90 㤡䟼˄൘࠶㹼‫ޜ‬䐟Ⲵᐖ䖖䚃㹼傦ˈ਼
ᰦᴹަԆ䖖䖶˅Ǆ
മ 4-1 ⭫ࠪҶ ALVINN ㌫㔏Ⲵањ⡸ᵜѝ֯⭘䗷Ⲵ⾎㓿㖁㔌㺘⽪ˈ䘉ҏᱟᖸཊ ANN ㌫

㔏Ⲵިර㺘⽪ᯩᔿǄ⾎㓿㖁㔌ᱮ⽪൘മⲴᐖ䗩ˈ䗃‫Ⲵޕ‬ᩴ‫ۿ‬ᵪമ‫ۿ‬൘ᆳⲴл䗩Ǆ㖁㔌മѝ⇿
њ㔃⛩ሩᓄањ㖁㔌অ‫˄ݳ‬unit˅Ⲵ䗃ࠪˈ㘼Ӿлᯩ䘋‫ޕ‬㔃⛩Ⲵᇎ㓯Ѫަ䗃‫ޕ‬Ǆਟԕⴻࡠˈ
‫ޡ‬ᴹഋњঅ‫᧕ⴤݳ‬Ӿമ‫᧕ۿ‬᭦ᡰᴹⲴ 30u32 њ‫ۿ‬㍐Ǆ䘉ഋњঅ‫ݳ‬㻛〠ѪĀ䳀㯿āঅ‫ˈݳ‬ഐѪ
ᆳԜⲴ䗃ࠪӵ൘㖁㔌޵䜘ˈнᱟᮤњ㖁㔌䗃ࠪⲴа䜘࠶Ǆ⇿њ䳀㯿অ‫ݳ‬ṩᦞ 960 њ䗃‫࣐Ⲵޕ‬
ᵳ઼䇑㇇ᗇࡠঅаⲴᇎᮠ٬䗃ࠪǄ❦ਾ䘉ഋњ䳀㯿অ‫Ⲵݳ‬䗃ࠪ㻛⭘֌ㅜҼቲ 30 њĀ䗃ࠪঅ
‫ݳ‬āⲴ䗃‫ޕ‬Ǆ⇿њ䗃ࠪঅ‫ݳ‬ሩᓄањ⢩ᇊⲴ傮傦ᯩੁˈ䘉Ӌঅ‫Ⲵݳ‬䗃ࠪߣᇊଚањ傮傦ᯩੁ
ᱟᴰᕪ⛸᧘㦀ⲴǄ
sharp left-ᙕࢗᐖ䖜
sharp right-ᙕࢗਣ䖜
straight ahead-↓ࡽᯩ
30 Output units-30 њ䗃ࠪঅ‫ݳ‬
4 Hidden units-4 њ䳀㯿অ‫ݳ‬
30u32 sensor input retina-30u32 Րᝏಘ㿶㖁㟌䗃‫ޕ‬
മ 4-1 ᆖҐ傮傦⊭䖖Ⲵ⾎㓿㖁㔌
ALVINN ㌫㔏֯⭘৽ੁՐ᫝㇇⌅ᶕᆖҐ傮傦⊭䖖˄кമ˅ˈᆳⲴᴰ儈ᰦ䙏䗮ࡠ⇿ሿᰦ 70 㤡䟼Ǆᐖ

മᱮ⽪Ҷᶕ㠚䖖ࡽᩴ‫ۿ‬ᵪⲴമ‫ۿ‬ᱟྲօ㻛᱐ሴࡠ 960 њ⾎㓿㖁㔌䗃‫ˈⲴޕ‬䘉Ӌ䗃‫ޕ‬৸ࡽ侸ࡠ 4 њ䳀
㯿অ‫޽ˈݳ‬䘎᧕ࡠ 30 њ䗃ࠪঅ‫ݳ‬Ǆ㖁㔌䗃ࠪ㕆⸱Ҷ᧘㦀Ⲵ傮傦ᯩੁǄਣമᱮ⽪Ҷ㖁㔌ѝањ䳀㯿
অ‫Ⲵݳ‬ᵳ٬Ǆ䘋‫ޕ‬䘉њ䳀㯿অ‫ Ⲵݳ‬30u32 њᵳ٬ᱮ⽪൘བྷⲴ⸙䱥ѝˈⲭ㢢ⲴᯩṶ㺘⽪↓ᵳ٬㘼唁㢢
ⲴᯩṶ㺘⽪䍏ᵳ٬ǄӾ䘉њ䳀㯿অ‫ ࡠݳ‬30 њ䗃ࠪঅ‫Ⲵݳ‬ᵳ٬㻛⭫൘䘉њབྷ⸙䱥кᯩⲴ䖳ሿ⸙ᖒѝǄ
Ӿ䘉Ӌ䗃ࠪᵳ٬ਟԕⴻࠪˈ◰⍫䘉њ䳀㯿অ‫ݳ‬Պ‫׳‬䘋ੁᐖ䖜Ǆ
മ 4-1 ѝⲴਣ‫ח‬䜘࠶᧿㔈ⲴᱟаӋᆖҐᗇࡠⲴᵳ٬ˈᆳԜо䘉њ ANN Ⲵഋњ䳀㯿অ‫ݳ‬

ѻа⴨㚄㌫Ǆл䶒Ⲵ唁ⲭᯩṬབྷ⸙䱥᧿䘠ⲴᱟӾ 30u32 ‫ۿ‬㍐䗃‫ࡠޕ‬䘉њ䳀㯿অ‫Ⲵݳ‬ᵳ٬Ǆ䘉
䟼ˈⲭᯩṬ㺘⽪↓ᵳ٬ˈ唁ᯩṬ㺘⽪䍏ᵳ٬ˈᯩṬⲴབྷሿ㺘⽪ᵳⲴᮠ䟿Ǆབྷ⸙䱥↓кᯩⲴ䖳
ሿⲴ⸙ᖒ㺘⽪Ӿ䘉њ䳀㯿অ‫ ࡠݳ‬30 њ䗃ࠪঅ‫Ⲵݳ‬ᵳǄ
ALVINN Ⲵ㖁㔌㔃ᶴᱟᖸཊ ANN ѝⲴިර㔃ᶴǄᡰᴹঅ‫࠶ݳ‬ቲӂ䘎ᖒᡀҶањᴹੁᰐ

⧟മǄ䙊ᑨˈANN മⲴ㔃ᶴਟԕᴹᖸཊ⿽㊫ර——ᰐ⧟Ⲵᡆᴹ⧟ⲴˈᴹੁⲴᡆᰐੁⲴǄᵜ
ㄐ䳶ѝ䇘䇪ԕ৽ੁՐ᫝㇇⌅Ѫส⹰Ⲵᴰᑨ㿱઼ᴰᇎ⭘Ⲵ ANN ᯩ⌅Ǆ৽ੁՐ᫝㇇⌅‫ٷ‬ᇊ㖁㔌
ᱟањപᇊ㔃ᶴˈሩᓄањᴹੁമˈਟ㜭वਜ਼⧟ǄANN ᆖҐቡᱟѪമѝⲴ⇿аᶑ䗩䘹ਆᵳ
٬Ǆቭ㇑Ḁ⿽㊫රⲴᗚ⧟ᱟ‫ݱ‬䇨ⲴˈབྷཊᮠⲴᇎ䱵ᓄ⭘䜭䟷⭘ᰐ⧟Ⲵࡽ侸㖁㔌ˈо ALVINN
֯⭘Ⲵ㖁㔌㔃ᶴ⴨լǄ
4.3 䘸ਾ⾔㔅㖇㔒ᆜҖⲺ䰤从
ANN ᆖҐ䶎ᑨ䘲ਸҾ䘉ṧⲴ䰞仈˖䇝㓳䳶ਸѪਜ਼ᴹಚ༠Ⲵ༽ᵲՐᝏಘᮠᦞˈֻྲᶕ㠚
ᩴ‫ۿ‬ᵪ઼哖‫ݻ‬仾ⲴᮠᦞǄᆳҏ䘲⭘Ҿ䴰㾱ᴤཊㅖਧ㺘⽪Ⲵ䰞仈ˈֻྲㅜ 3 ㄐ䇘䇪Ⲵߣㆆṁᆖ
Ґԫ࣑Ǆ䘉⿽ᛵߥл ANN ઼ߣㆆṁᆖҐ㓿ᑨӗ⭏㋮ᓖབྷփ⴨ᖃⲴ㔃᷌Ǆਟ৲㿱 Shavlik et al.
˄1991˅઼ Weiss and Kapouleas˄1989˅ѝ‫ޣ‬Ҿߣㆆṁ઼ ANN ᆖҐⲴᇎ傼∄䖳Ǆ৽ੁՐ᫝
㇇⌅ᱟᴰᑨ⭘Ⲵ ANN ᆖҐᢰᵟǄᆳ䘲ਸާᴹԕл⢩ᖱⲴ䰞仈˖
x ᇎֻᱟ⭘ᖸཊĀ኎ᙗ-٬āሩ㺘⽪ⲴǄ㾱ᆖҐⲴⴞḷ࠭ᮠᱟᇊѹ൘ਟԕ⭘ੁ
䟿᧿䘠ⲴᇎֻѻкⲴˈੁ䟿⭡亴‫ݸ‬ᇊѹⲴ⢩ᖱ㓴ᡀˈֻྲ ALVINN ֻᆀѝ
Ⲵ‫ۿ‬㍐٬Ǆ䘉Ӌ䗃‫ޕ‬኎ᙗѻ䰤ਟԕ儈ᓖ⴨‫ˈޣ‬ҏਟԕ⴨ӂ⤜・Ǆ䗃‫ޕ‬٬ਟԕ
ᱟԫօᇎᮠǄ
x ⴞḷ࠭ᮠⲴ䗃ࠪਟ㜭ᱟ⿫ᮓ٬ǃᇎᮠ٬ᡆ㘵⭡㤕ᒢᇎᮠ኎ᙗᡆ⿫ᮓ኎ᙗ㓴ᡀ
Ⲵੁ䟿Ǆֻྲˈ൘ ALVINN ㌫㔏ѝ䗃ࠪⲴᱟ 30 њ኎ᙗⲴੁ䟿ˈ⇿ањ࠶䟿
ሩᓄањᔪ䇞Ⲵ傮傦ᯩੁǄ⇿њ䗃ࠪ٬ᱟ 0 ઼ 1 ѻ䰤ⲴḀњᇎᮠˈሩᓄҾ൘
亴⍻⴨ᓄ傮傦ᯩੁᰦⲴ㖞ؑᓖ˄confidence˅ǄᡁԜҏਟԕ䇝㓳ањঅа㖁
㔌ˈ਼ᰦ䗃ࠪ㹼傦ᯩੁ઼ᔪ䇞Ⲵ࣐䙏ᓖˈ䘉ਚ㾱ㆰঅൠᢺ㕆⸱䘉є⿽䗃ࠪ亴
⍻Ⲵੁ䟿䘎᧕൘а䎧ቡਟԕҶǄ
x 䇝㓳ᮠᦞਟ㜭वਜ਼䭉䈟ǄANN ᆖҐ㇇⌅ሩҾ䇝㓳ᮠᦞѝⲴ䭉䈟ᴹ䶎ᑨྭⲴ
励ἂᙗǄ
x ਟᇩᗽ䮯ᰦ䰤Ⲵ䇝㓳Ǆ㖁㔌䇝㓳㇇⌅䙊ᑨ∄‫ߣۿ‬ㆆṁᆖҐ䘉ṧⲴ㇇⌅䴰㾱ᴤ
䮯Ⲵ䇝㓳ᰦ䰤Ǆ䇝㓳ᰦ䰤ਟ㜭Ӿࠐ。䫏ࡠࠐሿᰦˈ䘉㾱ⴻ㖁㔌ѝᵳ٬Ⲵᮠ䟿ǃ
㾱㘳㲁Ⲵ䇝㓳ᇎֻⲴᮠ䟿ǃԕ৺н਼ᆖҐ㇇⌅৲ᮠⲴ䇮㖞ㅹഐ㍐Ǆ
x ਟ㜭䴰㾱ᘛ䙏≲ࠪⴞḷ࠭ᮠ٬Ǆቭ㇑ ANN ⲴᆖҐᰦ䰤⴨ሩ䖳䮯ˈնሩᆖҐ
Ⲵ㖁㔌≲٬ˈԕ‫ׯ‬ᢺ㖁㔌ᓄ⭘ࡠਾ㔝Ⲵᇎֻˈ䙊ᑨᱟ䶎ᑨᘛ䙏ⲴǄֻྲˈ
ALVINN ൘䖖䖶ੁࡽ㹼傦ᰦˈ⇿。ᓄ⭘ᆳⲴ⾎㓿㖁㔌㤕ᒢ⅑ˈԕнᯝൠᴤᯠ
傮傦ᯩੁǄ
x Ӫ㊫㜭੖⨶䀓ᆖࡠⲴⴞḷ࠭ᮠᱟн䟽㾱ⲴǄ⾎㓿㖁㔌ᯩ⌅ᆖҐࡠⲴᵳ٬㓿ᑨ
ᱟӪ㊫䳮ԕ䀓䟺ⲴǄᆖࡠⲴ⾎㓿㖁㔌∄ᆖࡠⲴ㿴ࡉ䳮ҾՐ䗮㔉Ӫ㊫Ǆ
䘉аㄐⲴަ։䜘࠶ᱟ䘉ṧ㓴㓷Ⲵ˖ᡁԜ‫ݸ‬䇘䇪䇝㓳অњঅ‫Ⲵݳ‬ᆖҐ㇇⌅ˈ਼ᰦӻ㓽㓴ᡀ
⾎㓿㖁㔌Ⲵࠐ⿽ѫ㾱অ‫ˈݳ‬वᤜᝏ⸕ಘ˄perceptron˅ǃ㓯ᙗঅ‫˄ݳ‬linear unit˅઼ sigmoid অ
‫˄ݳ‬sigmoid unit˅Ǆ❦ਾ㔉ࠪ䇝㓳䘉Ӌঅ‫ݳ‬㓴ᡀⲴཊቲ㖁㔌Ⲵ৽ੁՐ᫝㇇⌅ˈᒦ㘳㲁ࠐња
㡜ᙗⲴ䰞仈ˈ∄ྲ ANN Ⲵ㺘ᖱ㜭࣋ǃ‫ٷ‬䇮オ䰤ᩌ㍒Ⲵᵜ䍘⢩ᖱǃ䗷ᓖᤏਸ䰞仈ǃԕ৺৽ੁ
Ր᫝㇇⌅ⲴਈփǄᵜㄐҏ㔉ࠪҶањᓄ⭘৽ੁՐ᫝㇇⌅䇶࡛Ӫ㝨Ⲵ䈖㓶ֻᆀˈᒦᤷሬ䈫㘵ྲ
օਆᗇ䘉њֻᆀⲴᮠᦞ઼ԓ⸱ˈᒦ䘋а↕ᇎ傼䘉њᓄ⭘Ǆ
4.4 ᝕⸛ಞ
а⿽㊫රⲴANN㌫㔏ᱟԕ㻛〠Ѫᝏ⸕ಘ˄perceptron˅Ⲵঅ‫ݳ‬Ѫส⹰Ⲵˈྲമ 4-2 ᡰ⽪Ǆ

ᝏ⸕ಘԕањᇎᮠ٬ੁ䟿֌Ѫ䗃‫ˈޕ‬䇑㇇䘉Ӌ䗃‫Ⲵޕ‬㓯ᙗ㓴ਸˈ❦ਾྲ᷌㔃᷌བྷҾḀњ䰸٬
ቡ䗃ࠪ 1ˈ੖ࡉ䗃ࠪ-1Ǆᴤ㋮⺞ൠˈྲ᷌䗃‫ޕ‬Ѫx1ࡠxnˈ䛓Ѹᝏ⸕ಘ䇑㇇Ⲵ䗃ࠪѪ˖
1 if w0 w1 x1 w2 x 2 wn x n ! 0
o( x1 ,, x n ) ®
¯ 1 otherwise
ަѝ⇿ањwiᱟањᇎᮠᑨ䟿ˈᡆਛ‫ڊ‬ᵳ٬˄weight˅ˈ⭘ᶕߣᇊ䗃‫ޕ‬xiሩᝏ⸕ಘ䗃ࠪⲴ
䍑 ⥞ ⦷ Ǆ 䈧 ⌘ ᜿ ˈ ᑨ 䟿 (-w0) ᱟ а њ 䰸 ٬ ˈ ᆳ ᱟ Ѫ Ҷ ֯ ᝏ ⸕ ಘ 䗃 ࠪ 1 ˈ 䗃 ‫ ࣐ Ⲵ ޕ‬ᵳ ઼
w1 x1 w2 x 2 wn x n ᗵ享䎵䗷Ⲵ䰸٬Ǆ
മ 4-2 ᝏ⸕ಘ
ѪҶㆰॆ㺘⽪ˈᡁԜ‫ٷ‬ᜣᴹањ䱴࣐Ⲵᑨ䟿䗃‫ޕ‬x0=1ˈ䛓ѸᡁԜቡਟԕᢺк䗩Ⲵнㅹᔿ
& &
¦
n
߉Ѫ i 0
wi xi ! 0 ˈᡆԕੁ䟿ᖒᔿ߉Ѫ w x ! 0 ǄѪҶㆰ⸝䎧㿱ˈᡁԜᴹᰦՊᢺᝏ⸕ಘ࠭
ᮠ߉Ѫ˖
& & &
o( x ) sgn( w x )
ަѝˈ
1 if y ! 0
sgn( y ) ®
¯ 1 otherwise
ᆖҐањᝏ⸕ಘ᜿ણ⵰䘹ᤙᵳw0, …, wnⲴ٬Ǆᡰԕᝏ⸕ಘᆖҐ㾱㘳㲁Ⲵ‫ى‬䘹‫ٷ‬䇮オ䰤H
ቡᱟᡰᴹਟ㜭Ⲵᇎᮠ٬ᵳੁ䟿Ⲵ䳶ਸǄ
& &
H {w | w ( n 1) }
4.4.1 ᝕⸛ಞⲺ㺞ᖷ㜳࣑
ᡁԜਟԕᢺᝏ⸕ಘⴻ֌ᱟ n 㔤ᇎֻオ䰤˄ণ⛩オ䰤˅ѝⲴ䎵ᒣ䶒ߣㆆ䶒ǄሩҾ䎵ᒣ䶒а
‫Ⲵח‬ᇎֻˈᝏ⸕ಘ䗃ࠪ 1ˈሩҾਖа‫Ⲵח‬ᇎֻ䗃ࠪ-1ˈྲമ 4-3 ᡰ⽪Ǆ䘉њߣㆆ䎵ᒣ䶒ᯩ〻
& &
ᱟ w x 0 Ǆᖃ❦ˈḀӋ↓৽ṧֻ䳶ਸнਟ㜭㻛ԫа䎵ᒣ䶒࠶ࢢǄ䛓Ӌਟԕ㻛࠶ࢢⲴ〠Ѫ
㓯ᙗਟ࠶˄linearly separable˅ṧֻ䳶ਸǄ
മ 4-3 є䗃‫ޕ‬ᝏ⸕ಘ㺘⽪Ⲵߣㆆ䶒
˄a˅а㓴䇝㓳ṧ઼ֻањ㜭↓⺞࠶㊫䘉ӋṧֻⲴᝏ⸕ಘߣㆆ䶒Ǆ˄b˅а㓴䶎㓯ᙗਟ࠶Ⲵ䇝㓳ṧֻ
˄ҏቡᱟн㜭⭘ԫаⴤ㓯↓⺞࠶㊫Ⲵṧֻ˅Ǆx1઼x2ᱟᝏ⸕ಘⲴ䗃‫ޕ‬ǄĀ+ā㺘⽪↓ֻˈĀ-ā㺘⽪
৽ֻǄ
অ⤜Ⲵᝏ⸕ಘਟԕ⭘ᶕ㺘⽪ᖸཊᐳቄ࠭ᮠǄֻྲˈ‫ٷ‬ᇊ⭘ 1˄ⵏ˅઼-1˄‫˅ٷ‬㺘⽪ᐳቄ
٬ˈ䛓Ѹ֯⭘ањє䗃‫Ⲵޕ‬ᝏ⸕ಘᶕᇎ⧠о࠭ᮠ˄AND˅Ⲵа⿽ᯩ⌅ᱟ䇮㖞ᵳw0= -0.8 ᒦф
w1=w2=0.5Ǆྲ᷌⭘䘉њᝏ⸕ಘᶕ㺘⽪ᡆ࠭ᮠ˄OR˅ˈ䛓Ѹਚ㾱᭩ਈᆳⲴ䰸٬w0=-0.3Ǆһᇎ
кˈAND઼ORਟ㻛ⴻ֌m-of-n࠭ᮠⲴ⢩ֻ˖ҏቡᱟ㾱֯࠭ᮠ䗃ࠪѪⵏˈ䛓Ѹᝏ⸕ಘⲴnњ䗃
‫ޕ‬ѝ㠣ቁmњᗵ享ѪⵏǄOR࠭ᮠሩᓄҾm=1ˈAND࠭ᮠሩᓄҾm=n.Ǆԫ᜿m-of-n࠭ᮠਟԕᖸ
ᇩ᱃ൠ⭘ᝏ⸕ಘ㺘⽪ˈਚ㾱䇮㖞ᡰᴹ䗃‫Ⲵޕ‬ᵳѪ਼ṧⲴ٬˄ྲ 0.5˅ˈ❦ਾᦞ↔ᚠᖃൠ䇮㖞䰸
٬Ǆ
ᝏ⸕ಘਟԕ㺘⽪ᡰᴹⲴ৏࿻ᐳቄ࠭ᮠ˄primitive boolean function˅˖оǃᡆǃо䶎˄NAND˅

઼ᡆ䶎˄NOR˅Ǆ❦㘼нᒨⲴᱟˈаӋᐳቄ࠭ᮠᰐ⌅⭘অаⲴᝏ⸕ಘ㺘⽪ˈֻྲᔲᡆ࠭ᮠ
˄XOR˅ˈᆳᖃфӵᖃx1zx2ᰦ䗃ࠪѪ 1Ǆ䈧⌘᜿മ 4-3˄b˅ѝ㓯ᙗнਟ࠶Ⲵ䇝㓳ṧᵜ䳶ሩᓄ
Ҿᔲᡆ࠭ᮠǄ
ᝏ⸕ಘ㺘⽪оǃᡆǃо䶎ǃᡆ䶎Ⲵ㜭࣋ᱟᖸ䟽㾱ⲴˈഐѪᡰᴹⲴᐳቄ࠭ᮠ䜭ਟ㺘⽪Ѫส
Ҿ䘉Ӌ৏࿻࠭ᮠⲴӂ䘎অ‫Ⲵݳ‬Ḁњ㖁㔌Ǆһᇎкˈӵ⭘єቲ␡ᓖⲴᝏ⸕ಘ㖁㔌ቡਟԕ㺘⽪ᡰ
ᴹⲴᐳቄ࠭ᮠˈ൘䘉Ӌ㖁㔌ѝ䗃‫ޕ‬㻛䘱ࡠཊњঅ‫ˈݳ‬䘉Ӌঅ‫Ⲵݳ‬䗃ࠪ㻛䗃‫ࡠޕ‬ㅜҼ㓗ˈҏᱟ
ᴰਾа㓗Ǆа⿽ᯩ⌅ᱟ⭘᷀ਆ㤳ᔿ˄disjunctive normal form˅˄ҏቡᱟሩ䗃‫઼ޕ‬ᆳԜⲴ੖ᇊ
Ⲵ‫ݸ‬䘋㹼ਸਆˈ޽ሩ䘉㓴ਸਆᔿ䘋㹼᷀ਆ˅ᶕ㺘⽪ᐳቄ࠭ᮠǄ⌘᜿ˈ㾱ᢺањ AND ᝏ⸕ಘ
Ⲵ䗃‫≲ޕ‬੖ᇊˈਚ㾱ㆰঅൠ᭩ਈ⴨ᓄ䗃‫ޕ‬ᵳⲴㅖਧǄ
ഐѪ䰸٬অ‫Ⲵݳ‬㖁㔌ਟԕ㺘⽪བྷ䟿Ⲵ࠭ᮠˈ㘼অ⤜Ⲵঅ‫ݳ‬н㜭‫ࡠڊ‬䘉а⛩ˈᡰԕ䙊ᑨᡁ
Ԝᝏ‫ޤ‬䏓ⲴᱟᆖҐ䰸٬অ‫ݳ‬㓴ᡀⲴཊቲ㖁㔌Ǆ
4.4.2 ᝕⸛ಞ䇣㓹⌋ࡏ
㲭❦ᡁԜⲴⴞⲴᱟᆖҐ⭡ཊњঅ‫ݳ‬ӂ䘎Ⲵ㖁㔌ˈնᡁԜ䘈ᱟӾྲօᆖҐঅњᝏ⸕ಘⲴᵳ
٬ᔰ࿻Ǆ߶⺞ൠ䈤ˈ䘉䟼ⲴᆖҐԫ࣑ᱟߣᇊањᵳੁ䟿ˈᆳਟԕ֯ᝏ⸕ಘሩҾ㔉ᇊⲴ䇝㓳ṧ
ֻ䗃ࠪ↓⺞Ⲵ 1 ᡆ-1Ǆ
ᐢ㓿⸕䚃ᴹࠐ⿽䀓ߣ䘉њᆖҐԫ࣑Ⲵ㇇⌅Ǆ䘉䟼ᡁԜ㘳㲁є⿽˖ᝏ⸕ಘ⌅ࡉ઼ delta ⌅
ࡉ˄delta rule˅˄ᱟㅜ 1 ㄐѝ⭘ᶕᆖҐ䇴ՠ࠭ᮠⲴᴰሿ൷ᯩ⌅ LMS Ⲵањਈփ˅Ǆ䘉є⿽㇇
⌅‫؍‬䇱᭦ᮋࡠਟ᧕ਇⲴ‫ٷ‬䇮ˈ൘н਼ⲴᶑԦл᭦ᮋࡠⲴ‫ٷ‬䇮⮕ᴹн਼Ǆ䘉є⿽ᯩ⌅ሩҾ ANN
ᱟᖸ䟽㾱ⲴˈഐѪᆳԜᨀ‫׋‬ҶᆖҐཊњঅ‫ݳ‬ᶴᡀⲴ㖁㔌Ⲵส⹰Ǆ
Ѫᗇࡠਟ᧕ਇⲴᵳੁ䟿ˈа⿽࣎⌅ᱟӾ䲿ᵪⲴᵳ٬ᔰ࿻ˈ❦ਾ৽༽ൠᓄ⭘䘉њᝏ⸕ಘࡠ
⇿њ䇝㓳ṧֻˈਚ㾱ᆳ䈟࠶㊫ṧֻቡ‫؞‬᭩ᝏ⸕ಘⲴᵳ٬Ǆ䟽༽䘉њ䗷〻ˈⴤࡠᝏ⸕ಘ↓⺞࠶
㊫ᡰᴹⲴ䇝㓳ṧֻǄ⇿а↕ṩᦞᝏ⸕ಘ䇝㓳⌅ࡉ˄perceptron training rule˅ᶕ‫؞‬᭩ᵳ٬ˈҏ
ቡᱟṩᦞл䶒Ⲵ⌅ࡉ‫؞‬᭩о䗃‫ޕ‬xiሩᓄⲴᵳwi:
wimwi+'wi
ަѝ
'wi =K(t-o)xi
䘉䟼 t ᱟᖃࡽ䇝㓳ṧֻⲴⴞḷ䗃ࠪˈo ᱟᝏ⸕ಘⲴ䗃ࠪˈKᱟањ↓Ⲵᑨᮠ〠ѪᆖҐ䙏
⦷˄learning rate˅ǄᆖҐ䙏⦷Ⲵ֌⭘ᱟ㕃઼⇿а↕䈳ᮤᵳⲴ〻ᓖǄᆳ䙊ᑨ㻛䇮ѪањሿⲴᮠ
٬˄ֻྲ 0.1˅ˈ㘼фᴹᰦՊ֯ަ䲿⵰ᵳ䈳ᮤ⅑ᮠⲴ໎࣐㘼㺠߿Ǆ
ѪӰѸ䘉њᴤᯠ⌅ࡉՊᡀ࣏᭦ᮋࡠ↓⺞Ⲵᵳ٬઒˛ѪҶᗇࡠⴤ㿲Ⲵᝏ㿹ˈ㘳㲁аӋ⢩
ֻǄ‫ٷ‬ᇊ䇝㓳ṧᵜᐢ㻛ᝏ⸕ಘ↓⺞࠶㊫Ǆ䘉ᰦˈ(t-o)ᱟ 0ˈ䘉֯'wiѪ 0ˈᡰԕ⋑ᴹᵳ٬㻛‫؞‬
᭩Ǆ㘼ྲ᷌ᖃⴞḷ䗃ࠪᱟ+1 ᰦᝏ⸕ಘ䗃ࠪањ-1ˈ䘉⿽ᛵߥѪ֯ᝏ⸕ಘ䗃ࠪањ+1 㘼нᱟ
& &
-1ˈᵳ٬ᗵ享㻛‫؞‬᭩ԕ໎བྷ w x Ⲵ٬Ǆֻྲˈྲ᷌xi>0ˈ䛓Ѹ໎བྷwiՊ֯ᝏ⸕ಘᴤ᧕䘁↓⺞
࠶㊫䘉њᇎֻǄ⌘᜿䘉⿽ᛵߥл䇝㓳⌅ࡉՊ໎䮯wiˈഐѪ(t-o)ˈK઼xi䜭ᱟ↓ⲴǄֻྲˈྲ᷌
xi=0.8ˈK=0.1ˈt=1ˈᒦфo= -1ˈ䛓Ѹᵳᴤᯠቡᱟ'wi =K(t-o)xi=0.1(1-(-1))0.8=0.16Ǆਖаᯩ
䶒ˈྲ᷌t=-1 㘼o=1ˈ䛓Ѹ઼↓Ⲵxi‫ޣ‬㚄Ⲵᵳ٬Պ㻛߿ሿ㘼нᱟ໎བྷǄ
һᇎкਟԕ䇱᰾ˈ൘ᴹ䲀⅑֯⭘ᝏ⸕ಘ䇝㓳⌅ࡉਾˈк䶒Ⲵ䇝㓳䗷〻Պ᭦ᮋࡠањ㜭↓
⺞࠶㊫ᡰᴹ䇝㓳ṧֻⲴᵳੁ䟿ˈࡽᨀᱟ䇝㓳ṧֻ㓯ᙗਟ࠶ˈᒦф֯⭘Ҷ‫࠶ݵ‬ሿⲴK ˄৲㿱
Minskey & Papert 1969˅Ǆྲ᷌ᮠᦞнᱟ㓯ᙗਟ࠶Ⲵˈ䛓Ѹн㜭‫؍‬䇱᭦ᮋǄ
4.4.3 ụᓜс䲃ૂ delta ⌋ࡏ
ቭ㇑ᖃ䇝㓳ṧֻ㓯ᙗਟ࠶ᰦˈᝏ⸕ಘ⌅ࡉਟԕᡀ࣏ൠ᢮ࡠањᵳੁ䟿ˈնྲ᷌ṧֻнᱟ
㓯ᙗਟ࠶ᰦᆳሶн㜭᭦ᮋǄഐ↔ˈӪԜ䇮䇑Ҷਖањ䇝㓳⌅ࡉᶕ‫ݻ‬ᴽ䘉њн䏣ˈ〠Ѫ delta
⌅ࡉ˄delta rule˅Ǆྲ᷌䇝㓳ṧᵜнᱟ㓯ᙗਟ࠶Ⲵˈ䛓Ѹ delta ⌅ࡉՊ᭦ᮋࡠⴞḷᾲᘥⲴᴰ֣
䘁լǄ(how to converge?)
delta ⌅ࡉⲴ‫ޣ‬䭞ᙍᜣᱟ֯⭘ởᓖл䱽˄gradient descent˅ᶕᩌ㍒ਟ㜭ᵳੁ䟿Ⲵ‫ٷ‬䇮オ䰤ˈ

ԕ᢮ࡠᴰ֣ᤏਸ䇝㓳ṧֻⲴᵳੁ䟿Ǆ䘉њ⌅ࡉᖸ䟽㾱ˈഐѪᆳᨀ‫׋‬Ҷ৽ੁՐ᫝㇇⌅Ⲵส⹰ˈ
㘼৽ੁՐ᫝㇇⌅㜭ཏᆖҐཊњঅ‫Ⲵݳ‬ӂ䘎㖁㔌Ǆ䘉њ⌅ࡉ䟽㾱ᙗⲴਖањ৏ഐᱟˈሩҾवਜ਼
ཊ⿽н਼㊫රⲴ䘎㔝৲ᮠॆ‫ٷ‬䇮Ⲵ‫ٷ‬䇮オ䰤ˈởᓖл䱽ᱟᗵ享䙽শ䘉ṧⲴ‫ٷ‬䇮オ䰤Ⲵᡰᴹᆖ
Ґ㇇⌅Ⲵส⹰Ǆ
ᴰྭᢺ delta 䇝㓳⌅ࡉ⨶䀓Ѫ䇝㓳ањᰐ䰸٬Ⲵᝏ⸕ಘˈҏቡᱟањ㓯ᙗঅ‫˄ݳ‬linear

unit˅ˈᆳⲴ䗃ࠪ o ྲл˖
& & &
o( x ) w x (4.1)
Ҿᱟˈањ㓯ᙗঅ‫ݳ‬ሩᓄҾᝏ⸕ಘⲴㅜа䱦⇥ˈнᑖᴹ䰸٬Ǆ
ѪҶ᧘ሬ㓯ᙗঅ‫Ⲵݳ‬ᵳ٬ᆖҐ⌅ࡉˈ‫ݸ‬ᤷᇊањᓖ䟿ḷ߶ᶕ㺑䟿‫ٷ‬䇮˄ᵳੁ䟿˅⴨ሩҾ
䇝㓳ṧֻⲴ䇝㓳䈟ᐞ˄training error˅
Ǆቭ㇑ᴹᖸཊ࣎⌅ᇊѹ䘉њ䈟ᐞˈањᑨ⭘Ⲵ⢩࡛ᯩ‫ׯ‬
Ⲵᓖ䟿ḷ߶Ѫ˖
& 1
E ( w) ¦
2 dD
(t d od ) 2 ˄4.2˅
ަѝDᱟ䇝㓳ṧֻ䳶ਸˈtdᱟ䇝㓳ṧֻdⲴⴞḷ䗃ࠪˈodᱟ㓯ᙗঅ‫ݳ‬ሩ䇝㓳ṧֻdⲴ䗃ࠪǄ
&
൘䘉њᇊѹѝˈ E (w) ᱟⴞḷ䗃ࠪtd઼㓯ᙗঅ‫ݳ‬䗃ࠪodⲴᐞᔲⲴᒣᯩ൘ᡰᴹⲴ䇝㓳ṧֻк≲
&
઼ਾ޽䲔ԕ 2Ǆ䘉䟼ᡁԜᢺEᇊѪ w Ⲵ࠭ᮠˈᱟഐѪ㓯ᙗঅ‫Ⲵݳ‬䗃ࠪo‫׍‬䎆Ҿ䘉њᵳੁ䟿Ǆᖃ
❦Eҏ‫׍‬䎆Ҿ⢩ᇊⲴ䇝㓳ṧֻ䳶ਸˈնᡁԜ䇔ѪᆳԜ൘䇝㓳ᵏ䰤ᱟപᇊⲴˈᡰԕнᗵ哫✖ൠ
ᢺE߉Ѫ䇝㓳ṧֻⲴ࠭ᮠǄㅜ 6 ㄐ㔉ࠪҶ䘹ᤙ䘉⿽EᇊѹⲴа⿽䍍ਦᯟ䇪䇱Ǆ⺞࠷ൠ䇢ˈ൘
䛓䟼ᡁԜᤷࠪҶ൘аᇊᶑԦлˈሩҾ㔉ᇊⲴ䇝㓳ᮠᦞ֯EᴰሿॆⲴ‫ٷ‬䇮ҏቡᱟHѝᴰਟ㜭Ⲵ
‫ٷ‬䇮Ǆ
4.4.3.1 ᖘ䊗ौ‫ٽ‬䇴グ䰪
ѪҶ⨶䀓ởᓖл䱽㇇⌅ˈᖒ䊑ൠ㺘⽪ᮤњ‫ٷ‬䇮オ䰤ᱟᴹᑞࣙⲴˈമ 4-4 ⭫ࠪҶवਜ਼ਟ㜭

ᵳੁ䟿Ⲵᮤњ‫ٷ‬䇮オ䰤઼ооᆳԜ⴨‫ޣ‬㚄ⲴE٬Ǆ䘉䟼ˈ඀ḷ䖤w0ˈw1㺘⽪ањㆰঅⲴ㓯ᙗ
অ‫ݳ‬ѝєњᵳⲴਟ㜭Ⲵਆ٬Ǆ㓥䖤ᤷࠪ⴨ሩҾḀപᇊⲴ䇝㓳ṧֻⲴ䈟ᐞEǄഐ↔മѝⲴ䈟ᐞ
ᴢ䶒ᾲᤜҶ‫ٷ‬䇮オ䰤ѝ⇿ањᵳੁ䟿ⲴԱᵋᓖ˄desirability˅
˄ᡁԜԱᵋᗇࡠањާᴹᴰሿ䈟
ᐞⲴ‫ٷ‬䇮˅Ǆྲ᷌㔉ᇊҶ⭘ᶕᇊѹEⲴᯩ⌅ˈ䛓ѸሩҾ㓯ᙗঅ‫ˈݳ‬䘉њ䈟ᐞᴢ䶒ᗵ❦ᱟާᴹ
অа‫ޘ‬ተᴰሿ٬Ⲵᣋ⢙䶒Ǆᖃ❦ˈާփⲴᣋ⢙䶒ᖒ⣦‫׍‬䎆ҾާփⲴ䇝㓳ṧֻ䳶ਸǄ
മ 4-4 н਼‫ٷ‬䇮Ⲵ䈟ᐞ
ሩҾᴹєњᵳ٬Ⲵ㓯ᙗঅ‫ٷˈݳ‬䇮オ䰤Hቡᱟw0,w1ᒣ䶒Ǆ㓥䖤㺘⽪оപᇊⲴ䇝㓳ṧֻ䳶ਸ⴨ᓄⲴᵳ
ੁ䟿‫ٷ‬䇮Ⲵ䈟ᐞǄ㇝ཤᱮ⽪Ҷ䈕⛩ởᓖⲴ⴨৽ᯩੁˈᤷࠪҶ൘w0ˈw1ᒣ䶒ѝ⋯䈟ᐞᴢ䶒ᴰ䲑ጝл䱽
ⲴᯩੁǄ
ởᓖл䱽ᩌ㍒⺞ᇊањ֯ E ᴰሿॆⲴᵳੁ䟿Ⲵᯩ⌅ᱟӾањԫ᜿Ⲵࡍ࿻ᵳੁ䟿ᔰ࿻ˈ
❦ਾԕᖸሿⲴ↕Հ৽༽‫؞‬᭩䘉њੁ䟿Ǆ൘⇿а↕ˈ᤹➗⋯䈟ᐞᴢ䶒ӗ⭏ᴰ䲑ጝл䱽Ⲵᯩੁ‫؞‬
᭩ᵳੁ䟿˄৲㿱മ 4-4˅
Ǆ㔗㔝䘉њ䗷〻ⴤࡠࡠ䗮‫ޘ‬ተⲴᴰሿ䈟ᐞǄ
4.4.3.2 ụᓜс䲃⌋ࡏⲺ᧞ሲ
&
ᡁԜᘾṧ㜭䇑㇇ࠪ⋯䈟ᐞᴢ䶒ᴰ䲑ጝл䱽Ⲵᯩੁ઒˛ਟԕ䙊䗷䇑㇇ E ⴨ሩੁ䟿 w Ⲵ⇿
&
њ࠶䟿Ⲵሬᮠᶕᗇࡠ䘉њᯩੁǄ䘉њੁ䟿ሬᮠ㻛〠Ѫ E ሩҾ w Ⲵởᓖ˄gradient˅ˈ䇠֌
&
E (w) Ǆ
& ª wE wE wE º
E ( w) « , , , » ˄4.3˅
¬ ww0 ww1 wwn ¼
&
⌘᜿ E (w) ᵜ䓛ᱟањੁ䟿ˈᆳⲴᡀઈᱟEሩ⇿њwiⲴ‫ٿ‬ሬᮠǄ当梯度被解释为权空间
的一个向量时，它确定了使E最陡峭上升的方向。ᡰԕ䘉њੁ䟿Ⲵ৽ᯩੁ㔉ࠪҶᴰ䲑ጝл䱽
&
ⲴᯩੁǄֻྲˈമ 4-4 ѝⲴ㇝ཤᱮ⽪Ҷw0ˈw1ᒣ䶒Ⲵањ⢩ᇊ⛩Ⲵ䍏ởᓖ E (w) Ǆ
ᰒ❦ởᓖ⺞ᇊҶ E ᴰ䲑ጝкॷⲴᯩੁˈ䛓Ѹởᓖл䱽Ⲵ䇝㓳⌅ࡉᱟ˖
& & &
w m w 'w
ަѝ
& &
'w K E (w) ˄4.4˅
䘉䟼Kᱟањ↓Ⲵᑨᮠਛ‫ڊ‬ᆖҐ䙏⦷ˈᆳߣᇊởᓖл䱽ᩌ㍒ѝⲴ↕䮯ǄަѝⲴ䍏ਧᱟഐ
ѪᡁԜᜣ㾱䇙ᵳੁ䟿ੁ E л䱽Ⲵᯩੁ〫ࣘǄ䘉њ䇝㓳⌅ࡉҏਟԕ߉ᡀᆳⲴ࠶䟿ᖒᔿ˖
wimwi+'wi
ަѝ
wE
'wi K ˄4.5˅
wwi
wE &
䘉ṧᖸ␵ᾊˈᴰ䲑ጝⲴл䱽ਟԕ䙊䗷᤹∄ֻ ᭩ਈ w Ⲵ⇿а࠶䟿wiᶕᇎ⧠Ǆ
wwi
㾱ᖒᡀањṩᦞㅹᔿ˄4.5˅䘝ԓᴤᯠᵳⲴᇎ⭘㇇⌅ˈᡁԜ䴰㾱ањ儈᭸Ⲵᯩ⌅൘⇿а
↕䇑㇇䘉њởᓖǄᒨ䘀Ⲵᱟˈ䇑㇇䗷〻ᒦнഠ䳮ǄᡁԜਟԕӾ‫ޜ‬ᔿ˄4.2˅ѝ䇑㇇ E Ⲵᗞ࠶ˈ
wE
Ӿ㘼ᗇࡠ㓴ᡀ䘉њởᓖੁ䟿Ⲵ࠶䟿 Ǆ䗷〻ྲл˖
wwi
wE w 1
wwi
¦
wwi 2 dD
(t d od ) 2
1 w
¦
2 dD wwi
(t d od ) 2
1 w
¦
2 dD
2(t d od )
wwi
(t d od )
w & &
¦ (t
d D
d od )
wwi
(t d w x d )
wE
wwi
¦ (t
d D
d o d )( xid ) ˄4.6˅
ަѝxid㺘⽪䇝㓳ṧֻdⲴањ䗃‫࠶ޕ‬䟿xiǄ⧠൘ᡁԜᴹҶањㅹᔿˈ㜭ཏ⭘㓯ᙗঅ‫Ⲵݳ‬
wE
䗃‫ޕ‬xidǃ䗃ࠪodǃԕ৺䇝㓳ṧֻⲴⴞḷ٬td㺘⽪ Ǆᢺㅹᔿ˄4.6˅ԓ‫ޕ‬ㅹᔿ˄4.5˅‫ׯ‬ᗇࡠ
wwi
Ҷởᓖл䱽ᵳ٬ᴤᯠ⌅ࡉǄ
'wi K ¦ (t d od ) xid ˄4.7˅

d D
ᾲ㘼䀰ѻˈ䇝㓳㓯ᙗঅ‫Ⲵݳ‬ởᓖл䱽㇇⌅ྲл˖䘹ਆањࡍ࿻Ⲵ䲿ᵪᵳੁ䟿˗ᓄ⭘㓯ᙗ
অ‫ࡠݳ‬ᡰᴹⲴ䇝㓳ṧֻˈ❦ਾṩᦞ‫ޜ‬ᔿ˄4.7˅䇑㇇⇿њᵳ٬Ⲵ'wi˗䙊䗷࣐к'wiᶕᴤᯠ⇿
њᵳ٬ˈ❦ਾ䟽༽䘉њ䗷〻Ǆ䘉њ㇇⌅㻛ᖂ㓣൘㺘˄4.1˅ѝǄഐѪ䈟ᐞᴢ䶒ӵवਜ਼ањ‫ޘ‬
ተⲴᴰሿ٬ˈᡰԕᰐ䇪䇝㓳ṧᵜᱟ੖㓯ᙗਟ࠶ˈ䘉њ㇇⌅Պ᭦ᮋࡠާᴹᴰሿ䈟ᐞⲴᵳੁ䟿ˈ
ᶑԦᱟᗵ享֯⭘ањ䏣ཏሿⲴᆖҐ䙏⦷KǄྲ᷌Kཚབྷˈởᓖл䱽ᩌ㍒ቡᴹ䎺䗷䈟ᐞᴢ䶒ᴰ
ሿ٬Ⲵড䲙ˈ㘼нᱟ‫⮉ڌ‬൘䛓а⛩Ǆഐ↔ˈሩ↔㇇⌅Ⲵа⿽ᑨ⭘Ⲵ᭩䘋ᱟ䲿⵰ởᓖл䱽↕ᮠ
Ⲵ໎࣐䙀⑀߿ሿKⲴ٬Ǆ
㺘 4-1 䇝㓳㓯ᙗঅ‫Ⲵݳ‬ởᓖл䱽㇇⌅
㾱ᇎ⧠ởᓖл䱽Ⲵ䲿ᵪ䘁լˈࡐ䲔ᔿ(T4.2)ˈᒦᢺᔿ(T4.1)ᴯᦒѪwi mwi +K(t-o)xiǄ
Gradient-Descent(training_examples, K)
& &
training_examples ѝ⇿ањ䇝㓳ṧֻᖒᔿѪᒿ‫ <ڦ‬x , t>ˈަѝ x ᱟ䗃‫ޕ‬٬ੁ䟿ˈt ᱟⴞ
ḷ䗃ࠪ٬ǄKᱟᆖҐ䙏⦷˄ֻྲ 0.05˅ Ǆ
z ࡍ࿻ॆ⇿њwiѪḀњሿⲴ䲿ᵪ٬
z 䙷ࡠ㓸→ᶑԦѻࡽˈ‫ڊ‬ԕл᫽֌˖
z ࡍ࿻ॆ⇿њ'wiѪ 0
&
z ሩҾ䇝㓳ṧֻ training_examples ѝⲴ⇿њ< x , t>ˈ‫˖ڊ‬
&
z ᢺᇎֻ x 䗃‫↔ࡠޕ‬অ‫ˈݳ‬䇑㇇䗃ࠪ o
z ሩҾ㓯ᙗঅ‫⇿Ⲵݳ‬њᵳwiˈ‫ڊ‬
'wi m'wi +K(t-o)xi ˄T4.1˅
z ሩҾ㓯ᙗঅ‫⇿Ⲵݳ‬њᵳwiˈ‫ڊ‬
wim wi +'wi ˄T4.2˅
4.4.3.3 ụᓜс䲃Ⲻ䳅ᵰ䘇ղ
ởᓖл䱽ᱟа⿽䟽㾱Ⲵ䙊⭘ᆖҐ㤳රǄᆳᱟᩌ㍒ᓎབྷ‫ٷ‬䇮オ䰤ᡆᰐ䲀‫ٷ‬䇮オ䰤Ⲵа⿽ㆆ
⮕ˈᆳਟᓄ⭘Ҿ┑䏣ԕлᶑԦⲴԫօᛵߥ˖˄1˅‫ٷ‬䇮オ䰤वਜ਼䘎㔝৲ᮠॆⲴ‫ٷ‬䇮˄ֻྲˈа
њ㓯ᙗঅ‫Ⲵݳ‬ᵳ٬˅˗˄2˅䈟ᐞሩҾ䘉Ӌ‫ٷ‬䇮৲ᮠਟᗞǄᓄ⭘ởᓖл䱽Ⲵѫ㾱ᇎ䐥䰞仈ᱟ˖
˄1˅ᴹᰦ᭦ᮋ䗷〻ਟ㜭䶎ᑨធ˄ᆳਟ㜭䴰㾱ᮠॳ↕Ⲵởᓖл䱽˅˗˄2˅ྲ᷌൘䈟ᐞᴢ䶒кᴹ
ཊњተ䜘ᶱሿ٬ˈ䛓Ѹн㜭‫؍‬䇱䘉њ䗷〻Պ᢮ࡠ‫ޘ‬ተᴰሿ٬Ǆ
㕃䀓䘉Ӌഠ䳮Ⲵањᑨ㿱Ⲵởᓖл䱽ਈփ㻛〠Ѫ໎䟿ởᓖл䱽˄incremental gradient
descent˅ˈᡆ䲿ᵪởᓖл䱽˄stochastic gradient descent˅Ǆ䢤Ҿ
‫ޜ‬ᔿ˄4.7˅㔉ࠪⲴởᓖл䱽䇝㓳⌅ࡉ൘ሩ D ѝⲴᡰᴹ䇝㓳ṧֻ≲઼ਾ䇑㇇ᵳ٬ᴤᯠˈ
䲿ᵪởᓖл䱽Ⲵᙍᜣᱟṩᦞ⇿њঅ⤜ṧֻⲴ䈟ᐞ໎䟿ൠ䇑㇇ᵳ٬ᴤᯠˈᗇࡠ䘁լⲴởᓖл䱽
ᩌ㍒Ǆ‫؞‬᭩ਾⲴ䇝㓳⌅ࡉо‫ޜ‬ᔿ˄4.7˅㔉ࠪⲴ⴨լˈਚᱟ൘䘝ԓ䇑㇇⇿њ䇝㓳ṧֻᰦṩᦞ
л䶒Ⲵ‫ޜ‬ᔿᶕᴤᯠᵳ٬
'wi =K(t-o)xi ˄4.10˅
ަѝtˈoˈ઼xi࠶࡛ᱟⴞḷ٬ǃঅ‫ݳ‬䗃઼ࠪㅜiњ䇝㓳ṧֻⲴ䗃‫ޕ‬Ǆ㾱‫؞‬᭩㺘 4-1 Ⲵởᓖ

л䱽㇇⌅ˈਚ㾱ㆰঅൠࡐ䲔˄T4.2˅ᔿᒦᢺᔿ˄T4.1˅ᴯᦒѪwi mwi +K(t-o)xiǄⴻᖵ䲿ᵪở
&
ᓖл䱽Ⲵа⿽ᯩ⌅ᱟ㘳㲁Ѫ⇿њঅ⤜Ⲵ䇝㓳ṧֻdᇊѹн਼Ⲵ䈟ᐞ࠭ᮠEd( w )˖
& 1
E d ( w) (t d o d ) 2 ˄4.11˅
2
ަѝtd઼odᱟ䇝㓳ṧֻdⲴⴞḷ䗃ࠪ٬઼অ‫ݳ‬䗃ࠪ٬Ǆ䲿ᵪởᓖл䱽䘝ԓ䇑㇇䇝㓳ṧֻ䳶
&
DⲴ⇿њṧֻdˈ൘⇿⅑䘝ԓ䗷〻ѝ᤹➗‫ޣ‬ҾEd( w )Ⲵởᓖᶕ᭩ਈᵳ٬Ǆ൘䘝ԓᡰᴹ䇝㓳ṧֻ
&
ᰦˈ䘉Ӌᵳ٬ᴤᯠⲴᒿࡇ㔉ࠪҶሩҾ৏ᶕⲴ䈟ᐞ࠭ᮠE( w )Ⲵởᓖл䱽Ⲵањਸ⨶䘁լǄ䙊
䗷֯K˄ởᓖл䱽Ⲵ↕䮯˅Ⲵ٬䏣ཏሿˈਟԕ֯䲿ᵪởᓖл䱽ԕԫ᜿〻ᓖ᧕䘁Ҿⵏᇎởᓖл
䱽Ǆḷ߶Ⲵởᓖл䱽઼䲿ᵪⲴởᓖл䱽ѻ䰤Ⲵ‫ޣ‬䭞४࡛ᱟ˖
x ൘ḷ߶Ⲵởᓖл䱽ѝˈᱟ൘ᵳ٬ᴤᯠࡽሩᡰᴹṧֻ≷ᙫ䈟ᐞˈ❦㘼൘䲿ᵪⲴ
ởᓖл䱽ѝˈᵳ٬ᱟ䙊䗷㘳ḕ⇿њ䇝㓳ᇎֻᶕᴤᯠⲴǄ
x ൘ḷ߶Ⲵởᓖл䱽ѝᵳ٬ᴤᯠⲴ⇿а↕ሩཊњṧֻ≲઼ˈ䘉䴰㾱ᴤཊⲴ䇑
㇇Ǆਖаᯩ䶒ˈഐѪ֯⭘ⵏ↓Ⲵởᓖˈḷ߶Ⲵởᓖл䱽ሩҾ⇿а⅑ᵳ٬ᴤᯠ
㓿ᑨ֯⭘∄䲿ᵪởᓖл䱽ᴹ䖳བྷⲴ↕䮯Ǆ
&
x ྲ᷌E( w )ᴹཊњተ䜘ᶱሿ٬ˈ䲿ᵪⲴởᓖл䱽ᴹᰦਟ㜭䚯‫ݽ‬䲧‫ޕ‬䘉Ӌተ䜘
& &
ᶱሿ٬ˈഐѪᆳ֯⭘н਼ⲴEd( w )㘼нᱟE˄ w ˅ᶕᕅሬᩌ㍒Ǆ
൘ᇎ䐥ѝˈᰐ䇪ᱟ䲿ᵪⲴ䘈ᱟḷ߶Ⲵởᓖл䱽ᯩ⌅䜭㻛ᒯ⌋ᓄ⭘Ǆ
‫ ޜ‬ᔿ ˄ 4.10 ˅ ѝ Ⲵ 䇝㓳 ⌅ ࡉ 㻛〠 Ѫ ໎ 䟿 ⌅ ࡉ ˄ delta rule ˅ˈ ᡆ ᴹ ᰦ ਛ LMS ⌅ ࡉ

˄least-mean-square ᴰሿ൷ᯩ˅ǃAdaline ⌅ࡉǃᡆ Windrow-Hoff ⌅ࡉ˄ԕᆳⲴਁ᰾㘵ભ਽˅Ǆ
൘ㅜ 1 ㄐѝ᧿䘠Ҷᆳ൘ᆖҐঊᔸ䰞仈Ⲵ䇴ՠ࠭ᮠѝⲴᓄ⭘ˈᖃᰦᡁԜ〠ᆳѪ LMS ᵳ٬ᴤᯠ
⌅ࡉǄ⌘᜿‫ޜ‬ᔿ˄4.10˅Ⲵ໎䟿⌅ࡉо 4.4.2 㢲Ⲵᝏ⸕ಘ䇝㓳⌅ࡉ⴨լǄһᇎкєњ㺘䗮ᔿ
& & &
ⴻ䎧ᶕᆼ‫ޘ‬а㠤Ǆ❦㘼ᆳԜᱟн਼Ⲵˈ ഐѪ൘໎䟿⌅ࡉѝ o ᱟᤷ㓯ᙗঅ‫Ⲵݳ‬䗃ࠪ o( x )= w x ˈ
& & &
㘼ሩҾᝏ⸕ಘ⌅ࡉˈo ᱟᤷ䰸٬䗃ࠪ o( x )=sgn( w x )Ǆ
ቭ㇑ᡁԜ㔉ࠪⲴ໎䟿⌅ࡉਟᆖҐ䶎䰸٬㓯ᙗঅ‫Ⲵݳ‬ᵳˈնᆳҏਟԕᯩ‫ׯ‬ൠ⭘ᶕ䇝㓳ᴹ䰸
& & & &
٬Ⲵᝏ⸕ಘঅ‫ݳ‬Ǆ‫ٷ‬ᇊ o= w x ᱟк䶒Ⲵ䶎䰸٬㓯ᙗঅ‫Ⲵݳ‬䗃ࠪˈᒦф oc=sgn˄ w x ˅ᱟ o
㻛䰸٬ॆⲴ㔃᷌ˈо൘ᝏ⸕ಘѝаṧǄ⧠൘ྲ᷌ᡁԜᐼᵋѪ oc䇝㓳ањᝏ⸕ಘ֯ަᤏਸⴞḷ
٬Ѫr1 Ⲵ䇝㓳ṧֻˈਟԕ֯⭘о䇝㓳 o аṧⲴⴞḷ٬઼䇝㓳ṧֻˈн䗷֯⭘໎䟿⌅ࡉǄᖸ
᰾ᱮˈྲ᷌䶎䰸٬䗃ࠪ o 㜭ཏ㻛䇝㓳ࡠᆼ㖾ᤏਸ䘉Ӌ٬ˈ䛓Ѹ䰸٬䗃ࠪ oc ҏՊᤏਸᆳԜ˄ഐ
Ѫ sgn(1)=1ˈ઼ sgn(-1)=-1˅Ǆণ֯н㜭ᆼ㖾ൠᤏਸⴞḷ٬ˈਚ㾱㓯ᙗঅ‫Ⲵݳ‬䗃ࠪާᴹ↓⺞Ⲵ
ㅖਧˈᴹ䰸٬Ⲵ oc٬Պ↓⺞ൠᤏਸⴞḷ٬r1Ǆ❦㘼⌘᜿ˈ⭡Ҿ䘉њ䗷〻Պᗇࡠ֯㓯ᙗঅ‫ݳ‬䗃
ࠪⲴ䈟ᐞᴰሿॆⲴᵳ٬ˈ䘉Ӌᵳ٬н㜭‫؍‬䇱ҏ֯ᴹ䰸٬䗃ࠪ ocⲴ䈟࠶㊫ṧֻᮠᴰሿॆǄ
4.4.4 ቅ㔉
ᡁԜᐢ㓿⹄ウҶ䘝ԓᆖҐᝏ⸕ಘᵳ٬Ⲵєњ⴨լⲴ㇇⌅Ǆ䘉єњ㇇⌅䰤Ⲵ‫ޣ‬䭞ᐞᔲᱟᝏ
⸕ಘ䇝㓳⌅ࡉṩᦞ䰸٬ॆ˄thresholded˅Ⲵᝏ⸕ಘ䗃ࠪⲴ䈟ᐞᴤᯠᵳ٬˄4.4.2˅ˈ❦㘼໎䟿
⌅ࡉṩᦞ䗃‫Ⲵޕ‬䶎䰸٬ॆ˄unthresholded˅㓯ᙗ㓴ਸⲴ䈟ᐞᶕᴤᯠᵳ(4.4.3)Ǆ
䘉єњ䇝㓳⌅ࡉ䰤Ⲵᐞᔲ৽᱐൘н਼Ⲵ᭦ᮋ⢩ᙗкǄᝏ⸕ಘ䇝㓳⌅ࡉ㓿䗷ᴹ䲀⅑Ⲵ䘝ԓ
᭦ᮋࡠањ㜭⨶ᜣ࠶㊫䇝㓳ᮠᦞⲴ‫ٷ‬䇮ˈնᶑԦᱟ䇝㓳ṧֻ㓯ᙗਟ࠶Ǆ໎䟿⌅ࡉ⑀䘁᭦ᮋࡠ
ᴰሿ䈟ᐞ‫ٷ‬䇮ˈਟ㜭䴰㾱ᰐ䲀Ⲵᰦ䰤ˈնᰐ䇪䇝㓳ṧֻᱟ੖㓯ᙗਟ࠶䜭Պ᭦ᮋǄ‫ޣ‬Ҿԕк᭦
ᮋᙗⲴ䈖㓶䇱᰾ਟԕ৲㘳 Hertz et al.˄1991˅Ǆ
ᆖҐᵳੁ䟿Ⲵㅜй⿽ਟ㜭ᯩ⌅ᱟ㓯ᙗ㿴ࡂ˄linear programming˅Ǆ㓯ᙗ㿴ࡂᱟ䀓㓯ᙗн
& & & &
ㅹᔿᯩ〻㓴Ⲵа⿽䙊⭘Ⲵᴹ᭸ᯩ⌅Ǆ⌘᜿⇿њ䇝㓳ṧֻሩᓄањᖒᔿѪ w x >0 ᡆ w x d0
ⲴнㅹᔿˈᒦфᆳԜⲴ䀓ቡᱟᡁԜᵏᵋⲴᵳੁ䟿ǄнᒨⲴᱟˈ䘉⿽ᯩ⌅ӵᖃ䇝㓳ṧֻ㓯ᙗਟ
࠶ᰦᴹ䀓ˈն Duda & Hart ˄1973ˈp.168˅ᔪ䇞Ҷа⿽ᴤᐗ࿉Ⲵᯩ⌅䘲ਸ䶎㓯ᙗਟ࠶Ⲵᛵ
ߥǄᰐ䇪ྲօˈ䘉⿽㓯ᙗ㿴ࡂⲴᯩ⌅н㜭ᢙኅࡠ䇝㓳ཊቲ㖁㔌ˈ䘉ᱟᡁԜᴰ‫ޣ‬ᗳⲴǄ⴨৽ˈ
↓ྲла㢲ᡰ䇘䇪ⲴˈสҾ໎䟿⌅ࡉⲴởᓖл䱽ᯩ⌅ਟԕㆰঅൠᢙኅࡠཊቲ㖁㔌Ǆ
4.5 ཐቸ㖇㔒ૂਃੇՖ᫣㇍⌋
↓ྲ 4.4.1 㢲ᡰᤷࠪⲴˈঅњᝏ⸕ಘӵ㜭㺘⽪㓯ᙗߣㆆ䶒Ǆ⴨৽ˈ৽ੁՐ᫝㇇⌅ᡰᆖҐ
Ⲵཊቲ㖁㔌㜭ཏ㺘⽪⿽㊫㑱ཊⲴ䶎㓯ᙗᴢ䶒Ǆֻྲˈമ 4-5 ᧿䘠ҶањިරⲴཊቲ㖁㔌઼ᆳ
Ⲵߣㆆᴢ䶒Ǆ䘉њ䈝丣䇶࡛ԫ࣑㾱४࠶ࠪ⧠൘Āh_dāкл᮷ѝⲴ 10 ⿽‫ݳ‬丣˄ֻྲˈĀhidāˈ
ĀhadāˈĀheadāˈĀhoodāㅹ˅Ǆ䗃‫Ⲵޕ‬䈝丣ؑਧ⭘єњ৲ᮠ㺘⽪ˈᆳԜᱟ䙊䗷ሩ༠丣Ⲵ仁䉡
࠶᷀ᗇࡠⲴˈ䘉ṧᡁԜਟԕᯩ‫ׯ‬ൠ൘Ҽ㔤ᇎֻオ䰤ѝᱮ⽪ࠪߣㆆ䶒Ǆྲമਟ㿱ˈཊቲ㖁㔌㜭
ཏ㺘⽪儈ᓖ䶎㓯ᙗⲴߣㆆ䶒ˈᆳ∄ࡽ䶒മ 4-3 ѝ⭫ࠪⲴঅњঅ‫Ⲵݳ‬㓯ᙗߣㆆ䶒㺘ᖱ㜭࣋ᴤ
ᕪǄ
മ 4-5 ཊቲࡽ侸㖁㔌Ⲵߣㆆ४ฏ
䘉䟼ᱮ⽪Ⲵ㖁㔌ᱟ⭘ᶕ䇝㓳䇶࡛ 10 ⿽ࠪ⧠൘Āh_dā˄ֻྲĀhadāˈĀhidā˅䰤Ⲵ‫ݳ‬丣Ǆ䘉њ㖁
㔌Ⲵ䗃‫⭡ޕ‬єњ৲ᮠ F1 ઼ F2 㓴ᡀˈᆳԜᱟ䙊䗷ሩ༠丣Ⲵ仁䉡࠶᷀ᗇࡠⲴǄ㖁㔌Ⲵ 10 њ䗃ࠪሩᓄ
Ҿ 10 њਟ㜭Ⲵ‫ݳ‬丣Ǆ䘉њ㖁㔌Ⲵ亴⍻ᱟަѝᴹᴰབྷ٬Ⲵ䗃ࠪǄਣമ⭫ࠪҶᆖࡠⲴ㖁㔌ᡰԓ㺘Ⲵ儈
ᓖ䶎㓯ᙗߣㆆ䶒ǄമѝⲴ⛩㺘⽪⍻䈅ṧֻˈᆳԜо⭘ᶕ䇝㓳䘉њ㖁㔌Ⲵṧֻᱟᆼ‫ޘ‬н਼ⲴǄ˄㓿䇨
ਟ᪈㠚 Haung & Lippmann˄1988˅˅
ᵜ㢲䇘䇪ྲօᆖҐ䘉ṧⲴཊቲ㖁㔌ˈ֯⭘Ⲵ㇇⌅઼ࡽ䶒䇘䇪Ⲵởᓖл䱽ᯩ⌅⴨լǄ
4.5.1 ਥᗤ䰾‫ٲ‬ঋ‫ݹ‬
ᓄ䈕֯⭘ӰѸ㊫රⲴঅ‫ݳ‬ᶕ֌Ѫᶴᔪཊቲ㖁㔌Ⲵส⹰˛䎧ࡍᡁԜਟԕቍ䈅䘹ᤙࡽ䶒䇘
䇪Ⲵ㓯ᙗঅ‫ˈݳ‬ഐѪᡁԜᐢ㓿Ѫ䘉⿽অ‫ݳ‬ሬࠪҶањởᓖл䱽ᆖҐ⌅ࡉǄ❦㘼ˈཊњ㓯ᙗঅ
‫Ⲵݳ‬䘎᧕ӽᰗӗ⭏㓯ᙗ࠭ᮠˈ㘼ᡁԜᴤᐼᵋ䘹ᤙ㜭ཏ㺘ᖱ䶎㓯ᙗ࠭ᮠⲴ㖁㔌Ǆᝏ⸕ಘঅ‫ݳ‬ᱟ
ਖа⿽䘹ᤙˈնᆳⲴн䘎㔝䰸٬֯ᆳнਟᗞˈᡰԕн䘲ਸởᓖл䱽㇇⌅ǄᡁԜᡰ䴰㾱Ⲵᱟ䘉
ṧⲴঅ‫ˈݳ‬ᆳⲴ䗃ࠪᱟ䗃‫Ⲵޕ‬䶎㓯ᙗ࠭ᮠˈᒦф䗃ࠪᱟ䗃‫Ⲵޕ‬ਟᗞ࠭ᮠǄа⿽ㆄṸᱟ sigmoid
অ‫˄ݳ‬sigmoid unit˅ˈ䘉ᱟа⿽䶎ᑨ㊫լҾᝏ⸕ಘⲴঅ‫ˈݳ‬նᆳสҾањᒣ━Ⲵਟᗞ䰸٬࠭
ᮠǄ
മ 4-6 sigmoid 䰸٬অ‫ݳ‬
മ 4-6 ⭫ࠪҶ sigmoid অ‫ݳ‬Ǆоᝏ⸕ಘ⴨լˈsigmoid অ‫ݸݳ‬䇑㇇ᆳⲴ䗃‫Ⲵޕ‬㓯ᙗ㓴ਸˈ

❦ਾᓄ⭘ањ䰸٬ࡠ↔㔃᷌Ǆ❦㘼ˈሩҾ sigmoid অ‫ˈݳ‬䰸٬䗃ࠪᱟ䗃‫Ⲵޕ‬䘎㔝࠭ᮠǄᴤ㋮
⺞ൠ䇢ˈsigmoid অ‫ݳ‬䘉ṧ䇑㇇ᆳⲴ䗃ࠪ˖
& &
o=V ( w x )
ަѝ
1
V ( y) ˄4.12˅
1 e y
V 㓿ᑨ㻛〠Ѫsigmoid࠭ᮠᡆ㘵ҏਟԕ〠Ѫlogistic࠭ᮠ˄logistic function˅Ǆ⌘᜿ᆳⲴ䗃
ࠪ㤳തѪ 0 ࡠ 1ˈ䲿䗃‫ޕ‬অ䈳䙂໎˄৲㿱മ 4-6 ѝⲴ䰸٬࠭ᮠᴢ㓯˅ǄഐѪ䘉њ࠭ᮠᢺ䶎ᑨ
བྷⲴ䗃‫ޕ‬٬ฏ᱐ሴࡠањሿ㤳തⲴ䗃ࠪˈᆳ㓿ᑨ㻛〠Ѫsigmoidঅ‫Ⲵݳ‬ᥔ঻࠭ᮠ˄squashing
function˅Ǆsigmoid࠭ᮠᴹањᴹ⭘Ⲵ⢩ᖱˈᆳⲴሬᮠᖸᇩ᱃ԕᆳⲴ䗃ࠪ㺘⽪[⺞࠷ൠ䇢ˈ
dV ( y )
=V(y)(1-V(y))]ǄᡁԜሶⴻࡠˈਾ䶒Ⲵởᓖл䱽ᆖҐ⌅ࡉ֯⭘Ҷ䘉њሬᮠǄᴹᰦҏਟ
dy
ԕ֯⭘ަԆ᱃䇑㇇ሬᮠⲴਟᗞ࠭ᮠԓᴯVǄֻྲˈsigmoid࠭ᮠᇊѹⲴe-y亩ᴹᰦ㻛ᴯᦒѪe-kyˈ
ަѝkѪḀњ↓ᑨᮠˈ⭘ᶕߣᇊ䘉њ䰸٬࠭ᮠⲴ䲑ጝᙗǄৼᴢ↓࠷࠭ᮠtanhᴹᰦҏ⭘ᶕԓᴯ
sigmoid࠭ᮠ˄৲㿱㓳Ґ 4.8˅Ǆ
4.5.2 ਃੇՖ᫣㇍⌋
ሩҾ⭡а㌫ࡇ⺞ᇊⲴঅ‫ݳ‬ӂ䘎ᖒᡀⲴཊቲ㖁㔌ˈ৽ੁՐ᫝㇇⌅ਟ⭘ᶕᆖҐ䘉њ㖁㔌Ⲵᵳ
٬Ǆᆳ䟷⭘ởᓖл䱽ᯩ⌅䈅മᴰሿॆ㖁㔌䗃ࠪ٬઼ⴞḷ٬ѻ䰤Ⲵ䈟ᐞᒣᯩǄ䘉а㢲㔉ࠪ৽ੁ
Ր᫝㇇⌅ˈла㢲᧘ሬࠪ৽ੁՐ᫝㇇⌅֯⭘Ⲵởᓖл䱽ᵳ٬ᴤᯠ⌅ࡉǄ
ഐѪᡁԜ㾱㘳㲁ཊњ䗃ࠪঅ‫Ⲵݳ‬㖁㔌ˈ㘼нᱟ䊑ԕࡽਚ㘳㲁অњঅ‫ˈݳ‬ᡰԕᡁԜ‫ݸ‬䟽ᯠ
ᇊѹ䈟ᐞ Eˈԕ‫ׯ‬ሩᡰᴹ㖁㔌䗃ࠪⲴ䈟ᐞ≲઼Ǆ
& 1
E( w ) { ¦ ¦ (t kd okd ) 2
2 d D koutputs
˄4.13˅
ަѝoutputsᱟ㖁㔌䗃ࠪঅ‫Ⲵݳ‬䳶ਸˈtkd઼okdᱟо䇝㓳ṧֻd઼ㅜkњ䗃ࠪঅ‫Ⲵޣ⴨ݳ‬䗃ࠪ
٬Ǆ
৽ੁՐ᫝㇇⌅䶒ѤⲴᆖҐ䰞仈ᱟᩌ㍒ањᐘབྷⲴ‫ٷ‬䇮オ䰤ˈ䘉њオ䰤⭡㖁㔌ѝᡰᴹঅ‫ݳ‬
Ⲵᡰᴹਟ㜭Ⲵᵳ٬ᇊѹǄ䘉⿽ᛵߥਟԕ⭘ањ䈟ᐞᴢ䶒ᶕᖒ䊑㺘⽪ˈоമ 4-4 㺘⽪Ⲵ㓯ᙗঅ
‫Ⲵݳ‬䈟ᐞᴢ䶒⴨լǄ䛓ᑵമѝⲴ䈟ᐞ㻛ᡁԜⲴᯠⲴ䈟ᐞᇊѹ E ᡰᴯԓˈᒦфオ䰤ѝⲴަԆ
㔤⧠൘ሩᓄ㖁㔌ѝоᡰᴹঅ‫Ⲵޣ⴨ݳ‬ᡰᴹᵳ٬Ǆ઼䇝㓳অњঅ‫Ⲵݳ‬ᛵߥаṧˈởᓖл䱽ਟ㻛
⭘ᶕቍ䈅ራ᢮ањ‫ٷ‬䇮֯ E ᴰሿॆǄ
ཊቲ㖁㔌Ⲵањѫ㾱н਼ᱟᆳⲴ䈟ᐞᴢ䶒ਟ㜭ᴹཊњተ䜘ᶱሿ٬ˈ㘼മ 4-4 㺘⽪Ⲵᣋ⢙

ᴢ䶒ӵᴹањᴰሿ٬ǄнᒨⲴᱟˈ䘉᜿ણ⵰ởᓖл䱽ӵ㜭‫؍‬䇱᭦ᮋࡠተ䜘ᶱሿ٬ˈ㘼ᵚᗵᗇ
ࡠ‫ޘ‬ተᴰሿⲴ䈟ᐞǄቭ㇑ᴹ䘉њ䳌⺽ˈᐢ㓿ਁ⧠ሩҾᇎ䐥ѝᖸཊᓄ⭘৽ੁՐ᫝㇇⌅䜭ӗ⭏Ҷ
ࠪ㢢Ⲵ㔃᷌Ǆ
㺘 4-2 वਜ਼єቲ sigmoid অ‫ࡽⲴݳ‬侸㖁㔌Ⲵ৽ੁՐ᫝㇇⌅˄䲿ᵪởᓖл䱽⡸ᵜ˅
Backpropagation(training_examples, K, nin, nout, nhidden)

& & &
trainning_exaples ѝ⇿ањ䇝㓳ṧֻᱟᖒᔿѪ< x , t >Ⲵᒿ‫ަˈڦ‬ѝ x ᱟ㖁㔌䗃
&
‫ޕ‬٬ੁ䟿ˈ t ᱟⴞḷ䗃ࠪ٬Ǆ
KᱟᆖҐ䙏⦷˄ֻྲ 0.05˅Ǆninᱟ㖁㔌䗃‫Ⲵޕ‬ᮠ䟿ˈnhiddenᱟ䳀㯿ቲঅ‫ݳ‬ᮠˈnoutᱟ
䗃ࠪঅ‫ݳ‬ᮠǄ
Ӿঅ‫ݳ‬iࡠঅ‫ݳ‬jⲴ䗃‫ޕ‬㺘⽪Ѫxjiˈঅ‫ݳ‬iࡠঅ‫ݳ‬jⲴᵳ٬㺘⽪ѪwijǄ
z ࡋᔪާᴹninњ䗃‫ˈޕ‬nhiddenњ䳀㯿অ‫ˈݳ‬noutњ䗃ࠪঅ‫Ⲵݳ‬㖁㔌
z ࡍ࿻ॆᡰᴹⲴ㖁㔌ᵳ٬ѪሿⲴ䲿ᵪ٬˄ֻྲ-0.05 ઼ 0.05 ѻ䰤Ⲵᮠ˅
z ൘䙷ࡠ㓸→ᶑԦࡽˈ‫ڊ‬
& &
z ሩҾ䇝㓳ṧֻ training_examples ѝⲴ⇿њ< x , t >ˈ‫ڊ‬
ᢺ䗃‫⋯ޕ‬㖁㔌ࡽੁՐ᫝
&
1ˊ ᢺᇎֻ x 䗃‫ޕ‬㖁㔌ˈᒦ䇑㇇㖁㔌ѝ⇿њঅ‫ݳ‬uⲴ䗃ࠪ
ouǄ
֯䈟ᐞ⋯㖁㔌৽ੁՐ᫝
2ˊ ሩҾ㖁㔌Ⲵ⇿њ䗃ࠪঅ‫ݳ‬kˈ䇑㇇ᆳⲴ䈟ᐞ亩Gk
Gk mok(1-ok)(tk-ok) ˄T4.3˅
3ˊ ሩҾ㖁㔌Ⲵ⇿њ䳀㯿অ‫ݳ‬hˈ䇑㇇ᆳⲴ䈟ᐞ亩Gh
Gh moh(1-oh) ¦
wkhGk ˄T4.4˅
koutputs
4ˊ ᴤᯠ⇿њ㖁㔌ᵳ٬wji
wjim wji +'wji
ަѝ
'wji=KGjxji ˄T4.5˅
㺘 4-2 㔉ࠪҶ৽ੁՐ᫝㇇⌅Ǆ䘉䟼᧿䘠Ⲵ㇇⌅䘲⭘Ҿवਜ਼єቲ sigmoid অ‫࠶Ⲵݳ‬ቲࡽ侸

㖁㔌ˈᒦф⇿аቲⲴঅ‫ݳ‬оࡽаቲⲴᡰᴹঅ‫⴨ݳ‬䘎Ǆ䘉ᱟ৽ੁՐ᫝㇇⌅Ⲵ໎䟿ởᓖл䱽˄ᡆ
䲿ᵪởᓖл䱽˅⡸ᵜǄ䘉䟼֯⭘Ⲵㅖਧоࡽа㢲֯⭘Ⲵаṧˈᒦ䘋㹼ҶྲлⲴᢙኅ˖
x 㖁㔌ѝ⇿њ㔃⛩㻛䍻Ҹањᒿਧ˄ֻྲањᮤᮠ˅ˈ䘉䟼Ⲵ㔃⛩㾱Ѹᱟ㖁㔌
Ⲵ䗃‫ˈޕ‬㾱Ѹᱟ㖁㔌ѝḀњঅ‫Ⲵݳ‬䗃ࠪǄ
x xji㺘⽪㔃⛩iࡠঅ‫ݳ‬jⲴ䗃‫ˈޕ‬ᒦфwji㺘⽪ሩᓄⲴᵳ٬Ǆ
x Gn㺘⽪оঅ‫ݳ‬n⴨‫ޣ‬㚄Ⲵ䈟ᐞ亩ǄᆳⲴ䀂㢢оࡽ䶒䇘䇪Ⲵdelta䇝㓳⌅ࡉѝⲴ
wE
(t-o)⴨լǄਾ䶒ᡁԜਟԕⴻࡠGn = Ǆ
wnet n
൘㺘 4-2 Ⲵ㇇⌅Ⲵᔰ࿻ˈᔪ・ањާᴹᵏᵋᮠ䟿Ⲵ䳀অ‫઼ݳ‬䗃ࠪঅ‫Ⲵݳ‬㖁㔌ˈᒦࡍ࿻ॆ
ᡰᴹ㖁㔌Ⲵᵳ٬ѪሿⲴ䲿ᵪᮠǄ㔉ᇊҶ䘉њപᇊⲴ㖁㔌㔃ᶴˈ㇇⌅Ⲵѫᗚ⧟ቡሩ䇝㓳ṧֻ䘋
㹼৽༽Ⲵ䘝ԓǄሩҾ⇿ањ䇝㓳ṧֻˈᆳᓄ⭘ⴞࡽⲴ㖁㔌ࡠ䘉њṧֻˈ䇑㇇ሩҾ䘉њṧֻ㖁
㔌䗃ࠪⲴ䈟ᐞˈ❦ਾᴤᯠ㖁㔌ѝᡰᴹⲴᵳ٬Ǆሩ䘉ṧⲴởᓖл䱽↕僔䘋㹼䘝ԓˈⴤࡠ㖁㔌Ⲵ
ᙗ㜭䗮ࡠਟ᧕ਇⲴ㋮ᓖ˄㓿ᑨᱟкॳ⅑ˈཊ⅑֯⭘਼ṧⲴ䇝㓳ṧֻ˅Ǆ
䘉䟼Ⲵởᓖл䱽ᵳᴤᯠ⌅ࡉ˄㺘 4-2 ѝⲴ‫ޜ‬ᔿ[T4.5]˅оdelta䇝㓳⌅ࡉ˄‫ޜ‬ᔿ[4.10]˅⴨

լǄቡ䊑delta⌅ࡉˈᆳ‫➗׍‬ԕлй㘵Ⲵ҈〟ᶕᴤᯠ⇿ањᵳ˖ᆖҐ䙏⦷Kǃ䈕ᵳ٬ᓄ⭘Ⲵ䗃
‫ޕ‬٬xjiǃ઼䘉њঅ‫ݳ‬䗃ࠪⲴ䈟ᐞǄᜏаⲴн਼ᱟdelta⌅ࡉѝⲴ䈟ᐞ亩˄t-o˅㻛ᴯᦒᡀањᴤ
༽ᵲⲴ䈟ᐞ亩GjǄ൘ 4.5.3 㢲Ⲵሩᵳᴤᯠ⌅ࡉⲴ᧘ሬѻਾᡁԜሶ㔉ࠪGjⲴ߶⺞ᖒᔿǄѪҶⴤ㿲
ൠ⨶䀓ᆳˈ‫ݸ‬㘳㲁㖁㔌Ⲵ⇿ањ䗃ࠪঅ‫ݳ‬kⲴGk˄൘㇇⌅Ⲵ‫ޜ‬ᔿ[T4.3]ѝ˅˅ᱟᘾṧ䇑㇇ⲴǄ
ᖸㆰঅˈGkоdelta⌅ࡉѝⲴ˄tk-ok˅⴨լˈն҈кҶsigmoidᥔ঻࠭ᮠⲴሬᮠok(1-ok)Ǆ⇿њ䳀
㯿অ‫ݳ‬hⲴGhⲴ٬ާᴹ⴨լⲴᖒᔿ˄㇇⌅Ⲵ‫ޜ‬ᔿ[T4.4]˅Ǆ❦㘼ˈഐѪ䇝㓳ṧֻӵሩ㖁㔌Ⲵ䗃
ࠪᨀ‫׋‬Ҷⴞḷ٬tkˈᡰԕ㕪ቁⴤ᧕Ⲵⴞḷ٬ᶕ䇑㇇䳀㯿অ‫Ⲵݳ‬䈟ᐞ٬Ǆഐ↔䟷ਆԕл䰤᧕࣎
⌅䇑㇇䳀㯿অ‫Ⲵݳ‬䈟ᐞ亩˖ሩਇ䳀㯿অ‫ݳ‬hᖡ૽Ⲵ⇿ањঅ‫Ⲵݳ‬䈟ᐞGk䘋㹼࣐ᵳ≲઼ˈ⇿њ
䈟ᐞGkᵳ٬ѪwkhˈwkhቡᱟӾ䳀㯿অ‫ݳ‬hࡠ䗃ࠪঅ‫ݳ‬kⲴᵳ٬Ǆ䘉њᵳ٬࡫⭫Ҷ䳀㯿অ‫ݳ‬hሩҾ
䗃ࠪঅ‫ݳ‬kⲴ䈟ᐞᓄĀ䍏䍓āⲴ〻ᓖǄ
㺘 4-2 ѝⲴ㇇⌅䲿⵰⇿њ䇝㓳ṧֻⲴࠪ⧠䙂໎ൠᴤᯠᵳǄ䘉а⛩оởᓖл䱽Ⲵ䲿ᵪ䘁լ
㇇⌅а㠤Ǆ㾱ਆᗇ䈟ᐞEⲴⵏᇎởᓖˈ䴰㾱൘‫؞‬᭩ᵳ٬ѻࡽሩᡰᴹ䇝㓳ṧֻⲴGjxji٬≲઼Ǆ
൘ިරⲴᓄ⭘ѝˈ৽ੁՐ᫝㇇⌅Ⲵᵳ٬ᴤᯠ䘝ԓՊ㻛䟽༽кॳ⅑Ǆᴹᖸཊ㓸→ᶑԦਟԕ
⭘ᶕ‫→ڌ‬䘉њ䗷〻Ǆа⿽ᯩ⌅ᱟ൘䘝ԓⲴ⅑ᮠࡠҶањപᇊ٬ᰦ‫˗→ڌ‬ᡆᖃ൘䇝㓳ṧֻкⲴ
䈟ᐞ䱽ࡠḀњ䰸٬ԕлᰦ˗ᡆ൘࠶⿫Ⲵ傼䇱ṧֻ䳶ਸкⲴ䈟ᐞㅖਸḀњḷ߶ᰦǄ㓸→ࡔᦞⲴ
䘹ᤙᱟᖸ䟽㾱ⲴˈഐѪཚቁⲴᗚ⧟ਟ㜭⋑ᴹᴹ᭸ൠ䱽վ䈟ᐞˈ㘼ཚཊⲴᗚ⧟Պሬ㠤ሩ䇝㓳ᮠ
ᦞⲴ䗷ᓖᤏਸǄ൘ 4.6.5 㢲ѝᡁԜՊᴤ䈖㓶ൠ䇘䇪䘉њ䰞仈Ǆ
4.5.2.1 ໔ࣖߨ䠅δMomentumε亯
ഐѪ৽ੁՐ᫝㇇⌅Ⲵᓄ⭘ྲ↔ᒯ⌋ˈᡰԕᐢ㓿ᔰਁࠪҶᖸཊ৽ੁՐ᫝㇇⌅ⲴਈփǄަѝ
ᴰᑨ㿱Ⲵᱟ‫؞‬᭩㇇⌅ѝ‫ޜ‬ᔿ˄T4.5˅Ⲵᵳ٬ᴤᯠ⌅ࡉˈ֯ㅜ n ⅑䘝ԓⲴᵳ٬ᴤᯠ䜘࠶ൠ‫׍‬䎆
Ҿਁ⭏൘ㅜ n-1 ⅑䘝ԓᰦⲴᴤᯠˈণᢺ‫ޜ‬ᔿ˄T4.5˅ᦒѪྲлⲴᖒᔿ˖
'wjin KGjxjiD'wjin̢ ˄˅
䘉䟼'wji(n)ᱟ㇇⌅ѫᗚ⧟ѝⲴㅜn⅑䘝ԓ䘋㹼Ⲵᵳ٬ᴤᯠˈᒦф 0dD<1 ᱟањ〠Ѫ冲量

˄momentum˅ⲴᑨᮠǄ⌘᜿䘉њ‫ޜ‬ᔿਣ‫Ⲵח‬ㅜа亩ቡᱟ৽ੁՐ᫝㇇⌅Ⲵ‫ޜ‬ᔿ˄T4.5˅ѝⲴ
ᵳ٬ᴤᯠǄਣ䗩ⲴㅜҼ亩ᱟᯠⲴˈ㻛〠Ѫߢ䟿亩ǄѪҶ⨶䀓䘉њߢ䟿亩Ⲵ֌⭘ˈ䇮ᜣởᓖл
䱽Ⲵᩌ㍒䖘䘩ቡྭ‫ۿ‬ањ˄ᰐߢ䟿Ⲵ˅⨳┊л䈟ᐞᴢ䶒ǄDⲴ֌⭘ᱟ໎࣐ߢ䟿֯䘉њ⨳Ӿа
⅑䘝ԓࡠла⅑䘝ԓᰦԕ਼ṧⲴᯩੁ┊ࣘǄߢ䟿ᴹᰦՊ֯䘉њ⨳┊䗷䈟ᐞᴢ䶒Ⲵተ䜘ᶱሿ
٬˗ᡆ֯ަ┊䗷䈟ᐞᴢ䶒кⲴᒣඖ४ฏˈྲ᷌⋑ᴹߢ䟿䘉њ⨳ᴹਟ㜭൘䘉њ४ฏ‫→ڌ‬Ǆᆳҏ
ާᴹ൘ởᓖнਈⲴ४ฏ䙀⑀໎བྷᩌ㍒↕䮯Ⲵ᭸᷌ˈӾ㘼ਟԕ࣐ᘛ᭦ᮋǄ
4.5.2.2 ᆜҖԱᝅⲺᰖ⧥㖇㔒
㺘 4-2 㔉ࠪⲴ৽ੁՐ᫝㇇⌅Ⲵᇊѹӵ䘲⭘ҾєቲⲴ㖁㔌Ǆ❦㘼䛓䟼㔉ࠪⲴ㇇⌅ਟԕㆰঅ
ൠ᧘ᒯࡠԫ᜿␡ᓖⲴࡽ侸㖁㔌Ǆ‫ޜ‬ᔿ˄T4.5˅Ⲵᵳ٬ᴤᯠ⌅ࡉ‫؍‬ᤱнਈˈᜏаⲴਈॆᱟ䇑㇇
G٬Ⲵ䗷〻Ǆᾲᤜൠ䈤ˈㅜmቲⲴঅ‫ݳ‬rⲴGr٬ᱟ⭡ᴤ␡Ⲵm+1 ቲⲴG٬ṩᦞлᔿ䇑㇇Ⲵ˖
Gr =or(1- or) ¦
sm 1ቲ
wsrGs ˄4.19˅
⌘᜿䘉њ‫ޜ‬ᔿо㺘 4-2 ㇇⌅Ⲵㅜ 3 ↕⴨਼ˈ䘉䟼㾱䈤᰾ⲴᱟሩҾ㖁㔌ѝⲴԫ᜿ᮠ䟿Ⲵ䳀

㯿অ‫ˈݳ‬䈕↕僔㾱㻛䟽༽ᖸཊ䙽Ǆ
ྲ᷌᧘ᒯࡠԫօᴹੁᰐ⧟㔃ᶴҏаṧⲴㆰঅˈ㘼н䇪㖁㔌ѝⲴঅ‫ݳ‬ᱟ੖䊑ᡁԜ㠣↔Ѫ→
‫ٷ‬ᇊⲴ䛓ṧ㻛㔏аൠᧂࡇ൘ቲкǄሩҾ㖁㔌অ‫⋑ݳ‬ᴹ᤹↔ᧂࡇⲴᛵߥˈ䇑㇇ԫ᜿޵䜘অ‫˄ݳ‬ҏ
ቡᱟᡰᴹ䶎䗃ࠪঅ‫Ⲵ˅ݳ‬GⲴ⌅ࡉᱟ˖
Gr =or(1- or) ¦
sDownStream (r )
wsrGs ˄4.20˅
ަѝ DownStream(r)ᱟ൘㖁㔌ѝঅ‫ ݳ‬r Ⲵ・ণл⑨˄immediately downstream˅অ‫Ⲵݳ‬䳶

ਸˈᡆ㘵䈤䗃‫ޕ‬ѝवᤜ r Ⲵ䗃ࠪⲴᡰᴹঅ‫ݳ‬Ǆ4.5.3 㢲ᡁԜ㾱᧘ሬⲴቡᱟ䘉⿽ᵳ٬ᴤᯠ⌅ࡉ
Ⲵа㡜ᖒᔿǄ
4.5.3 ਃੇՖ᫣⌋ࡏⲺ᧞ሲ
䘉а㢲㔉ࠪ৽ੁՐ᫝㇇⌅Ⲵᵳ٬䈳ᮤ⌅ࡉⲴ᧘ሬˈྲ᷌ᱟㅜа䙽䰵䈫ਟԕ䐣䗷䘉а㢲ˈ
㘼нཡ䘎㔝ᙗǄ
䘉䟼ᡁԜ㾱䀓ߣⲴ䰞仈ᱟ᧘ሬࠪ㺘 4-2 ㇇⌅֯⭘Ⲵ䲿ᵪởᓖл䱽⌅ࡉǄഎᗶ‫ޜ‬ᔿ˄4.11˅ˈ

䲿ᵪⲴởᓖл䱽㇇⌅䘝ԓ༴⨶䇝㓳ṧֻˈ⇿⅑༴⨶ањǄሩҾ⇿њ䇝㓳ṧֻdˈ࡙⭘‫ޣ‬Ҿ䘉
њṧֻⲴ䈟ᐞEdⲴởᓖ‫؞‬᭩ᵳ٬Ǆᦒਕ䈍䈤ˈሩҾ⇿ањ䇝㓳ṧֻdˈ⇿њᵳwji㻛໎࣐'wjiǄ
wE d
'wji= K ˄4.21˅
ww ji
ަѝˈEdᱟ䇝㓳ṧֻdⲴ䈟ᐞˈ䙊䗷ሩ㖁㔌ѝᡰᴹ䗃ࠪঅ‫઼≲Ⲵݳ‬ᗇࡠ
& 1
Ed( w ) { ¦ (t k ok ) 2
2 koutputs
䘉䟼outputsᱟ㖁㔌ѝ䗃ࠪঅ‫Ⲵݳ‬䳶ਸˈtkᱟঅ‫ݳ‬kሩҾ䇝㓳ṧֻdⲴⴞḷ٬ˈokᱟ㔉ᇊ䇝
㓳ṧֻdᰦঅ‫ݳ‬kⲴ䗃ࠪ٬Ǆ
䲿ᵪởᓖл䱽⌅ࡉⲴ᧘ሬᾲᘥкᱟ᱃៲Ⲵˈն䴰㾱⮉᜿ᖸཊлḷ઼ਈ䟿ǄᡁԜሶ䚥ᗚമ
4-6 ѝᡰ⭫ࠪⲴㅖਧˈ໎࣐ањлḷ j ⭘ᶕ㺘⽪㖁㔌ѝⲴㅜ j њঅ‫ާˈݳ‬փྲл˖
x xji=অ‫ݳ‬jⲴㅜiњ䗃‫ޕ‬
x wji=оঅ‫ݳ‬jⲴㅜiњ䗃‫ޣ⴨ޕ‬㚄Ⲵᵳ٬
x netj=¦iwjixji˄অ‫ݳ‬jⲴ䗃‫࣐Ⲵޕ‬ᵳ઼˅
x oj=অ‫ݳ‬j䇑㇇ࠪⲴ䗃ࠪ
x tj=অ‫ݳ‬jⲴⴞḷ䗃ࠪ
x V=sigmoid ࠭ᮠ
x outputs=㖁㔌ⲴᴰਾаቲⲴঅ‫ݳ‬䳶ਸ
x DownStream(j)=অ‫・Ⲵݳ‬ণ䗃‫˄ޕ‬immediate inputs˅ѝवਜ਼অ‫ ݳ‬j 䗃ࠪⲴঅ
‫ݳ‬䳶ਸ
wE d
⧠൘ᡁԜሬࠪ Ⲵањ㺘⽪ˈԕ‫ׯ‬ᇎ⧠‫ޜ‬ᔿ˄4.21˅ѝࠪ⧠Ⲵ䲿ᵪⲴởᓖл䱽⌅ࡉǄ
ww ji
俆‫⌘ˈݸ‬᜿ᵳ٬wjiӵ㜭䙊䗷netjᖡ૽㖁㔌ⲴަԆ䜘࠶ǄᡰԕˈᡁԜਟԕ֯⭘䬮ᔿ㿴ࡉ˄chain
rule˅ᗇࡠ
wE d wE d wnet j
=
ww ji wnet j ww ji
wE d
= xji ˄4.22˅
wnet j
wE d
ᐢ⸕ㅹᔿ˄4.22˅ˈᡁԜ࢙лⲴԫ࣑ቡᱟѪ ሬࠪањᯩ‫Ⲵׯ‬㺘⽪ǄᡁԜ‫⅑׍‬㘳㲁
wnet j
є⿽ᛵߥ˖а⿽ᛵߥᱟঅ‫ ݳ‬j ᱟ㖁㔌Ⲵањ䗃ࠪঅ‫ˈݳ‬ਖа⿽ᛵߥᱟ j ᱟањ޵䜘অ‫ݳ‬Ǆ
ᛵߥ 1˖䗃ࠪঅ‫Ⲵݳ‬ᵳ٬䇝㓳⌅ࡉǄቡ䊑wjiӵ㜭䙊䗷netjᖡ૽ަ։Ⲵ㖁㔌аṧˈnetjӵ㜭
䙊䗷ojᖡ૽ަ։Ⲵ㖁㔌ǄᡰԕᡁԜਟԕ޽⅑֯⭘䬮ᔿ㿴ࡉᗇࠪ
wE d wE d wo j
= ˄4.23˅
wnet j wo j wnet j
俆‫ݸ‬ӵ㘳㲁ㅹᔿ˄4.23˅Ⲵㅜа亩
wE d w 1
= ¦ (t k o k ) 2
wo j wo j 2 koutputs
w
䲔Ҷᖃ k=j ᰦˈᡰᴹ䗃ࠪঅ‫ ݳ‬k Ⲵሬᮠ (t k o k ) 2 Ѫ 0ǄᡰԕᡁԜнᗵሩཊњ䗃ࠪ
wo j
অ‫ˈ઼≲ݳ‬ਚ䴰Ԕ k=jǄ
wE d w 1
(t j o j ) 2
wo j wo j 2
1 w (t j o j )
2(t j o j )
2 wo j
= (-tj-oj) ˄4.24˅
wo j
᧕лᶕ㘳㲁ㅹᔿ˄4.23˅ⲴㅜҼ亩Ǆᰒ❦oj=V˄netj˅ˈሬᮠ ቡᱟsigmoid࠭ᮠⲴሬ
wnet j
ᮠˈ㘼ᡁԜᐢ㓿ᤷࠪ䗷sigmoid࠭ᮠⲴሬᮠѪV(netj)(1-V(netj))Ǆᡰԕˈ
wo j wV (net j )
wnet j wnet j ˄4.25˅
o j (1 o j )
ᢺ㺘䗮ᔿ˄4.24˅઼˄4.25˅ԓ‫˄ޕ‬4.23˅ˈᡁԜᗇࡠ
wE d
= -(tj-oj)oj(1-oj) ˄4.26˅
wnet j
❦ਾоㅹᔿ˄4.21˅઼˄4.22˅ਸᒦˈᡁԜ‫᧘ׯ‬ሬࠪҶ䗃ࠪঅ‫Ⲵݳ‬䲿ᵪởᓖл䱽⌅ࡉ˖
wE d
'wji= K =K˄tj-oj˅oj˄1-oj˅xji ˄4.27˅
ww ji
⌘᜿䘉њ䇝㓳⌅ࡉᚠᚠᱟ㺘 4-2 ㇇⌅ѝⲴ˄T4.3˅઼˄T4.5˅Ⲵᵳ٬ᴤᯠ⌅ࡉǄ↔ཆˈ
wE d
ᡁԜਟԕਁ⧠ᔿ˄T4.3˅ѝⲴGkо ٬⴨ㅹǄ൘䘉а㢲Ⲵަ։䜘࠶ᡁԜሶ֯⭘Giᶕ㺘
wnet k
wE d
⽪ԫ᜿অ‫ݳ‬iⲴ Ǆ
wnet i
ᛵߥ 2˖䳀㯿অ‫Ⲵݳ‬ᵳ٬䇝㓳⌅ࡉǄሩҾ㖁㔌ѝⲴ޵䜘অ‫ݳ‬ᡆ㘵䈤䳀㯿অ‫Ⲵݳ‬ᛵߥˈ᧘
ሬwjiᗵ享㘳㲁wji䰤᧕ൠᖡ૽㖁㔌䗃ࠪˈӾ㘼ᖡ૽EdǄ⭡Ҿ䘉њ৏ഐˈᡁԜਁ⧠ᇊѹ㖁㔌ѝঅ
‫ݳ‬jⲴᡰᴹ・ণл⑨˄immediately downstream˅অ‫Ⲵݳ‬䳶ਸ˄ҏቡᱟ・ণ䗃‫ޕ‬ѝवਜ਼অ‫ݳ‬j
Ⲵ䗃ࠪⲴᡰᴹঅ‫˅ݳ‬ᱟᴹ⭘ⲴǄᡁԜ⭘DownStream(j)㺘⽪䘉ṧⲴঅ‫ݳ‬䳶ਸǄ⌘᜿netjਚ㜭䙊
䗷Downstream(j)ѝⲴঅ‫ݳ‬ᖡ૽㖁㔌䗃ࠪ˄޽ᖡ૽Ed˅Ǆᡰԕਟԕྲл᧘ሬ
wE d wE d wnet k
wnet j
¦
kDownStream ( j ) wnet j wnet j
wnet k
¦
kDownStream ( j )
Gk
wnet j
wnet k wo j
¦
kDownStream ( j )
Gk
wo j wnet j
(4.28)
wo j
¦
kDownStream ( j )
G k wkj
wnet j
¦
kDownStream ( j )
G k wkj o j (1 o j )
wE d
䟽ᯠ㓴㓷਴亩ᒦ֯⭘Gj㺘⽪ ˈᡁԜᗇࡠ
wnet j
Gj o j (1 o j ) ¦G w
k kj
kDownstream ( j )
઼
'wji = K Gj xji
кᔿቡᱟ⭡‫ޜ‬ᔿ˄4.20˅ᗇࡠⲴа㡜⌅ࡉˈ⭘ᶕᴤᯠԫ᜿ᴹੁᰐ⧟㖁㔌㔃ᶴ޵䜘অ‫Ⲵݳ‬
ᵳ٬Ǆ⌘᜿㺘 4-2 Ⲵᔿ˄T4.4˅ӵᱟ䘉њ⌅ࡉᖃ Downstream(j)=outputs ᰦⲴањ⢩ֻǄ
4.6 ਃੇՖ᫣㇍⌋Ⲻ䈪᱄
4.6.1 ᭬ᮑᙝૂቶ䜞ᶷቅ‫ٲ‬
↓ྲࡽ䶒ᡰ᧿䘠Ⲵˈ৽ੁՐ᫝㇇⌅ᇎ⧠Ҷа⿽ሩਟ㜭Ⲵ㖁㔌ᵳ٬オ䰤Ⲵởᓖл䱽ᩌ㍒ˈ
ᆳ䘝ԓൠ߿ሿ䇝㓳ṧֻⲴⴞḷ٬઼㖁㔌䗃ࠪ䰤Ⲵ䈟ᐞǄഐѪሩҾཊቲ㖁㔌ˈ䈟ᐞᴢ䶒ਟ㜭ਜ਼
ᴹཊњн਼Ⲵተ䜘ᶱሿ٬ˈởᓖл䱽ਟ㜭䲧‫ޕ‬䘉Ӌተ䜘ᶱሿ٬ѝⲴањǄഐ↔ˈሩҾཊቲ㖁
㔌ˈ৽ੁՐ᫝㇇⌅ӵ㜭‫؍‬䇱᭦ᮋࡠ䈟ᐞ E ⲴḀњተ䜘ᶱሿ٬ˈнаᇊ᭦ᮋࡠ‫ޘ‬ተⲴᴰሿ䈟
ᐞǄ
ቭ㇑㕪ѿሩ᭦ᮋࡠ‫ޘ‬ተᴰሿ䈟ᐞⲴ‫؍‬䇱ˈ৽ੁՐ᫝㇇⌅൘ᇎ䐥ѝᱟ䶎ᑨᴹ᭸Ⲵ࠭ᮠ䙬䘁
㇇⌅ǄሩҾᖸཊᇎ䱵Ⲵᓄ⭘ˈӪԜਁ⧠ተ䜘ᶱሿ٬Ⲵ䰞仈⋑ᴹᜣ䊑Ⲵ䛓Ѹѕ䟽ǄѪҶሩ䘉њ
䰞仈ᴹаӋⴤ㿲Ⲵ䇔䇶ˈ㘳㲁ਜ਼ᴹབྷ䟿ᵳ٬Ⲵ㖁㔌ˈᆳሩᓄ⵰㔤ᮠ䶎ᑨ儈Ⲵオ䰤ѝⲴ䈟ᐞᴢ
䶒˄⇿њᵳ٬а㔤˅Ǆᖃởᓖл䱽䲧‫⴨ޕ‬ሩḀњᵳⲴተ䜘ᶱሿ٬ᰦˈ⴨ሩަԆⲴᵳ䘉䟼ᵚᗵ
ᱟተ䜘ᶱሿ٬Ǆһᇎкˈ㖁㔌Ⲵᵳ䎺ཊˈ䈟ᐞᴢ䶒Ⲵ㔤ᮠ䎺ཊˈҏቡ䎺ਟ㜭Ѫởᓖл䱽ᨀ‫׋‬
ᴤཊⲴĀ䘳䙨䐟㓯āˈ䇙ởᓖл䱽⿫ᔰ⴨ሩ䈕অњᵳ٬Ⲵተ䜘ᶱሿ٬༴Ǆ
ሩተ䜘ᶱሿ٬ⲴㅜҼ⿽㿲⛩ᱟˈ㘳㲁䲿⵰䇝㓳ѝ䘝ԓ⅑ᮠⲴ໎࣐㖁㔌ᵳ٬Ⲵ╄ॆᯩᔿǄ
⌘᜿൘㇇⌅ѝˈྲ᷌ᢺ㖁㔌Ⲵᵳ٬ࡍ࿻ॆѪ᧕䘁Ҿ 0 Ⲵ٬ˈ䛓Ѹ൘ᰙᵏⲴởᓖл䱽↕僔ѝˈ
㖁㔌ሶ㺘⧠Ѫањ䶎ᑨᒣ━Ⲵ࠭ᮠˈ䘁լѪ䗃‫Ⲵޕ‬㓯ᙗ࠭ᮠǄ䘉ᱟഐѪ sigmoid ࠭ᮠᵜ䓛൘
ᵳ٬䶐䘁 0 ᰦ᧕䘁㓯ᙗ˄㿱മ 4-6 ѝⲴ sigmoid ࠭ᮠᴢ㓯˅
Ǆӵᖃᵳ٬ᐢ㓿໎䮯Ҷаᇊᰦ䰤
ѻਾˈᆳԜ᡽Պࡠ䗮ਟԕ㺘⽪儈ᓖ䶎㓯ᙗ㖁㔌࠭ᮠⲴ〻ᓖǄᡆ䇨ਟԕᵏᖵ൘ᵳオ䰤Ⲵ䘉њ४
ฏᆈ൘ᴤཊⲴተ䜘ᶱሿ٬ˈ䘉ṧਟԕ㺘⽪ᴤ༽ᵲⲴ࠭ᮠǄҏਟᐼᵋᖃᵳࡠ䗮䘉а⛩ᰦᆳԜᐢ
㓿䏣ཏ䶐䘁‫ޘ‬ተᴰሿ٬ˈণ‫ׯ‬ᆳᱟ䘉њ४ฏⲴተ䜘ᶱሿ٬ҏᱟਟԕ᧕ਇⲴǄ
ቭ㇑ᴹк䶒Ⲵ䇴䇪ˈӪԜሩ⭘ ANN 㺘⽪Ⲵ༽ᵲ䈟ᐞᴢ䶒Ⲵởᓖл䱽⨶䀓ᗇ䘈нཏˈ䘈

н⸕䚃ᴹօᯩ⌅㜭⺞࠷ൠ亴⍻ተ䜘ᶱሿ٬ӰѸᰦ‫ى‬Պሬ㠤ഠ䳮Ǆ⭘ᶕ㕃䀓ተ䜘ᶱሿ٬䰞仈Ⲵ
аӋᑨ㿱Ⲵ੟ਁᔿ㿴ࡉवᤜ˖
x 䊑‫ޜ‬ᔿ˄4.18˅᧿䘠Ⲵ䛓ṧѪởᓖᴤᯠ⌅ࡉ࣐ањߢ䟿亩Ǆߢ䟿ᴹᰦਟԕᑖ
ࣘởᓖл䱽䗷〻ˈߢ䗷⤝ゴⲴተ䜘ᴰл٬˄❦㘼৏ࡉкᆳҏਟԕᑖࣘởᓖл
䱽䗷〻ߢ䗷⤝ゴⲴ‫ޘ‬ተᴰሿ٬ࡠަԆተ䜘ᶱሿ٬ʽ˅Ǆ
x ֯⭘䲿ᵪⲴởᓖл䱽㘼нᱟⵏ↓Ⲵởᓖл䱽Ǆṩᦞ 4.4.3.3 ሿ㢲䇘䇪Ⲵˈởᓖ
л䱽Ⲵ䲿ᵪ䘁լሩҾ⇿њ䇝㓳ṧֻ⋯ањн਼Ⲵ䈟ᐞᴢ䶒ᴹ᭸л䱽ˈᆳ‫׍‬䶐
䘉ӋởᓖⲴᒣ൷ᶕ䙬䘁ሩҾᮤњ䇝㓳䳶ਸⲴởᓖǄ䘉Ӌн਼Ⲵ䈟ᐞᴢ䶒䙊ᑨ
ᴹн਼Ⲵተ䜘ᶱሿ٬ˈ䘉֯ᗇл䱽䗷〻нཚਟ㜭䲧‫ޕ‬ԫањተ䜘ᶱሿ٬Ǆ
x ֯⭘਼ṧⲴᮠᦞ䇝㓳ཊњ㖁㔌ˈն⭘н਼Ⲵ䲿ᵪᵳ٬ࡍ࿻ॆ⇿њ㖁㔌Ǆྲ᷌
н਼Ⲵ䇝㓳䗷〻ӗ⭏н਼Ⲵተ䜘ᶱሿ٬ˈ䛓Ѹሩ࠶⿫Ⲵ傼䇱䳶ਸᙗ㜭ᴰྭⲴ
㖁㔌㻛䘹ᤙǄᡆ㘵‫⮉؍‬ᡰᴹⲴ㖁㔌ˈᒦфᢺᆳԜᖃ֌ањ㖁㔌ĀငઈՊāˈ
ᆳԜⲴ䗃ࠪᱟ⇿њ㖁㔌䗃ࠪⲴᒣ൷٬˄ਟ㜭࣐ᵳ˅Ǆ
4.6.2 ࢃ侾㖇㔒Ⲻ㺞ᖷ㜳࣑
ӰѸ㊫රⲴ࠭ᮠਟԕ֯⭘ࡽ侸㖁㔌ᶕ㺘⽪઒˛ᖃ❦䘉њ䰞仈ⲴㆄṸ‫׍‬䎆Ҿ㖁㔌Ⲵᇭᓖ
઼␡ᓖǄቭ㇑ⴞࡽሩଚа᯿࠭ᮠਟԕ⭘ଚ⿽㊫රⲴ㖁㔌᧿䘠䘈⸕䚃ᗇᖸቁˈնᐢ㓿⸕䚃Ҷй
ња㡜ᙗⲴ㔃䇪˖
x ᐳቄ࠭ᮠǄԫօᐳቄ࠭ᮠਟԕ㻛ާᴹєቲঅ‫Ⲵݳ‬㖁㔌߶⺞㺘⽪ˈቭ㇑ሩҾᴰ
ൿⲴᛵߥˈᡰ䴰䳀㯿অ‫Ⲵݳ‬ᮠ䟿䲿⵰㖁㔌䗃‫ޕ‬ᮠ䟿Ⲵ໎࣐ᤷᮠ㓗໎䮯ǄѪҶ
䈤᰾䘉ᱟྲօᇎ⧠Ⲵˈ㘳㲁л䶒㺘⽪ԫօᐳቄ࠭ᮠⲴ䙊⭘ᯩṸ˖ሩҾ⇿ањ
ਟ㜭Ⲵ䗃‫ੁޕ‬䟿ˈࡋᔪн਼Ⲵ䳀㯿অ‫ˈݳ‬ᒦ䇮㖞ᆳⲴᵳ٬֯ᖃфӵᖃ䘉њ⢩
ᇊⲴੁ䟿䗃‫ࡠޕ‬㖁㔌ᰦ䈕অ‫ݳ‬㻛◰⍫Ǆ䘉ṧቡӗ⭏ҶањሩҾԫօ䗃‫ޕ‬ӵᴹ
ањঅ‫ݳ‬㻛◰⍫Ⲵ䳀㯿ቲǄ᧕лᶕᢺ䗃ࠪঅ‫ݳ‬ᇎ⧠Ѫањᡆ䰘ˈӵ⭡ᡰᐼᵋ
Ⲵ䗃‫⁑ޕ‬ᔿ◰⍫Ǆ
x 䘎㔝࠭ᮠǄԫօᴹ⭼Ⲵ䘎㔝࠭ᮠਟԕ⭡ањєቲⲴ㖁㔌ԕԫ᜿ሿⲴ䈟ᐞ˄൘
ᴹ䲀Ⲵ㤳ᮠл˅䙬䘁˄Cybenko 1989˗Hornik et al. 1989˅Ǆ䘉њ⨶䇪䘲⭘Ҿ
䳀㯿ቲ֯⭘ sigmoid অ‫ݳ‬ǃ䗃ࠪቲ֯⭘˄䶎䰸٬Ⲵ˅㓯ᙗঅ‫Ⲵݳ‬㖁㔌Ǆᡰ䴰
Ⲵ䳀㯿অ‫ݳ‬ᮠ䟿‫׍‬䎆Ҿ㾱䙬䘁Ⲵ࠭ᮠǄ
x ԫ᜿࠭ᮠǄԫ᜿࠭ᮠਟԕ㻛ањᴹйቲঅ‫Ⲵݳ‬㖁㔌ԕԫ᜿㋮ᓖ䙬䘁˄Cybenko
1988˅Ǆоࡽ䶒⴨਼ˈ䗃ࠪቲ֯⭘㓯ᙗঅ‫ˈݳ‬єњ䳀㯿ቲ֯⭘ sigmoid অ‫ˈݳ‬
⇿аቲᡰ䴰Ⲵঅ‫ݳ‬ᮠ䟿а㡜н⺞ᇊǄ䘉а㔃䇪Ⲵ䇱᰾ᯩ⌅Ѫ˖俆‫ݸ‬䈤᰾ԫօ
࠭ᮠਟԕ㻛䇨ཊተ䜘ॆ࠭ᮠⲴ㓯ᙗ㓴ਸ䙬䘁ˈ䘉Ӌተ䜘࠭ᮠⲴ٬䲔ҶḀњሿ
㤳തཆ䜭Ѫ 0˗❦ਾ䈤᰾єቲⲴ sigmoid অ‫ݳ‬䏣ԕӗ⭏㢟ྭⲴተ䜘䙬䘁Ǆ
䘉Ӌ㔃䇪㺘᰾ᴹ䲀␡ᓖⲴࡽ侸㖁㔌Ѫ৽ੁՐ᫝㇇⌅ᨀ‫׋‬Ҷ䶎ᑨᴹ㺘ᖱ࣋Ⲵ‫ٷ‬䇮オ䰤Ǆ❦
㘼䇠տл䶒а⛩ᱟ䟽㾱Ⲵ˖ởᓖл䱽ᱟӾањࡍ࿻Ⲵᵳ٬ᔰ࿻Ⲵˈഐ↔ᩌ㍒㤳ത䟼Ⲵ㖁㔌ᵳ
ੁ䟿ਟ㜭нवਜ਼ᡰᴹⲴᵳੁ䟿ǄHertz et al.˄1991˅ᨀ‫׋‬Ҷк䶒㔃䇪Ⲵᴤ䈖㓶Ⲵ䇘䇪Ǆ
4.6.3 ‫ٽ‬䇴グ䰪ᩒ㍘ૂᖈ㓩‫څ‬㖤
ᢺ৽ੁՐ᫝㇇⌅Ⲵ‫ٷ‬䇮オ䰤ᩌ㍒઼ަԆᆖҐ㇇⌅䟷ਆⲴᩌ㍒⴨∄䖳ᖸᴹ᜿ѹǄሩҾ৽ੁ
Ր᫝㇇⌅ˈ㖁㔌ᵳⲴ⇿а⿽ਟ㜭䍻٬䜭㺘⽪Ҷањਕ⌅н਼Ⲵ‫ٷ‬䇮ˈ৏ࡉк䜭൘ᆖҐಘⲴ㘳
㲁㤳ത޵Ǆᦒਕ䈍䈤ˈ䘉њ‫ٷ‬䇮オ䰤ᱟ n њ㖁㔌ᵳ٬Ⲵ n 㔤⅗∿オ䰤Ǆ⌘᜿䘉њオ䰤ᱟ䘎㔝
Ⲵˈ䘉оߣㆆṁᆖҐ઼ަԆสҾ⿫ᮓ㺘⽪Ⲵᯩ⌅Ⲵ‫ٷ‬䇮オ䰤ᆼ‫ޘ‬н਼Ǆ‫ٷ‬䇮オ䰤Ⲵ䘎㔝ᙗԕ
৺䈟ᐞ E ‫ޣ‬Ҿ‫ٷ‬䇮Ⲵ䘎㔝৲ᮠਟᗞ䘉єњһᇎˈሬ㠤Ҷањ㢟ᇊѹⲴ䈟ᐞởᓖˈѪᴰ֣‫ٷ‬
䇮Ⲵᩌ㍒ᨀ‫׋‬Ҷањ䶎ᑨᴹ⭘Ⲵ㔃ᶴǄ䘉њ㔃ᶴоสҾㅖਧⲴᾲᘥᆖҐ㇇⌅ⲴĀа㡜ࡠ⢩↺
ᒿāᩌ㍒Ⲵ㔃ᶴˈᡆ ID3 ઼ C4.5 ㇇⌅ѝሩߣㆆṁⲴㆰঅࡠ༽ᵲᒿᩌ㍒ᡰ⭘Ⲵ㔃ᶴ䜭ᆼ‫ޘ‬н
਼Ǆ
৽ੁՐ᫝㇇⌅Ӿ㿲⍻ᮠᦞѝ⌋ॆⲴᖂ㓣‫ٿ‬㖞ᱟӰѸ઒˛㋮⺞ൠ࡫⭫৽ੁՐ᫝ᆖҐⲴᖂ
㓣‫ٿ‬㖞ᱟᴹ䳮ᓖⲴˈഐѪᆳ‫׍‬䎆Ҿởᓖл䱽ᩌ㍒઼ᵳオ䰤㾶ⴆਟ㺘ᖱ࠭ᮠオ䰤ⲴᯩᔿⲴ⴨ӂ
֌⭘ᙗǄ❦㘼ˈਟԕᢺ䘉а‫ٿ‬㖞㋇⮕ൠ࡫⭫Ѫ在数据点之间平滑插值˄smooth interpolation
between data points˅Ǆྲ᷌㔉ᇊєњ↓ֻˈᆳԜѻ䰤⋑ᴹ৽ֻˈ৽ੁՐ᫝㇇⌅Պ‫ੁٮ‬Ҿᢺ䘉
є⛩ѻ䰤Ⲵ⛩ҏḷ䇠Ѫ↓ֻǄֻྲˈ൘മ 4-5 ⭫ࠪⲴߣㆆ䶒ѝਟԕⴻࡠ䘉а⛩ˈ䇝㓳ṧֻⲴ
⢩ᇊṧᵜӗ⭏Ҷᒣ━ਈॆⲴߣㆆ४ฏǄ
4.6.4 䳆㰅ቸ㺞⽰
৽ੁՐ᫝㇇⌅Ⲵањ䘧ӪⲴ⢩ᙗᱟˈᆳ㜭ཏ൘㖁㔌޵䜘Ⲵ䳀㯿ቲਁ⧠ᴹ⭘Ⲵѝ䰤㺘⽪Ǆ
ഐѪ䇝㓳ṧֻӵवਜ਼㖁㔌䗃‫઼ޕ‬䗃ࠪˈᵳ٬䈳㢲Ⲵ䗷〻ਟԕ㠚⭡ൠ䇮㖞ᵳ٬ˈᶕᇊѹ൘ᴰሿ
ॆ䈟ᐞᒣᯩ E ѝᴰᴹ᭸Ⲵԫօ䳀㯿অ‫ݳ‬㺘⽪Ǆ䘉㜭ཏᕅሬ৽ੁՐ᫝㇇⌅ᇊѹᯠⲴ䳀㯿ቲ⢩
ᖱˈ䘉Ӌ⢩ᖱ൘䗃‫ޕ‬ѝ⋑ᴹ᰾⺞㺘⽪ࠪᶕˈնত㜭ᦅ᥹䗃‫ޕ‬ᇎֻѝоᆖҐⴞḷ࠭ᮠᴰ⴨‫Ⲵޣ‬
⢩ᖱǄ
ֻྲˈ㘳㲁മ 4-7 ᡰ⽪Ⲵ㖁㔌Ǆ䘉䟼ˈ8 њ㖁㔌䗃‫ޕ‬о 3 њ䳀㯿অ‫⴨ݳ‬䘎ˈ3 њ䳀㯿অ

‫ݳ‬৸‫⅑׍‬䘎᧕ࡠ 8 њ䗃ࠪঅ‫ݳ‬Ǆ⭡Ҿ䘉ṧⲴ㔃ᶴˈ3 њ䳀㯿অ‫ݳ‬ᗵ享䟽ᯠ㺘⽪ 8 њ䗃‫ޕ‬٬ˈ
ԕḀ⿽ᯩᔿᦅ᥹䗃‫ޣ⴨Ⲵޕ‬⢩ᖱˈԕ‫ׯ‬䘉њ䳀㯿ቲⲴ㺘⽪ਟԕ㻛䗃ࠪঅ‫⭘ݳ‬ᶕ䇑㇇↓⺞Ⲵⴞ
ḷ٬Ǆ
Inputs-䗃‫ޕ‬
Outputs-䗃ࠪ
Input-䗃‫ޕ‬٬
Output-䗃ࠪ٬
Hidden Values-䳀㯿٬
മ 4-7 ᆖҐࡠⲴ䳀㯿ቲ㺘⽪
䘉њ 8u3u8 Ⲵ㖁㔌㻛䇝㓳ԕᆖҐᚂㅹ࠭ᮠˈ֯⭘മѝᡰ⽪Ⲵ 8 њ䇝㓳ṧֻǄ൘ 5000 䖞˄epochs˅
䇝㓳ѻਾˈ3 њ䳀㯿অ‫⭘֯ݳ‬മਣ‫Ⲵח‬㕆⸱ᯩᔿᶕ㕆⸱ 8 њ⴨ӂн਼Ⲵ䗃‫ޕ‬Ǆ⌘᜿ྲ᷌ᢺ㕆⸱ਾⲴ
٬ഋ㠽ӄ‫ޕ‬Ѫ 0 ઼ 1ˈ䛓Ѹ㔃᷌ᱟ 8 њн਼٬Ⲵḷ߶Ҽ䘋٬㕆⸱Ǆ
& & &

㘳㲁䇝㓳മ 4-7 ᡰ⽪Ⲵ㖁㔌ˈᶕᆖҐㆰঅⲴⴞḷ࠭ᮠ f( x )= x ˈަѝ x ᱟਜ਼ᴹгњ 0 ઼
ањ 1 Ⲵੁ䟿Ǆ㖁㔌ᗵ享ᆖՊ൘ 8 њ䗃ࠪঅ‫ݳ‬䟽⧠䘉 8 њ䗃‫ޕ‬Ǆቭ㇑䘉ᱟањㆰঅⲴ࠭ᮠˈ
ն⧠൘䲀ࡦ㖁㔌ਚ㜭֯⭘ 3 њ䳀অ‫ݳ‬ǄᡰԕˈᆖҐࡠⲴ 3 њ䳀㯿অ‫ݳ‬ᗵ享ᦅ᥹տᶕ㠚 8 њ䗃
‫ޕ‬অ‫Ⲵݳ‬ᡰᴹ‫ޣ‬䭞ؑ᚟Ǆ
ᖃ৽ੁՐ᫝㇇⌅㻛⭘ᶕᆼᡀ䘉њԫ࣑ᰦˈ֯⭘ 8 њਟ㜭ੁ䟿֌Ѫ䇝㓳ṧֻˈᆳᡀ࣏ൠᆖ
ՊҶⴞḷ࠭ᮠǄởᓖл䱽Ⲵ৽ੁՐ᫝㇇⌅ӗ⭏Ⲵ䳀㯿ቲ㺘⽪ᱟӰѸ઒˛䙊䗷࠶᷀ᆖҐࡠⲴ㖁
㔌ሩҾ 8 њਟ㜭䗃‫ੁޕ‬䟿ӗ⭏Ⲵ䳀㯿অ‫Ⲵݳ‬٬ˈਟԕⴻࠪᆖࡠⲴ㕆⸱઼⟏⸕Ⲵሩ 8 њ٬֯⭘
3 սḷ߶Ҽ䘋ࡦ㕆⸱⴨਼˄ҏቡᱟ 000ˈ001ˈ010ˈĂĂˈ111˅Ǆമ 4-7 ᱮ⽪Ҷ৽ੁՐ᫝㇇
⌅Ⲵа⅑䘀㹼ѝ䇑㇇ࠪⲴ䘉 3 њ䳀㯿অ‫࠷⺞Ⲵݳ‬٬Ǆ
ཊቲ㖁㔌൘䳀㯿ቲ㠚ࣘਁ⧠ᴹ⭘㺘⽪Ⲵ㜭࣋ᱟ ANN ᆖҐⲴањ‫ޣ‬䭞⢩ᙗǄо䛓Ӌӵ䲀

Ҿ֯⭘Ӫ㊫䇮䇑㘵ᨀ‫Ⲵ׋‬亴ᇊѹ⢩ᖱⲴᆖҐᯩ⌅⴨∄ˈᆳᨀ‫׋‬Ҷа⿽⴨ᖃ䟽㾱Ⲵ⚥⍫ᙗüü
‫ݱ‬䇨ᆖҐಘࡋ䙐ࠪ䇮䇑㘵⋑ᴹ᰾⺞ᕅ‫Ⲵޕ‬⢩ᖱǄᖃ❦䘉Ӌࡋ䙐ࠪⲴ⢩ᖱаᇊᱟ㖁㔌䗃‫Ⲵޕ‬
sigmoid অ‫࠭ݳ‬ᮠਟԕ䇑㇇ࠪⲴǄ⌘᜿㖁㔌ѝ֯⭘Ⲵঅ‫ݳ‬ቲ䎺ཊˈቡਟԕࡋ䙐ࠪ䎺༽ᵲⲴ⢩
ᖱǄ4.7 㢲㾱䇘䇪ⲴӪ㝨䇶࡛ᓄ⭘ᨀ‫׋‬Ҷ䳀㯿অ‫ݳ‬⢩ᖱⲴਖањֻᆀǄ
ѪҶ໎ᕪሩ䘉њֻᆀѝ৽ੁՐ᫝㇇⌅᫽֌Ⲵⴤ㿲⨶䀓ˈ䇙ᡁԜᴤ䈖㓶ൠ࠶᷀ởᓖл䱽䗷
〻ѝⲴާփ᫽֌ķǄ֯⭘㺘 4-2 ѝⲴ㇇⌅䇝㓳മ 4-7 ѝⲴ㖁㔌ˈ䇮㖞ࡍ࿻Ⲵᵳ٬Ѫ४䰤˄-0.1,
0.1˅ѝⲴ䲿ᵪᮠˈᆖҐ䙏⦷K=0.3ˈ⋑ᴹᵳߢ䟿˄ণD=0˅Ǆ֯⭘ަԆⲴᆖҐ䙏⦷઼֯⭘䶎 0
Ⲵߢ䟿ᗇࡠⲴ㔃᷌⴨լǄ മ 4-7 ѝᱮ⽪Ⲵ䳀㯿অ‫ݳ‬㕆⸱ᱟ൘ᢗ㹼Ҷ㇇⌅Ⲵཆቲ䇝㓳䘝ԓ 5000
⅑ਾᗇࡠⲴ˄ҏቡᱟሩ 8 њ䇝㓳ṧֻⲴ⇿ањ䘝ԓ 5000 ⅑˅Ǆ❦㘼੨ᕅᡁԜ⌘᜿Ⲵབྷཊᮠᵳ
٬ਈॆᱟਁ⭏൘ࡽ 2500 ⅑ⲴǄ
ᡁԜਟԕ᧿㔈ࠪ䗃ࠪ䈟ᐞⲴᒣᯩ⴨ሩởᓖл䱽ᩌ㍒↕ᮠⲴ࠭ᮠᴢ㓯ˈ䘉ṧቡਟԕⴤ᧕㿲
ሏ৽ੁՐ᫝㇇⌅Ⲵởᓖл䱽ᩌ㍒Ⲵ᭸᷌Ǆᆳᱮ⽪൘മ 4-8 ѝᴰк䶒Ⲵᴢ㓯മѝǄ䘉ᑵമⲴ 8
ᶑᴢ㓯ሩᓄ 8 њ㖁㔌䗃ࠪˈ⇿аᶑᴢ㓯䜭ᱮ⽪Ҷ⴨ᓄⲴ㖁㔌䗃ࠪሩᡰᴹ䇝㓳ṧֻⲴ䈟ᐞᒣᯩ
઼Ǆ⁚䖤㺘⽪৽ੁՐ᫝㇇⌅Ⲵᴰཆቲ䘝ԓⲴ⅑ᮠǄྲമѝᡰᱮ⽪Ⲵˈ⇿њ䗃ࠪⲴ䈟ᐞᒣᯩ઼
䲿⵰ởᓖл䱽䗷〻㘼л䱽ˈḀӋঅ‫ݳ‬ᘛаӋˈḀӋঅ‫ݳ‬䖳ធǄ
䳀㯿অ‫ݳ‬㺘⽪Ⲵ╄ਈ䗷〻ਟԕ൘മ 4-8 ⲴㅜҼᑵമѝⴻࡠǄ䘉ᑵമᱮ⽪ҶሩҾањਟ㜭

Ⲵ䗃‫˄ޕ‬䘉ᑵമሩᓄⲴᱟ 01000000˅㖁㔌䇑㇇ࠪⲴйњ䳀㯿অ‫ݳ‬٬Ǆ઼ࡽ䶒аṧˈ⁚䖤㺘
⽪䇝㓳ᗚ⧟Ⲵ⅑ᮠǄྲമѝᡰᱮ⽪Ⲵˈ䘉њ㖁㔌᭦ᮋࡠമ 4-7 ѝ㔉ࠪⲴᴰ㓸Ⲵ㕆⸱ѻࡽ㓿শ
Ҷᖸཊн਼Ⲵ㕆⸱Ǆ
ᴰਾˈമ 4-8 ѝⲴㅜ 3 ᑵമ⭫ࠪҶ㖁㔌ѝ਴њᵳ٬Ⲵ╄ਈ䗷〻Ǆ䘉ᑵമᱮ⽪Ҷ䘎᧕ 8 њ

䗃‫ޕ‬অ‫઼˄ݳ‬ањᑨ䟿‫ٿ‬㖞䗃‫˄ޕ‬constant bias input˅˅ࡠ 3 њ䳀অ‫ݳ‬ѻаⲴᵳ٬Ⲵ╄ਈ䗷〻Ǆ
⌘᜿䘉њ䳀㯿অ‫ݳ‬ᵳ٬Ⲵᱮ㪇ਈॆо䳀㯿ቲ㕆⸱઼䗃ࠪ䈟ᐞᒣᯩⲴᱮ㪇ਈॆа㠤Ǆ䘉䟼᭦ᮋ
ķ
䘉њֻᆀⲴⓀԓ⸱ਟԕӾhttp://www.cs.cmu.edu/~tom/mlbook.htmlᗇࡠǄ
᧕䘁 0 Ⲵᵳ٬ᱟ‫ٿ‬㖞ᵳw0Ǆ
Sum of squared errors for each output unit-⇿њ䗃ࠪঅ‫Ⲵݳ‬䈟ᐞᒣᯩ઼
Hidden unit encoding for input 01000000-䗃‫ ޕ‬01000000 Ⲵ䳀㯿অ‫ݳ‬㕆⸱
Weights from inputs to one hidden unit-䗃‫ࡠޕ‬ањ䳀㯿অ‫Ⲵݳ‬ᵳ
മ 4-8 ᆖҐ 8u3u8 㖁㔌
ᴰкമᱮ⽪Ҷ䲿⵰䇝㓳䘝ԓ⅑ᮠ˄䖞ᮠ˅Ⲵ໎࣐ˈ8 њ䗃‫Ⲵޕ‬䈟ᐞᒣᯩ઼Ⲵ╄ਈǄѝമᱮ⽪ҶሩҾ
䗃‫ޕ‬ѢĀ01000000āⲴ䳀㯿ቲ㺘⽪Ⲵ╄ਈǄлമᱮ⽪Ҷ 3 њ䳀㯿অ‫ݳ‬ѻаⲴᵳ٬╄ਈ䗷〻Ǆ
4.6.5 ⌑ौθ䗽ᓜᤕਾૂ‫ᦤ࡚↘ڒ‬
൘㺘 4-2 ሩ৽ੁՐ᫝㇇⌅Ⲵ᧿䘠ѝˈ⋑ᴹᤷᇊ㇇⌅֯⭘Ⲵ㓸→ᶑԦǄ㓸→ᵳ٬ᴤᯠᗚ⧟
Ⲵਸ䘲ᶑԦᱟӰѸ઒˛ᖸ᰾ᱮˈа⿽䘹ᤙᱟ㔗㔝䇝㓳ⴤࡠሩ䇝㓳ṧֻⲴ䈟ᐞ E 䱽վ㠣Ḁњ
亴‫ݸ‬ᇊѹⲴ䰸٬ѻлǄһᇎкˈ䘉нᱟањྭⲴㆆ⮕ˈഐѪ৽ੁՐ᫝㇇⌅ᇩ᱃䗷ᓖᤏਸ䇝㓳
ṧֻˈ䱽վҶሩҾަԆᵚ㿱䗷ᇎֻⲴ⌋ॆ㋮ᓖǄ
ѪҶⴻࠪ֯䇝㓳ᮠᦞк䈟ᐞᴰሿॆⲴড䲙ˈ㘳㲁䈟ᐞ E ᱟྲօ䲿⵰ᵳ٬䘝ԓ⅑ᮠਈॆ
ⲴǄമ 4-9 ᱮ⽪Ҷєњ⴨ᖃިරⲴ৽ੁՐ᫝㇇⌅ᓄ⭘ѝⲴ䘉⿽ਈॆǄ俆‫ݸ‬㘳㲁മѝк䶒аᑵ
ᴢ㓯മǄєᶑᴢ㓯ѝ䖳վⲴаᶑᱮ⽪Ҷ൘䇝㓳䳶ਸкⲴ䈟ᐞ E 䲿⵰ởᓖл䱽䘝ԓ⅑ᮠⲴ໎
࣐㘼অ䈳л䱽Ǆ䖳儈Ⲵᴢ㓯ᱟ൘ањо䇝㓳ṧֻн਼Ⲵ傼䇱䳶ਸⲴᇎֻк⍻ࡠⲴ䈟ᐞ E Ⲵ
ᛵߥǄ䘉ᶑ㓯⍻䟿Ҷ㖁㔌Ⲵ⌋ॆ㋮ᓖ˄generalization accuracy˅üü㖁㔌ᤏਸ䇝㓳ᮠᦞཆⲴ
ᇎֻⲴ㋮ᓖǄ
⌘᜿൘傼䇱ṧֻк⍻䟿ࡠⲴⲴ䈟ᐞEķ‫ݸ‬л䱽ˈ❦ਾкॷˈቭ㇑൘䇝㓳ṧֻкⲴ䈟ᐞᤱ
㔝л䱽ǄѪӰѸՊਁ⭏䘉⿽⧠䊑઒˛䘉ᱟഐѪ䘉Ӌᵳ٬ᤏਸҶ䇝㓳ṧֻⲴĀ⢩ᔲᙗā
˄idiosyncrasy˅ˈ㘼䘉њĀ⢩ᔲᙗāሩҾṧֻⲴа㡜࠶ᐳ⋑ᴹԓ㺘ᙗǄANNѝབྷ䟿Ⲵᵳ٬৲
ᮠѪᤏਸ䘉ṧⲴĀ⢩ᔲᙗāᨀ‫׋‬ҶᖸབྷⲴ㠚⭡ᓖǄ
ѪӰѸ䗷ᓖᤏਸᖰᖰᱟਁ⭏൘䘝ԓⲴਾᵏˈ㘼нᱟ䘝ԓⲴᰙᵏ઒˛䇮ᜣ㖁㔌Ⲵᵳ٬ᱟ㻛
ࡍ࿻ॆѪሿ䲿ᵪ٬ⲴǄ֯⭘䘉ӋࠐѾаṧⲴᵳ٬ӵ㜭᧿䘠䶎ᑨᒣ━Ⲵߣㆆ䶒Ǆ䲿⵰䇝㓳Ⲵ䘋
㹼ˈаӋᵳ٬ᔰ࿻໎䮯ˈԕ䱽վ൘䇝㓳ᮠᦞкⲴ䈟ᐞˈ਼ᰦᆖҐࡠⲴߣㆆ䶒Ⲵ༽ᵲᓖҏ൘ᨀ
儈ǄҾᱟˈ䲿⵰ᵳ٬䈳ᮤ䘝ԓ⅑ᮠⲴ໎࣐ˈ৽ੁՐ᫝㇇⌅㧧ᗇⲴ‫ٷ‬䇮Ⲵᴹ᭸༽ᵲᓖҏ൘໎࣐Ǆ
ྲ᷌ᵳ٬䈳ᮤ䘝ԓ⅑ᮠ䏣ཏཊˈ৽ੁՐ᫝㇇⌅㓿ᑨՊӗ⭏䗷ᓖ༽ᵲⲴߣㆆ䶒ˈᤏਸҶ䇝㓳ᮠ
ᦞѝⲴಚ༠઼䇝㓳ṧֻѝ⋑ᴹԓ㺘ᙗⲴ⢩ᖱǄ䘉њ䗷ᓖᤏਸ䰞仈оߣㆆṁᆖҐѝⲴ䗷ᓖᤏਸ
ķ
䈁⌘˖৏Җ↔༴ᴹ䈟ˈ৏ਕѪgeneralization accuracy‫ݸ‬л䱽ਾкॷˈᱮ❦䘉䟼Ⲵgeneralization accuracyᓄ
Ѫerror E˅Ǆ
䰞仈⴨լ˄㿱ㅜ 3 ㄐ˅Ǆ
ᴹࠐ⿽ᢰᵟਟԕ⭘Ҿ䀓ߣ৽ੁՐ᫝ѝⲴ䗷ᓖᤏਸ䰞仈Ǆа⿽ᯩ⌅㻛〠Ѫᵳ٬㺠߿
˄weight decay˅ˈᆳ൘⇿⅑䘝ԓ䗷〻ѝԕḀњሿഐᆀ䱽վ⇿њᵳ٬Ǆ䘉ㅹ᭸Ҿ‫؞‬᭩ E Ⲵᇊѹˈ
࣐‫ޕ‬ањо㖁㔌ᵳ٬Ⲵᙫ䟿⴨ᓄⲴ᜙㖊亩Ǆ↔ᯩ⌅Ⲵࣘᵪ൘Ҿ‫؍‬ᤱᵳ٬䖳ሿˈӾ㘼֯ᆖҐ䗷
〻ੁ⵰༽ᵲߣㆆ䶒Ⲵ৽ᯩੁ‫ٿ‬㖞Ǆ
‫ݻ‬ᴽ䗷ᓖᤏਸ䰞仈Ⲵањᴰᡀ࣏Ⲵᯩ⌅ˈቡᱟ൘䇝㓳ᮠᦞཆ޽Ѫ㇇⌅ᨀ‫׋‬а྇傼䇱ᮠᦞ
˄validation data˅Ǆ㇇⌅൘֯⭘䇝㓳䳶ਸ傡ࣘởᓖл䱽ᩌ㍒Ⲵ਼ᰦˈⴁ㿶ሩҾ䘉њ傼䇱䳶ਸ
Ⲵ䈟ᐞǄᵜ䍘кˈ䘉⴨ᖃҾ‫ݱ‬䇨㇇⌅ᵜ䓛⭫ࠪമ 4-9 ѝᱮ⽪Ⲵєᶑᴢ㓯Ǆ㇇⌅ᓄ䈕䘋㹼ཊቁ
⅑ᵳ٬䈳ᮤ䘝ԓ઒˛ᱮ❦ˈᓄ䈕֯⭘൘傼䇱䳶ਸкӗ⭏ᴰሿ䈟ᐞⲴ䘝ԓ⅑ᮠˈഐѪ䘉ᱟ㖁㔌
ᙗ㜭ሩҾᵚ㿱䗷ᇎֻⲴᴰྭ㺘ᖱǄ൘䘉⿽ᯩ⌅Ⲵިරᇎ⧠ѝˈ㖁㔌Ⲵᵳ٬㻛‫⮉؍‬єԭᤧ䍍˖
аԭ⭘ᶕ䇝㓳ˈ㘼ਖаԭᤧ䍍֌ѪⴞࡽѪ→ᙗ㜭ᴰྭⲴᵳˈ㺑䟿Ⲵḷ߶ᱟᆳԜሩҾ傼䇱䳶ਸ
Ⲵ䈟ᐞǄаᰖ䇝㓳ࡠⲴᵳ٬൘傼䇱䳶ਸкⲴ䈟ᐞ∄‫؍‬ᆈⲴᵳ٬Ⲵ䈟ᐞ儈ˈ䇝㓳㻛㓸→ˈᒦф
䘄എ‫؍‬ᆈⲴᵳ٬֌Ѫᴰ㓸Ⲵ‫ٷ‬䇮Ǆᖃ䘉њ䗷〻㻛ᓄ⭘ࡠമ 4-9 ѝᴰкമⲴᛵߥᰦˈᆳሶ䗃ࠪ
൘ 9100 ⅑䘝ԓਾ㖁㔌ᗇࡠⲴᵳ٬Ǆമ 4-9 ⲴㅜҼᑵᴢ㓯മᱮ⽪ˈнᱟᙫ㜭᰾ᱮ⺞ᇊ傼䇱䳶
ਸօᰦ䗮ࡠᴰሿ䈟ᐞǄ൘䘉ᑵമѝˈ傼䇱䳶ਸⲴ䈟ᐞ‫ݸ‬л䱽ˈ❦ਾкॷˈ❦ਾ޽л䱽Ǆᡰԕ
ᗵ享⌘᜿䚯‫ݽ‬䭉䈟Ⲵ㔃䇪˖൘ 850 ⅑䘝ԓਾ㖁㔌ࡠ䗮ҶᆳⲴᴰሿ傼䇱䳶ਸ䈟ᐞǄ
Error versus weight updates(example 1)-䈟ᐞ⴨ሩᵳ٬ᴤᯠ⅑ᮠਈॆᴢ㓯˄ֻ 1˅
Error versus weight updates(example 2)- 䈟ᐞ⴨ሩᵳ٬ᴤᯠ⅑ᮠਈॆᴢ㓯˄ֻ 2˅
Error-䈟ᐞ
Number of weight updates-ᵳ٬ᴤᯠ⅑ᮠ
Training set error-䇝㓳䳶ਸⲴ䈟ᐞ
Validation set error-傼䇱䳶ਸⲴ䈟ᐞ
മ 4-9 єњн਼ᵪಘӪᝏ⸕ԫ࣑Ⲵ䈟ᐞ E ⴨ሩᵳ٬ᴤᯠ⅑ᮠⲴਈॆᴢ㓯
є⿽ᛵߥлˈ൘䇝㓳ṧֻкⲴ䈟ᐞ E 䜭অ䈳л䱽ˈഐѪởᓖл䱽Ⲵⴞḷᱟᴰሿॆ䘉њ䈟ᐞǄሩҾঅ
⤜Ⲵ傼䇱䳶ਸѝⲴṧֻˈ䈟ᐞ E 䙊ᑨ‫ݸ‬л䱽ˈ❦ਾ䈟ᐞਟ㜭ഐѪ䗷ᓖᤏਸ䇝㓳ṧֻ㘼кॷǄᴰᴹਟ
㜭↓⺞⌋ॆࡠᵚ㿱䗷ᮠᦞⲴ㖁㔌ᱟሩҾ傼䇱䳶ਸᴹᴰሿ䈟ᐞⲴ㖁㔌Ǆ⌘᜿൘ㅜҼᑵᴢ㓯മѝˈᗵ享
ሿᗳн㾱䗷ᰙ‫→ڌ‬䇝㓳ˈഐѪ൘傼䇱䳶ਸкⲴ䈟ᐞ E ൘䘝ԓࡠ 850 ⅑ᰦᔰ࿻кॷ㘼ਾ৸л䱽Ǆ
а㡜㘼䀰ˈ䗷ᓖᤏਸ䰞仈ԕ৺‫ݻ‬ᴽᆳⲴᯩ⌅ᱟањἈ᡻Ⲵ䰞仈Ǆк䶒ⲴӔ৹傼䇱ᯩ⌅൘
ਟ㧧ᗇ仍ཆⲴᮠᦞᨀ‫׋‬傼䇱䳶ਸᰦᐕ֌ᗇᴰྭǄ❦㘼нᒨⲴᱟˈ䗷ᓖᤏਸⲴ䰞仈ሩሿ䇝㓳䳶
ਸᴰѕ䟽Ǆ൘䘉⿽ᛵߥлˈᴹᰦ֯⭘а⿽〠ѪĀk-foldӔ৹傼䇱˄k-fold cross-validation˅āⲴ
ᯩ⌅ˈ䘉⿽ᯩ⌅䘋㹼k⅑н਼ⲴӔ৹傼䇱ˈ⇿⅑֯⭘ᮠᦞⲴн਼࠶ࢢ֌Ѫ䇝㓳䳶ਸ઼傼䇱䳶
ਸˈ❦ਾሩ㔃᷌䘋㹼ᒣ൷Ǆ൘䘉⿽ᯩ⌅Ⲵањ⡸ᵜѝˈᢺਟ‫Ⲵ⭘֯׋‬mњᇎֻ࠶ࢢᡀkњн
⴨ӔⲴᆀ䳶ˈ⇿њᆀ䳶ᴹm/kњᇎֻǄ❦ਾˈ䘀㹼k⅑Ӕ৹傼䇱䗷〻ˈ⇿а⅑֯⭘н਼Ⲵᆀ䳶
֌Ѫ傼䇱䳶ਸˈᒦਸᒦަԆⲴᆀ䳶֌Ѫ䇝㓳䳶ਸǄҾᱟˈ⇿ањṧֻՊ൘а⅑ᇎ傼ѝ㻛⭘֌
傼䇱䳶ਸⲴᡀઈˈ൘k-1 ⅑ᇎ傼ѝ⭘֌䇝㓳䳶ਸⲴᡀઈǄ൘⇿⅑䈅傼ѝˈ䜭֯⭘к䶒䇘䇪Ⲵ
Ӕ৹傼䇱䗷〻ˈᶕߣᇊ൘傼䇱䳶ਸкਆᗇᴰ֣ᙗ㜭Ⲵ䘝ԓ⅑ᮠiǄ❦ਾ䇑㇇䘉ӋiⲴ൷٬ i ˈ
ᴰਾ䘀㹼а⅑৽ੁՐ᫝㇇⌅ˈ䇝㓳ᡰᴹmњķᇎֻᒦ䘝ԓ i ⅑ˈ↔ᰦ⋑ᴹ傼䇱䳶ਸǄ䘉њ䗷〻
оㅜ 5 ㄐ᧿䘠ⲴสҾᴹ䲀ᮠᦞ∄䖳є⿽ᆖҐᯩ⌅Ⲵ䗷〻ᖸ⴨䘁Ǆ
4.7 ⽰ׁφӰ㝮䇼ࡡ
ѪҶ䈤᰾৽ੁՐ᫝㇇⌅ᓄ⭘ѝⲴаӋᇎ䱵Ⲵ䇮䇑䰞仈ˈ䘉а㢲䇘䇪ᢺ䘉њ㇇⌅ᓄ⭘ࡠӪ
㝨䇶࡛ⲴᆖҐԫ࣑Ǆ䘉а㢲⭘ᶕӗ⭏䘉њֻᆀⲴᡰᴹമ‫ۿ‬ᮠᦞ઼ԓ⸱䜭ਟԕӾԕл㖁൰ᗇ
ࡠ˖http://www.cs.cmu.edu//~tom/mlbook.htmlˈ਼ᰦ䘈ᴹྲօ֯⭘䘉Ӌԓ⸱Ⲵᆼᮤ᮷ẓǄ䈫
㘵ਟԕ㠚ᐡ䘋㹼䈅傼Ǆ
4.7.1 Աࣗ
䘉䟼ⲴᆖҐԫ࣑ᱟ࠶㊫н਼ӪⲴн਼ုᘱⲴᩴᖡമ‫ۿ‬ǄᡁԜ᭦䳶Ҷ 20 њн਼ⲴӪⲴᩴ
ᖡമ‫⇿ˈۿ‬њӪབྷ㓖ᴹ 32 ᕐമ‫ˈۿ‬ሩᓄ䘉њӪн਼Ⲵ㺘ᛵ˄ᘛҀˈ⋞їˈ᝔ᙂˈѝᙗ˅˗Ԇ
ԜⴻⲴн਼ᯩੁ˄ᐖˈਣˈ↓ࡽˈк˅˗઼ԆԜᱟ੖ᡤཚ䱣䮌ǄӾമ 4-10 Ⲵ⽪ֻമ‫ۿ‬ѝਟԕ
ⴻࡠˈӪਾ䶒Ⲵ㛼ᲟǃクⲴ㺓ᴽǃ઼Ӫ㝨൘മ‫ۿ‬ѝⲴս㖞ҏ䜭ᴹᐞᔲǄᡁԜ‫ޡ‬᭦䳶Ҷ 624
ᑵ⚠ᓖമ‫⇿ˈۿ‬аᑵⲴ࠶䗘⦷Ѫ 120u128ˈമ‫⇿Ⲵۿ‬њ‫ۿ‬㍐֯⭘ 0˄唁㢢˅ࡠ 255˄ⲭ㢢˅
Ⲵ⚠ᓖ٬᧿䘠Ǆ
Ӿ䘉Ӌമ‫ۿ‬ᮠᦞѝਟԕᆖҐᖸཊн਼Ⲵⴞḷ࠭ᮠǄֻྲˈᡁԜਟԕ䇝㓳ањ ANNˈ֯
㔉ᇊаᑵമ‫ۿ‬䗃‫ޕ‬ᰦ䗃ࠪ䘉њӪⲴᜏаḷ䇶˄identity˅ǃ㝨Ⲵᵍੁǃᙗ࡛ǃᱟ੖ᑖཚ䱣䮌ㅹǄ
ᡰᴹ䘉Ӌⴞḷ࠭ᮠਟԕԕᖸ儈Ⲵ㋮ᓖӾ䘉ӋᮠᦞѝᆖҐࡠˈ啃࣡䈫㘵Ԝ㠚㹼䈅傼Ǆ൘ᵜ㢲ਾ
䶒Ⲵ䜘࠶ˈᡁԜ㘳㲁ањ⢩ᇊⲴԫ࣑˖ᆖҐമ‫ۿ‬ѝӪ㝨Ⲵᵍੁ˄ᐖˈਣˈ↓ࡽˈ䘈ᱟк˅Ǆ
30u32resolution input images- 30u32 ࠶䗘⦷Ⲵ䗃‫ޕ‬മ‫ۿ‬
Network weights after 1 iteration through each training example- ሩ⇿њ䇝㓳ṧֻ䘝ԓ

1 ⅑ਾⲴ㖁㔌ᵳ٬
Network weights after 100 iteration through each training example- ሩ⇿њ䇝㓳ṧֻ䘝ԓ
100 ⅑ਾⲴ㖁㔌ᵳ٬
left: ᐖ
straight: ࡽ
ķ
䈁⌘˖৏Җ↔༴䈟Ѫn
right: ਣ
up: к
മ 4-10 ᆖҐ䇶࡛Ӫ㝨ᵍੁⲴӪᐕ⾎㓿㖁㔌
䘉䟼֯⭘Ӫ㝨Ⲵ⚠ᓖമ‫˄ۿ‬㿱ᴰка㹼˅䇝㓳ањ 960u3u4 Ⲵ㖁㔌ˈᶕ亴⍻ањӪᱟ൘ੁᐖǃੁਣǃ

ੁࡽ䘈ᱟੁкⴻǄ൘֯⭘Ҷ 260 ᑵ䘉ṧⲴമ‫ۿ‬䇝㓳ਾˈ䘉њ㖁㔌ሩҾ⤜・Ⲵ傼䇱䳶ਸ䗮ࡠҶ 90%
Ⲵ㋮ᓖǄമѝҏᱮ⽪Ҷ֯⭘䇝㓳ṧֻ䘝ԓ 1 ⅑ਾ઼䘝ԓ 100 ⅑ਾⲴ㖁㔌ᵳ٬Ǆ⇿њ䗃ࠪঅ‫˄ݳ‬ᐖˈ
ࡽˈਣˈк˅ᴹഋњᵳ٬ˈ⭘᳇˄䍏˅઼᰾˄↓˅Ⲵᯩඇᱮ⽪Ǆᴰᐖ‫Ⲵח‬ᯩඇሩᓄᵳw0ˈᆳߣᇊঅ
‫Ⲵݳ‬䰸٬ˈਣ䶒ⲴйњᯩඇሩᓄӾйњ䳀㯿অ‫ݳ‬䗃‫Ⲵޕ‬ᵳǄമѝҏᱮ⽪Ҷ⇿њ‫ۿ‬㍐䗃‫⇿ࡠޕ‬њ䳀㯿
অ‫Ⲵݳ‬ᵳ٬ˈ㻛⭫൘ሩᓄ‫ۿ‬㍐Ⲵս㖞Ǆ
4.7.2 䇴䇗㾷㍖
ᓄ⭘৽ੁՐ᫝㇇⌅ࡠањ㔉ᇊԫ࣑ᰦˈᗵ享ߣᇊࠐњ䇮䇑㾱㍐Ǆл䶒ᡁԜᖂ㓣ࠪҶᆖҐ
Ӫ㝨ᵍੁ䘉њᆖҐԫ࣑ⲴаӋ䇮䇑㾱㍐Ǆቭ㇑ᡁԜ⋑ᴹᢃ㇇৫䘹ᤙ㋮⺞ⲴᴰՈ䇮䇑ˈն䘉䟼
᧿䘠Ⲵ䇮䇑ሩⴞḷ࠭ᮠᆖҐᗇ⴨ᖃྭǄ൘䇝㓳Ҷ 260 ᑵമ‫ۿ‬ṧֻѻਾˈሩҾ⤜・⍻䈅䳶ਸⲴ
㋮ᓖ䗮ࡠ 90%Ǆ⴨ሩ㘼䀰ˈྲ᷌䲿ᵪ⥌⍻ഋњ㝨ᵍੁѝⲴањˈਚ㜭䗮ࡠ 25%Ⲵ↓⺞⦷Ǆ
䗃‫ޕ‬㕆⸱Ǆᐢ㓿⸕䚃 ANN Ⲵ䗃‫ޕ‬ᗵ❦ᱟമ‫Ⲵۿ‬Ḁ⿽㺘⽪ˈ䛓Ѹ䇮䇑Ⲵ‫ޣ‬䭞ᱟྲօ㕆

⸱䘉ᑵമ‫ۿ‬ǄֻྲᡁԜਟԕሩമ‫ۿ‬䘋㹼亴༴⨶ˈᶕ࠶䀓ࠪ䗩㕈ǃӞᓖа㠤Ⲵ४ฏᡆަԆተ䜘
മ‫ۿ‬⢩ᖱˈ❦ਾᢺ䘉Ӌ⢩ᖱ䗃‫ޕ‬㖁㔌Ǆ䘉⿽䇮䇑Ⲵањ䰞仈ᱟՊሬ㠤⇿ᑵമ‫ۿ‬ᴹн਼ᮠ䟿Ⲵ
⢩ᖱ৲ᮠ˄ֻྲ䗩㕈Ⲵᮠ䟿˅ˈ❦㘼 ANN ާᴹപᇊᮠ䟿Ⲵ䗃‫ޕ‬অ‫ݳ‬ǄሩҾ䘉⿽ᛵߥˈᡁԜ
Ⲵ䇮䇑ᱟᢺമ‫ۿ‬㕆⸱ᡀപᇊⲴ 30u32 ‫ۿ‬㍐ⲴӞᓖ٬ˈ⇿њ‫ۿ‬㍐ሩᓄањ㖁㔌䗃‫ޕ‬Ǆᒦфᢺ㤳
തᱟ 0 ࡠ 255 ⲴӞᓖ٬᤹∄ֻ㓯ᙗ㕙᭮ࡠ 0 ࡠ 1 Ⲵ४䰤޵ˈԕ֯㖁㔌䗃‫ޕ‬о䳀অ‫઼ݳ‬䗃ࠪঅ
‫ݳ‬൘਼ṧⲴ४䰤ਆ٬Ǆᇎ䱵к䘉䟼Ⲵ 30u32 ‫ۿ‬㍐മ‫ۿ‬ቡᱟ৏ᶕ 120u128 ‫ۿ‬㍐Ⲵമ‫Ⲵۿ‬վ࠶
䗘⦷ᾲᤜˈ⇿њվ࠶䗘⦷‫ۿ‬㍐ṩᦞሩᓄⲴ㤕ᒢ儈࠶䗘⦷‫ۿ‬㍐ӞᓖⲴ൷٬䇑㇇ᗇࡠǄ֯⭘䘉ṧ
Ⲵվ࠶䗘⦷മ‫ˈۿ‬ᢺ䗃‫ޕ‬њᮠ઼ᵳ٬Ⲵᮠ䟿߿ቁࡠҶањᴤ᱃Ҿ༴⨶Ⲵ㿴⁑ˈӾ㘼䱽վҶ䘀
㇇㾱≲ˈն਼ᰦҏ‫⮉؍‬Ҷ䏣ཏⲴ࠶䗘⦷ԕ↓⺞࠶㊫മ‫ۿ‬Ǆഎᗶമ 4-1 ѝ ALVINN ㌫㔏֯⭘
Ҷ⴨լⲴⲴ࠶䗘⦷മ‫֌ۿ‬Ѫ㖁㔌Ⲵ䗃‫ޕ‬Ǆањᴹ䏓Ⲵᐞ࡛ᱟˈ൘ ALVINN ѝˈ⇿ањվ࠶
䗘⦷‫ۿ‬㍐ⲴӞᓖㅹҾӾ儈࠶䗘⦷മ‫ۿ‬ሩᓄⲴ४ฏѝ䲿ᵪਆањ‫ۿ‬㍐ⲴӞᓖˈ㘼нᱟਆ䘉њ४
ฏѝᡰᴹ‫ۿ‬㍐ӞᓖⲴ൷٬ǄަࣘᵪᱟѪҶ᰾ᱮൠ߿ቁӾ儈࠶䗘⦷മ‫ۿ‬ӗ⭏վ࠶䗘⦷മ‫ۿ‬ᡰ䴰
Ⲵ䘀㇇Ǆ䘉њ᭸⦷ሩҾ ALVINN ㌫㔏ᱟ⢩࡛䟽㾱ⲴˈഐѪ൘㠚ࣘ傮傦䖖䖶Ⲵ䗷〻ѝˈALVINN
㌫㔏Ⲵ㖁㔌ᗵ享൘⇿。䫏༴⨶ᖸཊᑵമ‫ۿ‬Ǆ
䗃ࠪ㕆⸱ǄANN ᗵ享䗃ࠪഋњ٬ѝⲴањᶕ㺘⽪䗃‫ޕ‬മ‫ۿ‬ѝӪ㝨Ⲵᵍੁ˄ᐖˈਣˈкˈ
ࡽ˅Ǆ⌘᜿ᡁԜਟԕ֯⭘অаⲴ䗃ࠪঅ‫ݳ‬ᶕ㕆⸱䘉ഋ⿽ᛵߥⲴ࠶㊫ˈֻྲᤷᇊ䗃ࠪ٬ 0.2ˈ0.4ˈ
0.6 ઼ 0.8 ᶕ㕆⸱䘉ഋњਟ㜭٬Ǆн䗷䘉䟼ᡁԜ֯⭘ 4 њн਼Ⲵ䗃ࠪঅ‫⇿ˈݳ‬ањሩᓄഋ⿽
ਟ㜭ᵍੁѝⲴа⿽ˈਆާᴹᴰ儈٬Ⲵ䗃ࠪ֌Ѫ㖁㔌Ⲵ亴⍻٬Ǆ䘉⿽ᯩ⌅㓿ᑨ㻛〠Ѫ n ਆ 1
˄1-of-n˅䗃ࠪ㕆⸱Ǆ䘹ᤙ n ਆ 1 䗃ࠪ㕆⸱㘼н⭘অњঅ‫ݳ‬ᴹєњࣘᵪǄㅜаˈ䘉Ѫ㖁㔌㺘
⽪ⴞḷ࠭ᮠᨀ‫׋‬ҶᴤབྷⲴ㠚⭡ᓖ˄ণ൘䗃ࠪቲঅ‫ݳ‬ѝᴹ n ‫Ⲵؽ‬ਟ⭘ᵳ٬˅ǄㅜҼˈ൘ n ਆ 1
㕆⸱ѝˈᴰ儈٬䗃઼ࠪ⅑儈٬䗃ࠪ䰤Ⲵᐞᔲਟԕ֌Ѫሩ㖁㔌亴⍻Ⲵ㖞ؑᓖ˄н᰾⺞Ⲵ࠶㊫ਟ
㜭ሬ㠤㔃᷌⴨䘁ᡆ⴨ㅹ˅ Ǆ䘋а↕Ⲵ䇮䇑䰞仈ᱟĀ䘉 4 њ䗃ࠪঅ‫ⴞⲴݳ‬ḷ٬ᓄ䈕ᱟӰѸ˛ā
ањᱮ㘼᱃㿱Ⲵ࣎⌅ᱟ⭘ 4 њⴞḷ٬<1ˈ0ˈ0ˈ0>ᶕ㕆⸱㝨ᵍੁᐖˈ<0ˈ1ˈ0ˈ0>ᶕ㕆⸱
㝨ᵍੁ↓ࡽˈ‫↔׍‬㊫᧘ǄᡁԜ䘉䟼֯⭘ 0.1 ઼ 0.9ˈ㘼нᱟ 0 ઼ 1ˈণ<0.9ˈ0.1ˈ0.1ˈ0.1>
㺘⽪㝨ᵍੁᐖⲴⴞḷ䗃ࠪੁ䟿Ǆ䚯‫ ⭘֯ݽ‬0 ઼ 1 ֌Ѫⴞḷ٬Ⲵ৏ഐᱟ sigmoid অ‫ݳ‬ሩҾᴹ䲀
ᵳ٬н㜭ӗ⭏䘉ṧⲴ䗃ࠪǄྲ᷌ᡁԜԱമ䇝㓳㖁㔌ᶕ߶⺞३䝽ⴞḷ٬ 0 ઼ 1ˈởᓖл䱽ሶՊ
䘛֯ᵳ٬ᰐ⭼໎䮯Ǆਖаᯩ䶒ˈ٬ 0.1 ઼ 0.9 ᱟ sigmoid অ‫ݳ‬൘ᴹ䲀ᵳ٬ᛵߥлਟԕᆼᡀⲴǄ
㖁㔌㔃ᶴമǄ↓ྲࡽ䶒ᡰ᧿䘠Ⲵˈ৽ੁՐ᫝㇇⌅ਟԕ㻛ᓄ⭘ࡠԫօᴹੁᰐ⧟ sigmoid অ
‫Ⲵݳ‬㖁㔌ǄᡰԕˈᡁԜ䶒ѤⲴਖа䇮䇑䰞仈ᱟˈ䘉њ㖁㔌वਜ਼ཊቁњঅ‫ݳ‬ԕ৺ྲօӂ䘎Ǆᴰ
Პ䙽Ⲵа⿽㖁㔌㔃ᶴᱟ࠶ቲ㖁㔌ˈаቲⲴ⇿њঅ‫ࡽੁݳ‬䘎᧕ࡠлаቲⲴ⇿ањঅ‫ݳ‬ǄⴞࡽⲴ
䇮䇑䘹ᤙ䘉ṧⲴḷ߶㔃ᶴˈ֯⭘єቲ sigmoid অ‫˄ݳ‬ањ䳀㯿ቲ઼ањ䗃ࠪቲ˅Ǆ֯⭘аᡆ
єቲ sigmoid অ‫ݳ‬ᱟᖸᲞ䙽Ⲵˈ‫ڦ‬ቄ֯⭘йቲǄ֯⭘ᴤཊⲴቲᱟнᑨ㿱ⲴˈഐѪ䇝㓳ᰦ䰤Պ
ਈᗇᖸ䮯ˈ㘼фйቲ sigmoid অ‫Ⲵݳ‬㖁㔌ᐢ㓿㜭ཏ㺘⽪ᮠ䟿⴨ᖃབྷⲴⴞḷ࠭ᮠ˄㿱 4.6.2 㢲˅
Ǆ
ᡁԜᐢ㓿⺞ᇊ䘹ᤙањ࠶ቲⲴࡽ侸㖁㔌ˈ䛓Ѹަѝᓄ䈕वਜ਼ཊቁњ䳀㯿অ‫ݳ‬઒˛൘മ 4-10
ᣕ੺Ⲵ㔃᷌ѝˈӵ֯⭘Ҷйњ䳀㯿অ‫ˈݳ‬䗮ࡠҶሩ⍻䈅䳶ਸ 90%Ⲵ㋮ᓖǄ൘ਖањ֯⭘ 30
њ䳀㯿অ‫Ⲵݳ‬ᇎ傼ѝˈᗇࡠⲴ㋮ᓖᨀ儈ҶаࡠєњⲮ࠶⛩Ǆቭ㇑䘉єњᇎ傼ᗇࡠⲴ⌋ॆ㋮ᓖ
⴨ᐞᖸሿˈնਾањ䈅傼᰾ᱮ䴰㾱ᴤཊⲴ䇝㓳ᰦ䰤Ǆ֯⭘ 260 ᑵമ‫Ⲵۿ‬䇝㓳ṧֻˈ30 њ䳀
অ‫Ⲵݳ‬㖁㔌൘ Sun Sparc5 ᐕ֌ㄉкⲴ䇝㓳ᰦ䰤བྷ㓖ᱟањሿᰦǄ⴨ሩ㘼䀰ˈйњ䳀㯿অ‫ݳ‬
Ⲵ㖁㔌བྷ㓖ᱟ 5 ࠶䫏ǄӪԜᐢ㓿ਁ⧠൘ᖸཊᓄ⭘ѝ䴰㾱Ḁњᴰሿᮠ䟿Ⲵ䳀অ‫ݳ‬ᶕ㋮⺞ൠᆖҐ
ⴞḷ࠭ᮠˈᒦф䎵䗷䘉њᮠ䟿Ⲵཊ։Ⲵ䳀অ‫ݳ‬нՊᱮ㪇ൠᨀ儈⌋ॆ㋮ᓖˈᶑԦᱟ֯⭘Ӕ৹傼
䇱ᯩ⌅ᶕߣᇊᓄ䈕䘋㹼ཊቁ⅑ởᓖл䱽䘝ԓǄྲ᷌⋑ᴹ֯⭘Ӕ৹傼䇱ˈ䛓Ѹ໎࣐䳀㯿অ‫ݳ‬ᮠ
䟿㓿ᑨՊ໎࣐䗷ᓖᤏਸ䇝㓳ᮠᦞⲴ‫ˈੁٮ‬Ӿ㘼䱽վ⌋ॆ㋮ᓖǄ
ᆖҐ㇇⌅ⲴަԆ৲ᮠǄ൘䘉њᇎ傼ѝˈᆖҐ䙏⦷K㻛䇮ᇊѪ 0.3ˈߢ䟿D㻛䇮ᇊѪ 0.3Ǆ

䍻Ҹ䘉єњ৲ᮠᴤվⲴ٬Պӗ⭏བྷփ⴨ᖃⲴ⌋ॆ㋮ᓖˈն䴰㾱ᴤ䮯Ⲵ䇝㓳ᰦ䰤Ǆྲ᷌䘉єњ
٬㻛䇮ᇊᗇཚ儈ˈ䇝㓳ሶн㜭᭦ᮋࡠањާᴹਟ᧕ਇ䈟ᐞ˄൘䇝㓳䳶ਸк˅Ⲵ㖁㔌Ǆ൘ᮤњ
䈅傼ѝᡁԜ֯⭘ᆼ‫Ⲵޘ‬ởᓖл䱽˄઼㺘 4-2 ㇇⌅ѝ䲿ᵪ䘁լⲴởᓖл䱽н਼˅Ǆ䗃ࠪঅ‫Ⲵݳ‬
㖁㔌ᵳ٬㻛ࡍ࿻ॆѪሿⲴ䲿ᵪ٬Ǆ❦㘼䗃‫ޕ‬অ‫Ⲵݳ‬ᵳ٬㻛ࡍ࿻ॆѪ 0ˈഐѪ䘉ṧਟԕ֯ᆖҐ
ࡠⲴᵳ٬Ⲵമ‫˄ॆۿ‬㿱മ 4-10˅ᴤ᱃Ҿ⨶䀓ˈ㘼ሩ⌋ॆ㋮ᓖ⋑ᴹ᰾ᱮⲴᖡ૽Ǆ䇝㓳Ⲵ䘝ԓ
⅑ᮠⲴ䘹ᤙਟԕ䙊䗷࠶ࢢਟ⭘ⲴᮠᦞѪ䇝㓳䳶ਸ઼⤜・Ⲵ傼䇱䳶ਸǄởᓖл䱽ᯩ⌅㻛⭘Ҿᴰ
ሿॆ䇝㓳䳶ਸкⲴ䈟ᐞˈᒦф⇿䳄 50 ⅑ởᓖл䱽䘝ԓṩᦞ傼䇱䳶ਸ䇴ՠа⅑㖁㔌Ⲵᙗ㜭Ǆ
ᴰ㓸䘹ᤙⲴ㖁㔌ᱟሩ傼䇱䳶ਸ㋮ᓖᴰ儈Ⲵ㖁㔌Ǆਟԕ৲㿱 4.6.5 㢲ᗇࡠ‫ޣ‬Ҿ䘉њ䗷〻Ⲵ䀓䟺
઼‫ᦞ׍‬Ǆᴰ㓸ᣕ੺Ⲵ㋮ᓖ˄ҏቡᱟ 90%ˈሩҾമ 4-10 ѝⲴ㖁㔌˅ᱟ൘⋑ᴹሩ䇝㓳ӗ⭏ԫօ
ᖡ૽Ⲵㅜйњ䳶ਸüü⍻䈅䳶ਸк⍻䟿ᗇࡠⲴǄ
4.7.3 ᆜҖࡦⲺ䳆㰅ቸ㺞⽰
ᴹᗵ㾱࠶᷀ал㖁㔌ѝᆖҐᗇࡠⲴ 2899 њķᵳ٬Ǆമ 4-10 ᧿㔈Ҷሩᡰᴹ䇝㓳ṧֻ䘋㹼

а⅑ᵳ٬ᴤᯠਾⲴ⇿њᵳ٬ˈ઼ 100 ⅑ᴤᯠਾⲴᵳ٬Ǆ
ѪҶ⨶䀓䘉Ӌമ‫ݸˈۿ‬㘳㲁മѝ㍗䛫Ӫ㝨മ‫ۿ‬лⲴഋњ⸙ᖒǄ⇿ањ⸙ᖒ᧿㔈Ҷ㖁㔌ѝ
ഋњ䗃ࠪঅ‫˄ݳ‬㕆⸱Ҷᐖǃࡽǃਣ઼к˅ѝⲴањᵳ٬Ǆ⇿њ⸙ᖒѝⲴഋњሿᯩᖒ㺘⽪઼䘉
ķ
䈁⌘˖2899=䗃‫ޕ‬অ‫ݳ‬ойњ䳀অ‫ݳ‬䰤䘎᧕ሩᓄⲴᵳ˄960u3˅+йњ䳀অ‫ݳ‬оഋњ䗃ࠪঅ‫ݳ‬䰤䘎᧕ሩᓄ
Ⲵᵳ˄3u4˅+йњ䳀অ‫઼ݳ‬ഋњ䗃ࠪঅ‫Ⲵݳ‬w0ᵳ˄3+4˅
њ䗃ࠪঅ‫ޣݳ‬㚄Ⲵഋњᵳ٬üüᴰᐖ䗩ᱟᵳw0ˈᆳߣᇊঅ‫Ⲵݳ‬䰸٬ˈ❦ਾᱟ䘎᧕йњ䳀㯿অ
‫ࡠݳ‬䘉њ䗃ࠪⲴйњᵳ٬ǄᯩᖒⲴӞᓖ㺘⽪ᵳ٬ˈӞⲭ㺘⽪䖳བྷⲴ↓ᵳ٬ˈ᳇唁㺘⽪䖳བྷⲴ
䍏ᵳ٬ˈӻҾѝ䰤Ⲵ⚠㢢䱤ᖡ㺘⽪ѝㅹⲴᵳ٬ǄֻྲˈḷѪĀкāⲴ䗃ࠪঅ‫Ⲵݳ‬䰸٬ᵳw0᧕
䘁 0ˈӾㅜањ䳀㯿অ‫ݳ‬ᶕⲴᵳ٬Ѫ䖳བྷⲴ↓٬ˈӾㅜҼњ䳀㯿অ‫ݳ‬ᶕⲴᵳ٬Ѫ䖳བྷⲴ䍏٬Ǆ
䳀㯿অ‫Ⲵݳ‬ᵳ٬ᱮ⽪൘䗃ࠪঅ‫Ⲵݳ‬л䗩Ǆഎᗶалˈ⇿њ䳀㯿অ‫᧕ݳ‬ਇᡰᴹ 30u32 њ‫ۿ‬

㍐䗃‫ޕ‬Ǆо䘉Ӌ䗃‫ޣޕ‬㚄Ⲵ 30u32 њᵳ٬㻛ᱮ⽪൘ᆳԜሩᓄⲴ‫ۿ‬㍐Ⲵս㖞˄䰸٬ᵳw0㻛䟽ਐ
ᱮ⽪൘മ‫ۿ‬䱥ࡇⲴᐖк䀂˅Ǆ䶎ᑨᴹ䏓ⲴᱟˈਟԕⴻࡠᵳⲴਆ٬䙊ᑨሩӪ㝨઼䓛փࠪ⧠Ⲵമ
‫ۿ‬४ฏⲴ⢩࡛᭿ᝏǄ
䪸ሩ⇿ањ䇝㓳ṧֻởᓖл䱽䘝ԓ 100 ⅑ਾⲴ㖁㔌ᵳ٬ᱮ⽪൘മⲴл䜘Ǆ⌘᜿ᴰᐖ䗩Ⲵ

䳀㯿অ‫Ⲵݳ‬ᵳ٬઼䘝ԓа⅑ᰦⲴᵳ٬ᴹᖸབྷн਼ˈਖєњ䳀㯿অ‫Ⲵݳ‬ᵳ٬ҏᴹᡰਈॆǄ⧠൘
ਟԕ࠶᷀ал䘉њᴰ㓸ᵳ٬䳶ਸѝⲴ㕆⸱Ǆֻྲˈ㘳㲁䗃ࠪঅ‫ݳ‬ᤷࠪањӪᱟ൘ੁਣⴻǄ䘉
њঅ‫ݳ‬оㅜҼњ䳀㯿অ‫ݳ‬䰤ާᴹањ䖳བྷⲴ↓ᵳ٬ˈоㅜйњ䳀অ‫ݳ‬䰤ާᴹањབྷⲴ䍏ᵳ
٬Ǆ࠶᷀䘉єњ䳀অ‫Ⲵݳ‬ᵳ٬ˈᇩ᱃ⴻࡠྲ᷌ањӪⲴ㝨ᱟ䖜ੁԆⲴਣ䶒˄ҏቡᱟᡁԜⲴᐖ
䶒˅ˈ䛓ѸԆⲴӞᓖ儈ⲴⳞ㛔Պབྷ㠤о䘉њ䳀㯿অ‫ݳ‬ѝⲴ䖳བྷ↓٬ሩ喀ˈ਼ᰦԆⲴӞᓖվⲴ
ཤਁՊབྷ㠤о䍏ᵳ٬ሩ喀ˈ䘉ሬ㠤↔অ‫ݳ‬䗃ࠪањ䖳བྷⲴ٬Ǆ਼ṧⲴമ‫ۿ‬Պ֯ㅜйњ䳀অ‫ݳ‬
䗃ࠪањ᧕䘁 0 Ⲵ٬ˈഐѪӞᓖ儈Ⲵ㝨䜘‫ੁٮ‬ҾоབྷⲴ䍏ᵳሩ喀Ǆ
4.8 Ӱᐛ⾔㔅㖇㔒Ⲻ儎㓝䈓从
4.8.1 ެԌਥ䘿Ⲻ䈥ᐤ࠳ᮦ
↓ྲࡽ䶒ᡰᤷࠪⲴˈਚ㾱࠭ᮠ E ⴨ሩ৲ᮠॆⲴ‫ٷ‬䇮オ䰤ਟᗞˈ䛓Ѹቡਟԕᢗ㹼ởᓖл
䱽Ǆ㲭❦สᵜⲴ৽ੁՐ᫝㇇⌅ԕ㖁㔌䈟ᐞᒣᯩ઼Ⲵᖒᔿᇊѹ EˈնҏᴹӪᨀࠪަԆⲴᇊѹˈ
ԕ‫ׯ‬ᢺަԆⲴ㓖ᶏᕅ‫ޕ‬ᵳ٬䈳ᮤ⌅ࡉǄྲ᷌ᇊѹҶањᯠⲴ Eˈ䛓Ѹቡᗵ享᧘ሬࠪањᯠⲴ
ᵳ٬䈳ᮤ⌅ࡉ‫׋‬ởᓖл䱽֯⭘ǄE ⲴަԆਟ䘹ᇊѹवᤜ˖
x Ѫᵳ٬໎࣐ањ᜙㖊亩Ǆྲ਼ࡽ䶒䇘䇪ⲴˈᡁԜਟԕ࣐‫ޕ‬ањ䲿⵰ੁ䟿ᑵᓖ໎
䮯Ⲵ亩ࡠ E ѝǄ䘉ሬ㠤ởᓖл䱽ᩌራ䖳ሿⲴᵳ٬ੁ䟿ˈӾ㘼߿ሿ䗷ᓖᤏਸⲴ仾
䲙Ǆа⿽࣎⌅ᱟ᤹➗л䶒Ⲵㅹᔿ䟽ᯠᇊѹ E˖
& 1
E ( w) { ¦ ¦ (t kd o kd ) 2 J ¦ w ji
2
2 dD koutputs i, j
䘉ᗇࡠҶањо৽ੁՐ᫝⌅ࡉสᵜа㠤Ⲵᵳᴤᯠ⌅ࡉˈਚᱟ൘⇿⅑䘝ԓᰦѪ⇿
њᵳ҈ԕᑨ䟿˄1-2JK˅Ǆഐ↔䘹ᤙ䘉⿽ E Ⲵᇊѹ઼֯⭘ᵳ㺠߿ㆆ⮕˄㿱㓳Ґ 4.10˅
ᱟㅹԧⲴǄ
x ሩ䈟ᐞ໎࣐а亩ⴞḷ࠭ᮠⲴᯌ⦷˄slope˅ᡆሬᮠǄḀӋᛵߥлˈ䇝㓳ؑ᚟ѝн
ӵᴹⴞḷ٬ˈ㘼ф䘈ᴹ‫ޣ‬Ҿⴞḷ࠭ᮠⲴሬᮠǄֻྲˈSimard et al.˄1992˅᧿䘠
Ҷањᆇㅖ䇶࡛Ⲵᓄ⭘ˈ൘䘉њᓄ⭘ѝ֯⭘ҶаӋ䇝㓳ሬᮠᶕᕪ䘛㖁㔌ᆖҐ䛓
Ӌ൘മ‫ۿ‬ᒣ〫ѝнਈⲴᆇㅖ䇶࡛࠭ᮠǄMitchell and Thrun˄1993˅᧿䘠Ҷṩᦞᆖ
ҐಘԕࡽⲴ⸕䇶䇑㇇䇝㓳ሬᮠⲴᯩ⌅Ǆ൘䘉єњ㌫㔏ѝ˄൘ㅜ 12 ㄐѝ᧿䘠˅ˈ
䈟ᐞ࠭ᮠ䜭㻛໎࣐Ҷа亩ˈ⭘ᶕ㺑䟿䘉Ӌ䇝㓳ሬᮠ઼㖁㔌Ⲵᇎ䱵ሬᮠ䰤ⲴᐞᔲǄ
䘉ṧⲴ䈟ᐞ࠭ᮠⲴањֻᆀᱟ
& 1 ª § wt wo ·
2
º
E ( w) { ¦ ¦ «(t kd okd ) 2 P ¦ ¨¨ kdj kdj ¸¸ »
2 dD koutputs « jinputs © wxd wxd ¹ »¼
¬
wt kd
wx j
j
x d
䘉䟼ˈ 㺘⽪ሩҾ䇝㓳ᇎֻdㅜjњ䗃‫ޕ‬অ‫Ⲵݳ‬٬ǄҾᱟ d ᱟ᧿䘠ⴞḷ䗃ࠪ٬
wokd
wx dj 㺘⽪ᇎ䱵ⲴᆖҐ㖁㔌
j
x d
tkdᓄ䈕ྲօ䲿䗃‫ޕ‬٬ ਈॆⲴ䇝㓳ሬᮠǄ㊫լⲴˈ
ⲴሩᓄሬᮠǄᑨᮠPߣᇊ३䝽䇝㓳٬ሩҾ३䝽䇝㓳ሬᮠⲴ⴨ሩᵳ٬Ǆ
x ֯㖁㔌ሩⴞḷ٬ⲴӔ৹⟥˄cross entropy˅ᴰሿॆǄ㘳㲁ᆖҐањᾲ⦷࠭ᮠˈ∄
ྲ亴⍻ањُ䍧⭣䈧㘵Պ੖䘈䍧ˈṩᦞᱟ䘉њ⭣䈧㘵Ⲵᒤ喴઼ᆈⅮ։仍Ǆቭ㇑
䘉䟼Ⲵ䇝㓳ṧֻӵᨀ‫׋‬ҶᐳቄරⲴⴞḷ٬˄㾱Ѹᱟ 1ˈ㾱Ѹᱟ 0ˈṩᦞ䘉њ⭣䈧
㘵ᱟ੖䘈䍧˅ˈնสᵜⲴⴞḷ࠭ᮠᴰྭԕ⭣䈧㘵䘈䍧Ⲵᾲ⦷Ⲵᖒᔿ䗃ࠪˈ㘼н
ᱟሩ⇿њ䗃‫ޕ‬ᇎֻ䜭Աമ䗃ࠪ᰾⺞Ⲵ 0 ᡆ 1 ٬Ǆ൘䘉⿽ᛵߥлˈᡁԜᐼᵋ㖁㔌
䗃ࠪањᾲ⦷ՠ䇑ˈਟԕ䇱᰾ᴰሿॆӔ৹⟥˄cross entropy˅Ⲵ㖁㔌ਟԕ㔉ࠪᴰ
ྭⲴ˄ҏቡᱟᴰབྷլ❦˅ᾲ⦷ՠ䇑ˈӔ৹⟥Ⲵᇊѹྲл˖
¦ t d log od (1 t d ) log(1 od )
d D
䘉䟼odᱟ㖁㔌ሩҾ䇝㓳ṧֻd䗃ࠪⲴᾲ⦷ՠ䇑ˈtdᱟሩҾ䇝㓳ṧֻdⲴⴞḷ٬˄0
ᡆ 1˅Ǆㅜ 6 ㄐ䇘䇪Ҷօᰦ৺ѪӰѸᴰਟ㜭Ⲵ㖁㔌‫ٷ‬䇮ቡᱟ֯Ӕ৹⟥ᴰሿॆⲴ‫ٷ‬
䇮ˈᒦ᧘ሬҶ⴨ᓄⲴsigmoidঅ‫Ⲵݳ‬ởᓖл䱽ᵳ٬䈳ᮤ⌅ࡉǄㅜ 6 ㄐҏ᧿䘠Ҷ൘
ӰѸᶑԦлᴰਟ㜭Ⲵ‫ٷ‬䇮ቡᱟ֯䈟ᐞᒣᯩ઼ᴰሿॆⲴ‫ٷ‬䇮Ǆ
x ᭩ਈᴹ᭸䈟ᐞ࠭ᮠҏਟԕ䙊䗷ᵳ٬‫ޡ‬ӛ˄weight sharing˅ᆼᡀˈҏቡᱟᢺон
਼অ‫ݳ‬ᡆ䗃‫ޣ⴨ޕ‬㚄ⲴᵳĀ᥶㔁൘а䎧āǄ䘉䟼Ⲵᜣ⌅ᱟᕪ䘛н਼Ⲵ㖁㔌ᵳ٬
ਆа㠤Ⲵ٬ˈ䙊ᑨᱟѪҶᇎᯭӪ㊫䇮䇑㘵һ‫⸕ݸ‬䚃ⲴḀњ㓖ᶏǄֻྲˈWaibel et
al.˄1989˅઼ Lang et al.˄1990˅᧿䘠Ҷ⾎㓿㖁㔌൘䈝丣䇶࡛ᯩ䶒Ⲵањᓄ⭘ ˈ
ަѝ㖁㔌Ⲵ䗃‫ޕ‬ᱟ൘ањ 144 ∛。Ⲵᰦ䰤デѝн਼ᰦ䰤Ⲵ䈝丣仁⦷࠶䟿Ǆ൘䘉
њᓄ⭘ѝਟԕ‫Ⲵڊ‬ањ‫ٷ‬ᇊᱟ˖ањ⢩ᇊ䈝丣˄ֻྲĀeeeā˅Ⲵ仁⦷࠶䟿Ⲵ䇶
࡛ᱟо䘉њ䈝丣൘ 144 ∛。ᰦ䰤デѝࠪ⧠Ⲵ⺞࠷ᰦ䰤ᰐ‫Ⲵޣ‬ǄѪҶᇎᯭ䘉њ㓖
ᶏˈᗵ享ᕪ䘛᧕᭦䘉њᰦ䰤デн਼䜘࠶Ⲵн਼অ‫ޡݳ‬ӛᵳ٬Ǆ䘉ṧ‫Ⲵڊ‬᭸᷌ᱟ
㓖ᶏҶ‫ٷ‬䇮Ⲵ▌൘オ䰤ˈӾ㘼߿ሿҶ䗷ᓖᤏਸⲴ仾䲙ˈᨀ儈Ҷ߶⺞⌋ॆࡠᵚ㿱
䗷ᛵᖒⲴਟ㜭ᙗǄᵳ٬‫ޡ‬ӛ䙊ᑨ䘉ṧᇎ⧠˖俆‫ݸ‬൘‫ޡ‬ӛᵳ٬Ⲵ⇿њঅ‫࡛࠶ݳ‬ᴤ
ᯠ਴њᵳ٬ˈ❦ਾਆ䘉Ӌᵳ٬Ⲵᒣ൷ˈ޽⭘䘉њᒣ൷٬ᴯᦒ⇿њ䴰‫ޡ‬ӛⲴᵳ٬Ǆ
䘉њ䗷〻Ⲵ㔃᷌ᱟ㻛‫ޡ‬ӛⲴᵳ٬о⋑ᴹ㻛‫ޡ‬ӛⲴᵳ٬⴨∄֯⭘Ҷн਼Ⲵ䈟ᐞ࠭
ᮠǄ
4.8.2 ެԌਥ䘿Ⲻ䈥ᐤᴶቅौ䗽ぁ
㲭❦ởᓖл䱽ᱟᩌራ֯䈟ᐞ࠭ᮠᴰሿॆⲴ‫ٷ‬䇮Ⲵᴰ䙊⭘Ⲵᩌ㍒ᯩ⌅ѻаˈնᆳнᙫᱟᴰ
儈᭸ⲴǄᖃ䇝㓳༽ᵲⲴ㖁㔌ᰦˈн䳮㿱ࡠ৽ੁՐ᫝㇇⌅㾱䘋㹼кз⅑Ⲵᵳ٬ᴤᯠ䘝ԓǄ⭡Ҿ
䘉њ৏ഐˈӪԜ᧒㍒ᒦᨀࠪҶᖸཊަԆⲴᵳ٬Ոॆ㇇⌅ǄѪҶ亶ՊަԆⲴਟ㜭ᯩ⌅ˈᡁԜн
࿘ᢺᵳ٬ᴤᯠᯩ⌅ቡⴻ֌ᱟ㾱ߣᇊєњ䰞仈˖䘹ᤙањ᭩ਈᖃࡽᵳ٬ੁ䟿Ⲵᯩੁ˗䘹ᤙ㾱〫
ࣘⲴ䐍⿫Ǆ൘৽ੁՐ᫝㇇⌅ѝˈ䘉њᯩੁᱟ䙊䗷ਆởᓖⲴ䍏٬ᶕ䘹ᤙⲴˈ䐍⿫ᱟ䙊䗷ᑨ䟿Ⲵ
ᆖҐ䙏⦷KߣᇊⲴǄ
а⿽㻛〠ѪĀ㓯㓯ᩌ㍒˄line search˅āⲴՈॆᯩ⌅ˈ䟷⭘Ҷн਼Ⲵᯩ⌅䘹ᤙᵳ٬ᴤᯠⲴ䐍
⿫Ǆ⺞࠷ൠ䇢ˈ⇿ᖃ䘹ᇊҶаᶑ⺞ᇊᵳ٬ᴤᯠᯩੁⲴ䐟㓯ˈ䛓ѸᵳᴤᯠⲴ䐍⿫ᱟ䙊䗷ራ᢮⋯
䘉ᶑ㓯Ⲵ䈟ᐞ࠭ᮠⲴᴰሿ٬ᶕ䘹ᤙⲴǄ⌘᜿䘉ਟ㜭ሬ㠤ᖸབྷᑵᓖҏਟ㜭ᱟᖸሿᑵᓖⲴᵳ٬ᴤ
ᯠˈ㾱ⴻ⋯䘉ᶑ㓯Ⲵᴰሿ䈟ᐞ⛩Ⲵս㖞Ǆਖа⿽ᯩ⌅ˈᱟṩᦞĀ㓯ᩌ㍒āⲴᙍᜣᔪ・Ⲵˈ㻛
〠Ѫ‫ޡ‬
‫ޡ‬䖝ởᓖ˄conjugate gradient˅⌅Ǆ䘉⿽ᯩ⌅䘋㹼а㌫ࡇ㓯ᩌ㍒ᶕᩌ㍒䈟ᐞᴢ䶒Ⲵᴰሿ
٬Ǆ䘉а㌫ࡇᩌ㍒Ⲵㅜа↕ӽ❦֯⭘ởᓖⲴ৽ᯩੁ֌ѪᯩੁǄ൘ਾᶕⲴ⇿а↕ˈ䘹ᤙ֯䈟ᐞ
ởᓖ࠶䟿ࡊྭѪ 0 ᒦ‫؍‬ᤱѪ 0 ⲴᯩੁǄ
㲭❦ަԆⲴ䈟ᐞᴰሿॆᯩ⌅ᨀ儈Ҷ䇝㓳㖁㔌Ⲵ᭸⦷ˈն䊑‫ޡ‬䖝ởᓖ䘉ṧⲴᯩ⌅ሩҾᴰ㓸
㖁㔌Ⲵ⌋ॆ䈟ᐞ⋑ᴹ᰾ᱮⲴᖡ૽Ǆሩᴰ㓸䈟ᐞᜏаਟ㜭Ⲵᖡ૽ᱟˈн਼Ⲵ䈟ᐞᴰሿॆ䗷〻Պ
䲧‫ޕ‬н਼Ⲵተ䜘ᶱሿ٬ǄBishop˄1996˅वਜ਼Ҷ‫ޣ‬Ҿ䇝㓳㖁㔌Ⲵࠐ⿽৲ᮠՈॆᯩ⌅Ⲵа㡜ᙗ
䇘䇪Ǆ
4.8.3 䙈ᖈ㖇㔒δRecurrent Networksε
ⴤࡠ⧠൘ᡁԜ㘳㲁Ⲵਚᱟᴹੁᰐ⧟Ⲵ㖁㔌ᤃᢁ㔃ᶴǄ䙂ᖂ㖁㔌ᱟᴹྲл⢩ᖱⲴӪᐕ⾎㓿
㖁㔌˖䘲⭘Ҿᰦᒿᮠᦞ˗֯⭘㖁㔌অ‫ݳ‬൘ᰦ䰤 t Ⲵ䗃ࠪ֌ѪަԆঅ‫ݳ‬൘ᰦ䰤 t+1 Ⲵ䗃‫ޕ‬Ǆԕ
䘉⿽ᯩᔿˈ䙂ᖂ㖁㔌᭟ᤱ൘㖁㔌ѝ֯⭘Ḁ⿽ᖒᔿⲴᴹੁ⧟˄directed cycles˅ǄѪҶ╄⽪䙂ᖂ
㖁㔌ˈ㘳㲁ањᰦᒿ亴⍻ԫ࣑üüṩᦞᖃཙⲴ㓿⍾ᤷḷ x(t)ˈ亴⍻лаཙⲴ㛑⾘ᒣ൷ᐲ٬
y(t+1)Ǆ㔉ᇊҶ䘉ṧⲴᰦᒿᮠᦞˈањᱮ㘼᱃㿱Ⲵ࣎⌅ᱟṩᦞ䗃‫ޕ‬٬ x(t)䇝㓳ањࡽ侸㖁㔌
亴⍻䗃ࠪ y(t+1)Ǆањ䘉ṧⲴ㖁㔌ᱮ⽪൘മ 4-11˄a˅ѝǄ
䘉ṧⲴ㖁㔌Ⲵ㕪⛩ᱟӵ‫׍‬䎆 x(t)֌ࠪሩ y(t+1)亴⍻ˈ㘼н㜭ᦅ᥹ y(t+1)ሩ x Ⲵԕࡽ٬Ⲵ‫׍‬

䎆ᙗǄ㘼䘉ਟ㜭ᱟᗵ䴰Ⲵˈֻྲˈ᰾ཙⲴ㛑⾘ᒣ൷ᐲ٬ਟ㜭‫׍‬䎆ҾӺཙⲴ㓿⍾ᤷḷ઼᱘ཙⲴ
㓿⍾ᤷḷⲴᐞᔲǄᖃ❦ᡁԜਟԕ䙊䗷ᢺ x(t)઼ x(t-1)䜭֌Ѫࡽ侸㖁㔌Ⲵ䗃‫ˈޕ‬ᶕᕕ㺕䘉њн
䏣Ǆնᱟྲ᷌ᡁԜᐼᵋ䘉њ㖁㔌亴⍻ y(t+1)ᰦ㘳㲁ԫ᜿䗷৫Ⲵᰦ䰤デ޵Ⲵؑ᚟઒˛䛓Ѹቡ䴰
㾱⭘н਼Ⲵ䀓ߣᯩṸҶǄമ 4-11˄b˅ᱮ⽪Ⲵ䙂ᖂ㖁㔌ᨀ‫׋‬Ҷањ䘉ṧⲴ䀓ߣᯩṸǄ䘉䟼ᡁ
Ԝੁ䳀㯿ቲ࣐ҶањᯠⲴঅ‫ ݳ‬b ઼ᯠⲴ䗃‫ޕ‬অ‫ ݳ‬c(t)Ǆc(t)Ⲵ٬㻛ᇊѹѪঅ‫ ݳ‬b ൘ᰦ䰤 t-1 Ⲵ
٬˗ҏቡᱟ䈤ˈ㖁㔌൘Ḁањᰦ䰤↕˄time step˅Ⲵ䗃‫ޕ‬٬ c(t)ᤧ䍍㠚অ‫ ݳ‬b ൘ࡽаᰦ䰤↕
Ⲵ٬Ǆ⌘᜿䘉ᇎ⧠Ҷа⿽䙂ᖂ‫ޣ‬㌫ˈަѝ b 㺘⽪‫ޣ‬Ҿ㖁㔌䗃‫Ⲵޕ‬শਢؑ᚟ǄഐѪ b ᰒ‫׍‬䎆Ҿ
x(t)৸‫׍‬䎆Ҿ c(t)ˈᡰԕ b ਟ㜭ᾲᤜҶ x ԕࡽԫ᜿ᰦ䰤䐍⿫Ⲵ٬ǄᖸཊަԆⲴ㖁㔌ᤃᢁҏਟԕ
⭘ᶕ㺘⽪䙂ᖂ㖁㔌ǄֻྲˈᡁԜਟԕ൘䗃‫઼ޕ‬অ‫ ݳ‬b 䰤ᨂ‫ޕ‬㤕ᒢቲঅ‫ˈݳ‬ҏਟԕ൘࣐‫ޕ‬অ‫ݳ‬
b ઼䗃‫ޕ‬অ‫ ݳ‬c Ⲵൠᯩ޽ᒦ㹼ᨂ‫ࠐޕ‬њঅ‫ݳ‬Ǆ
Feedforward network-ࡽ侸㖁㔌
Recurrent network-䙂ᖂ㖁㔌
Recurrent network unfolded in time-᤹ᰦ䰤ኅᔰⲴ䙂ᖂ㖁㔌
മ 4-11 䙂ᖂ㖁㔌
ྲօ䇝㓳䘉ṧⲴ䙂ᖂ㖁㔌઒˛䙂ᖂ㖁㔌ᴹཊ⿽ਈփˈഐ↔ӪԜҏ࠶࡛ᨀࠪҶн਼Ⲵ䇝㓳
ᯩ⌅˄ֻྲ৲㿱Jordan 1986; Elman 1990; Mozer 1995; Williams & Zipser 1995˅Ǆᴹ䏓Ⲵᱟˈ
䊑മ 4-11˄b˅䛓ṧⲴ䙂ᖂ㖁㔌ਟԕㆰঅ֯⭘৽ੁՐ᫝㇇⌅Ⲵਈփᶕ䇝㓳ǄѪҶ⨶䀓ྲօᇎ
ᯭˈ㘳㲁മ 4-11˄c˅ˈᱮ⽪Ҷ䙂ᖂ㖁㔌᤹➗ᰦ䰤ኅᔰⲴᮠᦞ⍱Ǆ䘉䟼ᡁԜᢺ䙂ᖂ㖁㔌ᤧ䍍ᡀ
ࠐԭˈ⭘н਼ᤧ䍍䰤Ⲵ䘎᧕ᴯᦒᦹ৽侸⧟Ǆ⌘᜿䘉њབྷⲴ㖁㔌н޽वਜ਼എ䐟Ǆᡰԕኅᔰ㖁㔌
Ⲵᵳ٬ਟԕⴤ᧕֯⭘৽ੁՐ᫝㇇⌅ᶕ䇝㓳Ǆᖃ❦ᇎ䐥ѝᡁԜᐼᵋӵ‫⮉؍‬аԭ䙂ᖂ㖁㔌઼ᵳ٬
䳶ਸⲴᤧ䍍Ǆᡰԕˈ൘䇝㓳ҶኅᔰⲴ㖁㔌ਾˈਟԕਆн਼ᤧ䍍ѝᵳ٬wjiⲴᒣ൷٬֌Ѫᴰ㓸㖁
㔌ⲴሩᓄⲴᵳ٬wjiǄMozer˄1995˅䶎ᑨ䈖㓶ൠ᧿䘠Ҷ䘉њ䇝㓳䗷〻Ǆᇎ䐥ѝˈ䙂ᖂ㖁㔌∄
⋑ᴹ৽侸⧟Ⲵ㖁㔌䳮ԕ䇝㓳ˈ⌋ॆⲴਟ䶐ᙗҏнྲਾ㘵Ǆ❦㘼ᆳԜӽ❦ഐ䖳ᕪⲴ㺘ᖱ࣋‫؍‬ᤱ
⵰䟽㾱ᙗǄ
4.8.4 ࣞᘷ‫ؤ‬᭯㖇㔒㔉ᶺ
ⴤࡠ⧠൘ᡁԜ㘳㲁Ⲵ⾎㓿㖁㔌ᆖҐ䰞仈ᱟ䈳ᮤањപᇊ㖁㔌㔃ᶴѝⲴᵳ٬ǄѪҶ᭩ழ⌋
ॆ㋮ᓖ઼䇝㓳᭸⦷ˈӪԜᨀࠪҶᖸཊࣘᘱ໎䮯ᡆ঻㕙㖁㔌অ‫઼ݳ‬অ‫ݳ‬䰤䘎᧕ᮠ䟿Ⲵᯩ⌅Ǆ
а⿽ᜣ⌅ᱟӾањнवਜ਼䳀㯿অ‫Ⲵݳ‬㖁㔌ᔰ࿻ˈ❦ਾṩᦞ䴰㾱໎࣐䳀অ‫ݳ‬໎䮯㖁㔌ˈⴤ
ࡠ䇝㓳䈟ᐞл䱽ࡠḀњਟ᧕ਇⲴ≤ᒣǄ㓗㚄⴨‫˄ޣ‬Cascade-Correlation˅㇇⌅˄Fahlman &
Lebiere 1990˅ቡᱟ䘉ṧа⿽㇇⌅Ǆ㓗㚄⴨‫ޣ‬㇇⌅Ӿࡋᔪањ⋑ᴹ䳀অ‫Ⲵݳ‬㖁㔌ᔰ࿻Ǆֻྲˈ
ሩҾᡁԜⲴӪ㝨ᵍੁⲴᆖҐԫ࣑ˈᆳՊᔪ・ањӵवਜ਼ഋњ䗃ࠪঅ‫ޘݳ‬䘎᧕ࡠ 30u32 њ䗃‫ޕ‬
㔃⛩Ⲵ㖁㔌Ǆ൘䘉њ㖁㔌㻛䇝㓳Ҷа⇥ᰦ䰤ਾˈᡁԜਟԕᖸᇩ᱃ൠਁ⧠䘈ᴹ䖳བྷⲴ↻⮉䈟ᐞˈ
ഐѪһᇎк䘉њⴞḷ࠭ᮠнਟ㜭㻛ањঅቲ㔃ᶴⲴ㖁㔌⨶ᜣൠ㺘⽪Ǆ൘䘉⿽ᛵߥлˈ㇇⌅໎
࣐ањ䳀㯿অ‫ˈݳ‬䘹ᤙᆳⲴᵳ٬֯䘉њ䳀㯿অ‫Ⲵݳ‬٬઼ᮤњ㖁㔌Ⲵ↻⮉䈟ᐞⲴ⴨‫ޣ‬ᙗᴰབྷ
ॆǄ⧠൘ањᯠⲴঅ‫ݳ‬㻛ᆹ㻵䘋Ҷ㖁㔌ˈᆳⲴᵳ٬‫؍‬ᤱнਈˈᒦф໎࣐䘉њᯠঅ‫⇿ࡠݳ‬ањ
䗃ࠪঅ‫ݳ‬䰤Ⲵ䘎᧕Ǆ䟽༽䘉њ䗷〻Ǆ৏࿻Ⲵᵳ٬㻛޽⅑䇝㓳˄‫؍‬ᤱ䳀㯿অ‫Ⲵݳ‬ᵳ٬нਈ˅ˈ
Ựḕ↻⮉䈟ᐞˈྲ᷌↻⮉䈟ᐞ䘈儈Ҿ䰸٬ቡ࣐‫ޕ‬ㅜҼњ䳀অ‫ݳ‬Ǆ⇿ᖃ࣐‫ޕ‬ањᯠⲴ䳀㯿অ‫ˈݳ‬
ᆳⲴ䗃‫ޕ‬वᤜᡰᴹ৏࿻Ⲵ㖁㔌䗃‫઼ޕ‬ᐢ㓿ᆈ൘Ⲵ䳀㯿অ‫Ⲵݳ‬䗃ࠪǄ㖁㔌ԕ䘉⿽ᯩᔿ໎䮯ˈ〟
㚊䳀㯿অ‫ࡠⴤˈݳ‬㖁㔌Ⲵ↻։䈟ᐞл䱽ࡠḀњਟ᧕ਇⲴ≤ᒣǄFahlman & Lebiere˄1990˅ᣕ
੺Ҷ㓗㚄⴨‫ޣ‬㇇⌅ᱮ㪇߿ቁ䇝㓳ᰦ䰤Ⲵֻᆀˈ৏ഐᱟ⇿а↕ӵᴹаቲ㖁㔌൘㻛䇝㓳Ǆ䘉њ㇇
⌅Ⲵањᇎ䱵ഠ䳮ᱟഐѪ㇇⌅ਟԕᰐ䲀ࡦൠ໎࣐অ‫ˈݳ‬ᆳቡᖸᇩ᱃䗷ᓖᤏਸ䇝㓳ᮠᦞˈᡰԕ
ᗵ享䟷ਆ䚯‫ݽ‬䗷ᓖᤏਸⲴ亴䱢᧚ᯭǄ
ࣘᘱ‫؞‬᭩㖁㔌㔃ᶴⲴㅜҼњᜣ⌅ᱟ֯⭘⴨৽Ⲵ䙄ᖴǄн޽Ӿਟ㜭Ⲵᴰㆰঅ㖁㔌ᔰ࿻໎࣐
༽ᵲᙗˈ㘼ᱟӾањ༽ᵲⲴ㖁㔌ᔰ࿻‫ᦹ࢚؞‬ḀӋᰐ‫ޣ‬㍗㾱Ⲵ䘎᧕ǄࡔᯝḀњᵳᱟ੖ᰐ‫ޣ‬㍗㾱
Ⲵа⿽ᯩ⌅ᱟⴻᆳⲴ٬ᱟ੖᧕䘁 0ǄㅜҼ⿽ⴻᶕ൘ᇎ䐥ѝᴤ࣐ᡀ࣏Ⲵᯩ⌅ᱟ㘳㲁䘉њᵳ٬Ⲵ
wE
ањሿⲴਈॆሩ䈟ᐞ E Ⲵᖡ૽Ǆਈॆ w ሩ E Ⲵᖡ૽˄ҏቡᱟ ˅ਟԕ㻛ⴻ֌㺑䟿䘉њ䘎
ww
᧕Ⲵᱮ㪇ᙗ˄salient˅ⲴቪᓖǄLeCun et al.˄1990˅᧿䘠Ҷањ㖁㔌㻛䇝㓳Ⲵ䗷〻ˈᴰнᱮ
㪇Ⲵ䘎᧕㻛᣶䲔ˈ䟽༽䘉њ䗷〻ⴤࡠ䙷ࡠḀњ㓸→ᶑԦǄԆԜ〠䘉⿽ᯩ⌅ѪĀᴰՈ㝁ᦏՔ
˄optimal brain damage˅ā⌅ˈഐѪ൘⇿а↕㇇⌅䜭䈅മ৫䲔ᴰ⋑ᴹ⭘Ⲵ䘎᧕ǄԆԜᣕ੺Ҷ
൘ањᆇㅖ䇶࡛ᓄ⭘ѝ䘉⿽ᯩ⌅ሶањབྷⲴ㖁㔌ѝᵳ٬߿ቁࡠഋ࠶ѻаˈሩ⌋ॆ㋮ᓖᴹᗞሿ
Ⲵ᭩ழˈᒦфབྷབྷ᭩ழҶਾᶕⲴ䇝㓳᭸⦷Ǆ
а㡜㘼䀰ˈࣘᘱ‫؞‬᭩㖁㔌㔃ᶴⲴᯩ⌅ᐢ㓿ਆᗇҶаӋᡀ࣏ˈնҏᴹн䏣Ǆ䘉⿽ᯩ⌅ᱟ੖
㜭っᇊൠᨀ儈৽ੁՐ᫝㇇⌅Ⲵ⌋ॆ㋮ᓖ䘈ᴹᖵ⹄ウǄ❦㘼ᐢ㓿䇱᰾൘аӋᛵᖒлᆳਟԕᱮ㪇
ൠ䱽վ䇝㓳ᰦ䰤Ǆ
䘉аㄐⲴ㾱⛩वᤜ˖
x Ӫᐕ⾎㓿㖁㔌ᆖҐѪᆖҐᇎᮠ٬઼ੁ䟿٬࠭ᮠᨀ‫׋‬Ҷа⿽ᇎ䱵Ⲵᯩ⌅ˈሩҾ
䘎㔝Ⲵ઼⿫ᮓ٬Ⲵ኎ᙗ䜭ਟԕ֯⭘ˈᒦфሩ䇝㓳ᮠᦞѝⲴಚ༠ᴹᖸྭⲴ励ἂ
ᙗǄ৽ੁՐ᫝㇇⌅ᱟᴰᑨ㿱Ⲵ㖁㔌ᆖҐ㇇⌅ˈᐢ㓿ᡀ࣏ᓄ⭘ࡠᖸཊᆖҐԫ࣑ˈ
∄ྲ᡻߉䇶઼࡛ᵪಘӪ᧗ࡦǄ
x ৽ੁՐ᫝㇇⌅㘳㲁Ⲵ‫ٷ‬䇮オ䰤ᱟപᇊ䘎᧕Ⲵᴹᵳ㖁㔌ᡰ㜭㺘⽪Ⲵᡰᴹ࠭ᮠ
オ䰤Ǆवਜ਼йቲঅ‫ࡽⲴݳ‬侸㖁㔌㜭ཏԕԫ᜿㋮ᓖ䙬䘁ԫ᜿࠭ᮠˈਚ㾱⇿аቲ
ᴹ䏣ཏᮠ䟿˄ਟ㜭䶎ᑨཊ˅Ⲵঅ‫ݳ‬Ǆণ֯ᱟањᇎ䱵བྷሿⲴ㖁㔌ҏ㜭ཏ㺘⽪
ᖸབྷ㤳തⲴ儈ᓖ䶎㓯ᙗⲴ࠭ᮠˈ䘉֯ᗇࡽ侸㖁㔌ᡀѪᆖҐ亴‫ݸ‬ᵚ⸕Ⲵа㡜ᖒ
ᔿⲴ⿫ᮓ઼䘎㔝࠭ᮠⲴᖸྭ䘹ᤙǄ
x ৽ੁՐ᫝㇇⌅֯⭘ởᓖл䱽ᯩ⌅ᩌ㍒ਟ㜭‫ٷ‬䇮Ⲵオ䰤ˈ䘝ԓ߿ሿ㖁㔌Ⲵ䈟ᐞ
ԕᤏਸ䇝㓳ᮠᦞǄởᓖл䱽᭦ᮋࡠ䇝㓳䈟ᐞ⴨ሩ㖁㔌ᵳ٬Ⲵተ䜘ᶱሿ٬Ǆᴤ
а㡜Ⲵˈởᓖл䱽ᱟа⿽ᴹᓄ⭘▌࣋Ⲵᯩ⌅ˈᆳਟ⭘ᶕᩌ㍒ᖸཊ䘎㔝৲ᮠⲴ
‫ٷ‬䇮オ䰤ˈਚ㾱䇝㓳䈟ᐞᱟ‫ٷ‬䇮৲ᮠⲴਟᗞ࠭ᮠǄ
x ৽ੁՐ᫝㇇⌅ᴰԔӪᝏ‫ޤ‬䏓Ⲵ⢩ᖱѻаᱟˈᆳ㜭ཏࡋ䙐ࠪ㖁㔌䗃‫ޕ‬ѝ⋑ᴹ᰾
⺞ࠪ⧠Ⲵ⢩ᖱǄ⺞࠷ൠ䇢ˈཊቲ㖁㔌Ⲵ޵䜘˄䳀㯿˅ቲ㜭ཏ㺘⽪ሩᆖҐⴞḷ
࠭ᮠᴹ⭘Ⲵն䳀ਜ਼൘㖁㔌䗃‫ޕ‬ѝⲴѝ䰤⢩ᖱǄ䘉⿽㜭࣋㻛ֻᆀྲ 4.6.4 㢲Ⲵ
8u3u8 㖁㔌ѝࡋ䙐Ⲵᮠᆇ 1 ࡠ 8 Ⲵᐳቄ㕆⸱˗ԕ৺ 4.7 㢲Ӫ㝨䇶࡛ᓄ⭘ѝ䳀
㯿ቲ㺘⽪Ⲵമ‫ۿ‬⢩ᖱǄ
x 䗷ᓖᤏਸ䇝㓳ᮠᦞᱟ ANN ᆖҐѝⲴањ䟽㾱䰞仈Ǆ䗷ᓖᤏਸሬ㠤㖁㔌⌋ॆ
ࡠᯠⲴᮠᦞᰦᙗ㜭ᖸᐞˈቭ㇑㖁㔌ሩҾ䇝㓳ᮠᦞ㺘⧠䶎ᑨྭǄӔ৹傼䇱ᯩ⌅
ਟԕ⭘ᶕՠ䇑ởᓖл䱽ᩌ㍒Ⲵਸ䘲㓸→⛩ˈӾ㘼ᴰሿॆ䗷ᓖᤏਸⲴ仾䲙Ǆ
x ቭ㇑৽ੁՐ᫝㇇⌅ᱟᴰᑨ㿱Ⲵ ANN ᆖҐ㇇⌅ˈӪԜҏᨀࠪᖸཊަԆⲴ㇇⌅ˈ
वᤜሩҾ⢩↺ԫ࣑ⲴаӋ㇇⌅Ǆֻྲˈ䙂ᖂ㖁㔌ᯩ⌅䇝㓳वਜ਼ᴹੁ⧟Ⲵ㖁㔌ˈ
㊫լ㓗㚄⴨‫Ⲵޣ‬㇇⌅᭩ਈᵳⲴ਼ᰦҏ᭩ਈ㖁㔌㔃ᶴǄ
ᵜҖⲴަԆㄐ㢲ҏӻ㓽ҶаӋ‫ޣ‬Ҿ ANN ᆖҐⲴަԆؑ᚟Ǆㅜ 6 ㄐ㔉ࠪҶ䘹ᤙᴰሿॆ䈟

ᐞᒣᯩ઼Ⲵ䍍ਦᯟ䇪䇱ˈԕ৺൘ަԆᛵߥл⭘ᴰሿॆӔ৹⟥˄cross entropy˅ԓᴯᴰሿॆ䈟
ᐞᒣᯩ઼Ⲵᯩ⌅Ǆㅜ 7 ㄐ䇘䇪ҶѪਟ䶐ᆖҐᐳቄ࠭ᮠᡰ䴰㾱Ⲵ䇝㓳ᇎֻᮠ䟿Ⲵ⨶䇪㔃᷌ˈԕ
৺ḀӋ㊫ර㖁㔌Ⲵ Vapnik-Chervonenkis 㔤Ǆ‫ޣ‬Ҿ䗷ᓖᤏਸԕ৺ྲօ䚯‫Ⲵݽ‬䇘䇪ਟԕ൘ㅜ 5
ㄐѝ᢮ࡠǄㅜ 12 ㄐ䇘䇪Ҷ֯⭘ԕࡽⲴ⸕䇶ᶕᨀ儈⌋ॆ㋮ᓖⲴᯩ⌅Ǆ
ሩӪᐕ⾎㓿㖁㔌Ⲵ⹄ウਟԕ䘭ⓟࡠ䇑㇇ᵪ、ᆖⲴᰙᵏǄMcCulloch & Pitts˄1943˅ᨀࠪ

Ҷањ⴨ᖃҾᝏ⸕ಘⲴ⾎㓿‫⁑ݳ‬රˈ60 ᒤԓⲴབྷ䟿ᐕ֌᧒㍒Ҷ䘉њ⁑රⲴᖸཊਈփǄ60 ᒤ
ԓᰙᵏ Widrow & Hoff˄1960˅᧒㍒Ҷᝏ⸕ಘ㖁㔌˄ԆԜ〠ѪĀadelinesā
˅઼ delta ⌅ࡉˈ
Rosenblatt˄1962˅䇱᰾Ҷᝏ⸕ಘ䇝㓳⌅ࡉⲴ᭦ᮋᙗǄ❦㘼ˈⴤࡠ 60 ᒤԓᲊᵏˈӪԜᔰ࿻␵
ᾊঅቲⲴᝏ⸕ಘ㖁㔌Ⲵ㺘ᖱ㜭࣋ᖸᴹ䲀ˈ㘼ф᢮нࡠ䇝㓳ཊቲ㖁㔌Ⲵᴹ᭸ᯩ⌅ǄMinsky &
Papert˄1969˅䈤᰾ণ֯ᱟ䊑 XOR 䘉ṧㆰঅⲴ࠭ᮠҏн㜭⭘অቲⲴᝏ⸕ಘ㖁㔌㺘⽪ᡆᆖҐˈ
൘ᮤњ 70 ᒤԓ ANN Ⲵ⹄ウ㺠䘰ҶǄ
൘ 80 ᒤԓѝᵏ ANN Ⲵ⹄ウ㓿শҶа⅑༽‫ˈޤ‬ѫ㾱ᱟഐѪ䇝㓳ཊቲ㖁㔌Ⲵ৽ੁՐ᫝㇇
⌅Ⲵਁ᰾˄Rumelhart & McClelland 1986˗Parker 1985˅Ǆ䘉Ӌᙍᜣਟԕ㻛䘭ⓟࡠᴹ‫Ⲵޣ‬ᰙᵏ
⹄ウ˄ֻྲ Werbos 1975˅Ǆ㠚Ӿ 80 ᒤԓˈ৽ੁՐ᫝㇇⌅ቡᡀѪᓄ⭘ᴰᒯ⌋ⲴᆖҐᯩ⌅ˈ㘼
фӪԜҏ〟ᶱ᧒㍒ࠪҶᖸཊަԆⲴ ANN ᯩ⌅Ǆ൘਼аᰦᵏˈ䇑㇇ᵪਈᗇн޽䍥䟽ˈ䘉‫ݱ‬䇨
ӪԜ䈅傼䛓Ӌ൘ 60 ᒤԓнਟ㜭㻛ᆼ‫᧒ޘ‬㍒Ⲵ䇑㇇ᇶ䳶ᙗⲴ㇇⌅Ǆ
ᖸཊᮉ、Җу䰘䇪䘠Ҷ⾎㓿㖁㔌ᆖҐǄаᵜᰙᵏⲴնӽᴹ⭘Ⲵ‫ޣ‬Ҿ⁑ᔿ䇶࡛Ⲵ৲ᮠᆖҐ
ᯩ⌅ⲴҖᱟ Duda & Hart˄1973˅ǄWindrow & Stearns˄1985˅Ⲵᮉ、Җ㾶ⴆҶᝏ⸕ಘ઼⴨‫ޣ‬
Ⲵঅቲ㖁㔌ԕ৺ᆳԜⲴᓄ⭘ǄRumelhart & McClelland˄1986˅᭦㕆Ҷ 80 ᒤԓѝᵏᔰ࿻Ⲵ䟽
ᯠ◰ਁ䎧ӪԜሩ⾎㓿㖁㔌ᯩ⌅‫ޤ‬䏓Ⲵ䇪᮷Ǆ‫ޣ‬Ҿ⾎㓿㖁㔌ᴰ䘁ࠪ⡸ⲴҖ㉽वᤜ Bishop
˄1996˅˗Chauvin & Rumelhart˄1995˅˗Freeman & Skapina˄1991˅
˗Fu˄1994˅˗Hecht-Nielson
˄1990˅઼ Hertz et al.˄1991˅Ǆ
Ґ仈
4.1 ሩമ 4-3 ⭫ࠪⲴ䈟ᐞᴢ䶒ˈᝏ⸕ಘⲴᵳw0ˈw1઼w2Ⲵ٬ᱟӰѸ˛‫ٷ‬ᇊ䘉њ䈟ᐞᴢ䶒

оx1䖤⴨Ӕ൘x1= -1ˈᒦоx2䖤⴨Ӕ൘x2 = 2Ǆ
4.2 䇮䇑ањє䗃‫Ⲵޕ‬ᝏ⸕ಘᶕᇎ⧠ᐳቄ࠭ᮠ ABǄ䇮䇑ањєቲⲴᝏ⸕ಘ㖁㔌ᶕᇎ

⧠ᐳቄ࠭ᮠ A XOR BǄ
4.3 㘳㲁֯⭘䰸٬㺘䗮ᔿw0 + w1x1 + w2x2 > 0 ᇊѹⲴєњᝏ⸕ಘǄᝏ⸕ಘAⲴᵳ٬Ѫ
w0=1ˈw1=2ˈw2=1
ᝏ⸕ಘ B Ⲵᵳ٬Ѫ
w0=0ˈw1=2ˈw2=1
䈧 ࡔ ᯝ ԕ л 㺘䗮 ሩ ᡆ 䭉 Ǆ ᝏ ⸕ ಘ A ᱟ more_general_than ᝏ ⸕ ಘ B Ⲵ Ǆ
˄more_general_than ൘ㅜ 2 ㄐѝᇊѹ˅
4.4 ᇎ⧠ањє䗃‫ޕ‬㓯ᙗঅ‫Ⲵݳ‬delta䇝㓳⌅ࡉǄ䇝㓳ᆳᶕᤏਸⴞḷᾲᘥ-2+x1+2x2>0Ǆ⭫
ࠪ䈟ᐞE⴨ሩ䇝㓳䘝ԓ⅑ᮠⲴ࠭ᮠᴢ㓯Ǆ⭫ࠪ 5ˈ10ˈ50ˈ100ˈĂĂ⅑䘝ԓਾⲴߣㆆ䶒Ǆ
˄a˅ ѪK䘹ਆн਼Ⲵᑨ䟿٬ˈᒦ֯⭘㺠߿ⲴᆖҐ䙏⦷üüҏቡᱟㅜi⅑䘝ԓ
֯⭘K0/iˈ޽䘋㹼䈅傼Ǆଚањ᭸᷌ᴤྭ˛
˄b˅ 䈅傼໎䟿˄incremental˅઼ᢩ䟿˄batch˅ᆖҐǄ䛓њ᭦ᮋᗇᴤᘛ˛㘳
㲁ᵳ٬ᴤᯠ⅑ᮠ઼ᙫᢗ㹼ᰦ䰤Ǆ
4.5 ᧘ሬ䗃ࠪѪ o Ⲵঅњঅ‫Ⲵݳ‬ởᓖл䱽䇝㓳⌅ࡉˈަѝ
o w0 w1 x1 w1 x12 wn x n wn x n2
4.6 ㆰ⮕Ⲵ䀓䟺ѪӰѸ‫ޜ‬ᔿ˄4.10˅ѝⲴ delta ⌅ࡉӵᱟ‫ޜ‬ᔿ˄4.7˅㺘⽪Ⲵⵏ↓ởᓖл䱽

⌅ࡉⲴ䘁լ˛
4.7 㘳㲁ањєቲⲴࡽ侸ANNˈᆳާᴹєњ䗃‫ޕ‬a઼bˈањ䳀অ‫ݳ‬cˈ઼ањ䗃ࠪঅ‫ݳ‬dǄ
䘉њ㖁㔌ᴹӄњᵳ٬˄wcaˈwcbˈwc0ˈwdcˈwd0˅ˈަѝwx0㺘⽪অ‫ݳ‬xⲴ䰸٬ᵳǄ‫ݸ‬ᢺ䘉Ӌᵳ
Ⲵ٬ࡍ࿻ॆѪ˄0.1ˈ0.1ˈ0.1ˈ0.1ˈ0.1˅ˈ❦ਾ㔉ࠪ֯⭘৽ੁՐ᫝㇇⌅䇝㓳䘉њ㖁㔌Ⲵࡽє
⅑䘝ԓѝ⇿а⅑䘉Ӌᵳ٬Ⲵ٬Ǆ‫ٷ‬ᇊᆖҐ䙏⦷K=0.3ˈߢ䟿D=0.9ˈ䟷⭘໎䟿Ⲵᵳ٬ᴤᯠˈ઼
ԕл䇝㓳ṧֻ˖
a b d
1 0 1
0 1 0
4.8 ‫؞‬᭩㺘 4-2 ѝⲴ৽ੁՐ᫝㇇⌅ˈ֯⭘ৼᴢ↓࠷tanh࠭ᮠਆԓsigmoid࠭ᮠ֌Ѫᥔ঻࠭

& &
ᮠǄҏቡᱟ䈤ˈ‫ٷ‬ᇊঅњঅ‫Ⲵݳ‬䗃ࠪᱟo=tanh˄ w x ˅Ǆ㔉ࠪ䗃ࠪቲᵳ٬઼䳀㯿ቲᵳ٬Ⲵᵳ
ᴤᯠ⌅ࡉǄᨀ⽪˖tanhc(x)=1-tanh2(x)Ǆ
4.9 എᗶമ 4-7 ᧿䘠Ⲵ 8u3u8 㖁㔌Ǆ㘳㲁䇝㓳ањ 8u1u8 Ⲵ㖁㔌ᶕᆼᡀ਼ṧⲴԫ࣑˗ҏ

ቡᱟӵᴹањ䳀㯿অ‫Ⲵݳ‬㖁㔌Ǆ⌘᜿ˈമ 4-7 ѝⲴ 8 њ䇝㓳ṧֻਟԕ㻛㺘⽪Ѫঅњ䳀অ‫Ⲵݳ‬
8 њн਼Ⲵ٬˄ֻྲ 0.1ˈ0.2ˈĂĂˈ0.8˅Ǆ䛓Ѹӵᴹањ䳀অ‫Ⲵݳ‬㖁㔌㜭ཏṩᦞ䘉Ӌ䇝㓳
ṧֻᆖҐᚂㅹ࠭ᮠੇ˛ᨀ⽪˖㘳㲁㊫լ䘉ṧⲴ䰞仈Āᱟ੖ᆈ൘䘉ṧⲴ䳀㯿অ‫ݳ‬ᵳ٬ˈ㜭ӗ⭏
к䶒ᔪ䇞Ⲵ䳀㯿অ‫ݳ‬㕆⸱˛āˈĀᱟ੖ᆈ൘䘉ṧⲴ䗃ࠪঅ‫ݳ‬ᵳ٬ˈ㜭↓⺞䀓⸱䘉ṧⲴ䗃‫ޕ‬㕆
⸱˛ā઼Āởᓖл䱽ᩌ㍒ਟ㜭ਁ⧠䘉ṧⲴᵳ٬ੇ˛ā
4.10 㘳㲁 4.8.1 ሿ㢲ѝ᧿䘠Ⲵਖа⿽䈟ᐞ࠭ᮠ˖
& 1
E ( w) { ¦ ¦ (t kd o kd )2 J ¦ w ji
2
2 dD koutputs i, j
Ѫ䘉њ䈟ᐞ E ᧘ሬࠪởᓖл䱽ᵳᴤᯠ⌅ࡉǄ䇱᰾䘉њᵳ٬ᴤᯠ⌅ࡉⲴᇎ⧠ਟ䙊䗷൘䘋
㹼㺘 4-2 Ⲵḷ߶ởᓖл䱽ᵳᴤᯠࡽᢺ⇿њᵳ٬҈ԕањᑨᮠǄ
4.11 ᓄ ⭘ ৽ ੁ Ր ᫝ ㇇ ⌅ ᶕ ᆼ ᡀ Ӫ 㝨䇶 ࡛ ԫ ࣑ Ǆ ৲ 㿱 ӂ 㚄㖁亥
http://www.cs.cmu.edu/~tom/mlbook.html ᶕ㧧ᗇަ㓶㢲ˈवᤜӪ㝨മ‫ۿ‬ᮠᦞˈ৽ੁՐ᫝〻ᒿ
Ⓚԓ⸱઼ާփⲴԫ࣑Ǆ
4.12 ᧘ሬࠪᆖҐ xˈy ᒣ䶒кⲴ⸙ᖒ䘉аⴞḷᾲᘥⲴởᓖл䱽㇇⌅Ǆ֯⭘ xˈy Ⲵ඀ḷ᧿

䘠⇿ањ‫ٷ‬䇮ˈ⸙ᖒⲴᐖл䀂઼ਣк䀂࠶࡛㺘⽪Ѫ llxˈllyˈurx ઼ uryǄᇎֻ<x,y>㻛‫ٷ‬䇮<llxˈ
llyˈurxˈury>ḷ䇠Ѫ↓ֻⲴ‫ݵ‬㾱ᶑԦᱟ⛩<x,y>սҾሩᓄⲴ⸙ᖒ޵䜘Ǆ᤹ᵜㄐѝⲴ࣎⌅ᇊѹ
䈟ᐞ EǄ䈅䇮䇑ањởᓖл䱽㇇⌅ᶕᆖҐ䘉ṧⲴ⸙ᖒ‫ٷ‬䇮Ǆ⌘᜿䈟ᐞ E нᱟ llxˈllyˈurx ઼
ury Ⲵ䘎㔝࠭ᮠˈ䘉оᝏ⸕ಘᆖҐⲴᛵߥаṧǄ˄ᨀ⽪˖㘳㲁ᝏ⸕ಘѝ֯⭘Ⲵєњ䀓ߣ࣎⌅˖
˄1˅᭩ਈ࠶㊫⌅ࡉᶕ֯䗃ࠪ亴⍻ᡀѪ䗃‫Ⲵޕ‬䘎㔝࠭ᮠ˗˄2˅ਖཆᇊѹањ䈟ᐞüü∄ྲࡠ
⸙ᖒѝᗳⲴ䐍⿫üüቡ‫ۿ‬䇝㓳ᝏ⸕ಘⲴ delta ⌅ࡉǄ˅ᖃ↓઼ֻ৽ֻਟ㻛⸙ᖒ࠶ࢢᰦˈ䇮䇑Ⲵ
㇇⌅Պ᭦ᮋࡠᴰሿ䈟ᐞ‫ٷ‬䇮ੇ˛օᰦнՊ˛䈕㇇⌅ᴹተ䜘ᶱሿ٬Ⲵ䰞仈ੇ˛䈕㇇⌅оᆖҐ
⢩ᖱ㓖ᶏਸਆⲴㅖਧᯩ⌅⴨∄ྲօ˛
৲㘳᮷⥞
ㅢ5ㄖ䇺զ‫ٽ‬䇴
ᇍ‫؛‬䆒ⱘ㊒ᑺ䖯㸠㒣偠ⱘ䆘Ԅᰃᴎ఼ᄺдЁⱘ෎ᴀ䯂乬Ǆᴀゴҟ㒡њ⫼㒳䅵ᮍ⊩Ԅ䅵‫؛‬
䆒㊒ᑺˈЏ㽕Ў㾷‫އ‬ҹϟϝϾ䯂乬˖佪‫ˈܜ‬ᏆⶹϔϾ‫؛‬䆒೼᳝䰤᭄᥂ḋᴀϞ㾖ᆳࠄⱘ㊒ᑺˈ
ᗢḋԄ䅵ᅗ೼݊Ҫᅲ՟Ϟⱘ㊒ᑺǄ݊⃵ˈབᵰϔϾ‫؛‬䆒೼ᶤѯ᭄᥂ḋᴀϞདѢ঺ϔϾˈ䙷М
ϔ㠀ᚙ‫މ‬ϟᰃ৺䆹‫؛‬䆒᳈‫⹂ޚ‬Ǆ㄀ϝˈᔧ᭄᥂᳝䰤ᯊˈᗢḋ催ᬜഄ߽⫼䖭ѯ᭄᥂ˈ䗮䖛ᅗӀ
᮶㛑ᄺдࠄ‫؛‬䆒ˈ䖬㛑Ԅ䅵݊㊒ᑺ˛⬅Ѣ᳝䰤ⱘ᭄᥂ḋᴀৃ㛑ϡҷ㸼᭄᥂ⱘϔ㠀ߚᏗˈ᠔ҹ
Ң䖭ѯ᭄᥂ϞԄ䅵ߎⱘ‫؛‬䆒㊒ᑺৃ㛑᳝䇃ᏂǄ㒳䅵ⱘᮍ⊩ˈ㒧ড়᭄᳝݇᥂෎‫ߚޚ‬Ꮧⱘ‫؛‬ᅮˈ
Փ៥Ӏৃҹ⫼᳝䰤᭄᥂ḋᴀϞⱘ㾖ᆳ㊒ᑺᴹ䘐䖥ᭈϾ᭄᥂ߚᏗϞⱘⳳᅲ㊒ᑺǄ
5.1 ࣞᵰ
໮᭄ᚙ‫މ‬ϟˈᇍᄺдࠄⱘ‫؛‬䆒䖯㸠ሑৃ㛑‫ⱘ⹂ޚ‬ᗻ㛑䆘Ԅकߚ䞡㽕Ǆॳ಴Пϔᕜㅔऩˈ
ᰃЎњⶹ䘧ᰃ৺ৃҹՓ⫼䆹‫؛‬䆒Ǆ՟བˈҢϔϾ䭓ᑺ᳝䰤ⱘ᭄᥂ᑧЁᄺдˈҹњ㾷ϡৠए⭫
᠟↉ⱘᬜᵰˈህ᳝ᖙ㽕ሑৃ㛑‫⹂ޚ‬ഄⶹ䘧ᄺд㒧ᵰⱘℷ⹂ᗻǄ঺ϔॳ಴೼Ѣˈᇍ‫؛‬䆒ⱘ䆘Ԅ
ᰃ䆌໮ᄺдᮍ⊩ⱘ䞡㽕㒘៤䚼ߚǄ՟བ೼‫އ‬ㄪᷥᄺдЁˈЎ䙓‫ܡ‬䖛ᑺᢳড়䯂乬ᖙ乏䖯㸠ৢׂ
࠾ˈ䖭ᯊ៥Ӏᖙ乏䆘Ԅ↣ϔℹׂ࠾ᇍᷥⱘ㊒ᑺѻ⫳ⱘᕅડǄ಴ℸˈ᳝ᖙ㽕њ㾷Ꮖׂ࠾੠᳾ׂ
࠾ᷥⱘ㊒ᑺԄ䅵Ё೎᳝ⱘৃ㛑䇃ᏂǄ
ᔧ᭄᥂कߚ‫ܙ‬䎇ᯊˈ‫؛‬䆒㊒ᑺⱘԄ䅵ⳌᇍᆍᯧǄ✊㗠ᔧ㒭ᅮⱘ᭄᥂䲚䴲ᐌ᳝䰤ᯊˈ㽕ᄺ
дϔϾὖᗉᑊԄ䅵݊ᇚᴹⱘ㊒ᑺˈᄬ೼ϸϾᕜ݇䬂ⱘೄ䲒˖
z Ԅ䅵ⱘ‫أ‬Ꮒ(Bias in the estimate)Ǆ佪‫ˈܜ‬ᄺдࠄⱘὖᗉ೼䆁㒗ḋ՟Ϟⱘ㾖ᆳ㊒ᑺ
䗮ᐌϡ㛑ᕜདഄ⫼ѢԄ䅵೼ᇚᴹḋ՟Ϟⱘ㊒ᑺǄ಴Ў‫؛‬䆒ᰃҢ䖭ѯḋ՟Ёᕫߎ
ⱘˈ಴ℸᇍᇚᴹḋ՟ⱘ㊒ᑺԄ䅵䗮ᐌ‫أ‬ѢФ㾖Ǆᇸ݊೼ᄺд఼䞛⫼њᕜ໻ⱘ‫؛‬
䆒ぎ䯈ˈᑊ䖛ᑺᢳড়䆁㒗ḋ՟ᯊˈ䖭ϔᚙ‫މ‬᳈ৃ㛑ߎ⦄Ǆ㽕ᇍᇚᴹⱘ㊒ᑺ䖯㸠
᮴‫أ‬Ԅ䅵ˈ‫݌‬ൟⱘᮍ⊩ᰃ䗝ᢽϢ䆁㒗ḋ՟੠‫؛‬䆒᮴݇ⱘẔ偠ḋ՟ˈ೼䖭Ͼḋ՟
䲚ড়ϞẔ偠‫؛‬䆒Ǆ
z Ԅ䅵ⱘᮍᏂ(Variance in the estimate)Ǆ݊⃵ˈेՓ‫؛‬䆒㊒ᑺ೼⣀ゟⱘ᮴‫أ‬Ẕ偠ḋ
՟Ϟ⌟䞣ˈᕫࠄⱘ㊒ᑺҡৃ㛑Ϣⳳᅲ㊒ᑺϡৠˈ䖭প‫އ‬Ѣ⡍ᅮẔ偠ḋ՟䲚ড়ⱘ
㒘៤ǄẔ偠ḋ՟䍞ᇥˈѻ⫳ⱘᮍᏂ䍞໻Ǆ
ᴀゴ䅼䆎њᇍᄺࠄⱘ‫؛‬䆒ⱘ䆘ԄǃᇍϸϾ‫؛‬䆒㊒ᑺⱘ↨䕗ǃ੠᳝䰤᭄᥂ḋᴀᚙ‫މ‬ϟϸϾ
ᄺдㅫ⊩㊒ᑺⱘ↨䕗Ǆ݊Ёⱘ䅼䆎໮᭄䲚Ё೼㒳䅵੠䞛ḋ⧚䆎ⱘ෎ᴀᅮᕟˈ㗠ᴀゴ‫؛‬ᅮ䇏㗙
೼㒳䅵ᄺᮍ䴶≵᳝㚠᱃ⶹ䆚Ǆ‫؛‬䆒ⱘ㒳䅵Ẕ偠䳔㽕䕗໮ⱘ⧚䆎ⶹ䆚Ǆᴀゴᦤկњҟ㒡ᗻⱘ㓐
䗄ˈ䲚Ё䅼䆎䙷ѯϢ‫؛‬䆒ⱘᄺдǃ䆘Ԅ੠↨䕗Ⳍ݇ⱘ䯂乬Ǆ
5.2 զ䇗‫ٽ‬䇴㋴ᓜ
೼䆘ԄϔϾ‫؛‬䆒ᯊˈ៥Ӏϔ㠀᳈ᛳ݈䍷ѢԄ䅵݊ᇍ᳾ᴹᅲ՟ⱘߚ㉏㊒ᑺǄৠᯊˈг䳔㽕
ⶹ䘧䖭ϔ㊒ᑺԄ䅵Ёⱘৃ㛑ⱘ䇃Ꮒ˄ेϢℸԄ䅵Ⳍ㘨㋏ⱘ䇃Ꮒ䮼䰤˅Ǆ
ᴀゴՓ⫼ⱘᄺд䯂乬ⱘḚᶊབϟǄ᳝ϔ᠔᳝ৃ㛑ᅲ՟ⱘぎ䯈 X˄བ᠔᳝Ҏⱘ䲚ড়˅ ˈ݊
ϞᅮНњ໮ϾⳂᷛߑ᭄˄བ䅵ߦᴀᑈ䌁ф⒥䲾ᵓ㗙˅Ǆ៥Ӏ‫؛‬ᅮ X Ёϡৠᅲ՟‫݋‬᳝ϡৠⱘߎ
⦄乥⥛ˈᇍℸˈϔ⾡ড়䗖ⱘᓎ῵ᮍᓣᰃˈ‫؛‬ᅮᄬ೼ϔ᳾ⶹⱘὖ⥛ߚᏗ DˈᅗᅮНњ X Ё↣ϔ
ᅲ՟ߎ⦄ⱘὖ⥛˄བ 19 ቕⱘҎⱘὖ⥛↨ 109 ቕⱘҎὖ⥛催˅
Ǆ⊼ᛣ D ᑊ≵᳝䇈ᯢ x ᰃϔℷ՟
䖬ᰃϔড՟ˈা⹂ᅮњ݊ߎ⦄ὖ⥛Ǆᄺдӏࡵᰃ೼‫؛‬䆒ぎ䯈 H ϞᄺдϔϾⳂᷛὖᗉ˄ेⳂ
ᷛߑ᭄˅fǄⳂᷛߑ᭄ f ⱘ䆁㒗ḋ՟⬅ᮑᬭ㗙ᦤկ㒭ᄺд఼˖↣ϔϾᅲ՟ᣝ✻ߚᏗ D 㹿⣀ゟ
ഄᢑপˈ✊ৢᅗ䖲ৠ݊ℷ⹂ⱘⳂᷛؐ f(x)㹿ᦤկ㒭ᄺд఼Ǆ
Ў䇈ᯢ䖭ϔ⚍ˈ㗗㰥Ⳃᷛߑ᭄Ā䅵ߦᴀᑈ䌁ф⒥䲾ᵓ㗙āˈৃҹ䇗ᶹএ⒥䲾ᵓଚᑫⱘ乒
ᅶˈ䗮䖛ℸ䇗ᶹᴹᬊ䲚䆁㒗ḋ՟Ǆ೼䖭䞠ᅲ՟ぎ䯈 X Ў᠔᳝Ҏ㒘៤ⱘ䲚ড়ˈ↣Ͼᅲ՟ৃ⬅
Ҏⱘ৘⾡ሲᗻᦣ䗄ˈབᑈ啘ǃ㘠Ϯǃ↣ᑈ⒥䲾⃵᭄ㄝǄߚᏗᚙ‫ މ‬D ᣛᅮњ೼⒥䲾ᵓଚᑫЁ
䘛ࠄⱘ↣ϾҎⱘὖ⥛ǄⳂᷛߑ᭄ f:Xė{0,1}ᇚ↣ϾҎ䖯㸠ߚ㉏ˈ߸ᮁᅗᰃ৺Ӯ೼ᴀᑈ‫ݙ‬䌁ф
⒥䲾ᵓǄ
೼䖭Ͼϔ㠀ⱘḚᶊЁˈ៥Ӏᛳ݈䍷ⱘᰃҹϟϸϾ䯂乬˖
1. 㒭ᅮ‫؛‬䆒 h ੠ࣙ৿㢹ᑆᣝ D ߚᏗ䱣ᴎᢑপⱘḋ՟ⱘ᭄᥂䲚ˈབԩ䩜ᇍᇚᴹᣝৠḋߚᏗᢑ

পⱘᅲ՟ˈᕫࠄᇍ h ⱘ㊒ᑺⱘ᳔དԄ䅵Ǆ
2. 䖭ϔ㊒ᑺԄ䅵ⱘৃ㛑ⱘ䇃Ꮒᰃ໮ᇥ˛
5.2.1 ṭᵢ䭏䈥⦽ૂⵕᇔ䭏䈥⦽
Ў㾷‫އ‬Ϟ䗄ⱘϸϾ䯂乬ˈ䳔㽕⹂ߛഄऎߚߎϸ⾡㊒ᑺ˄៪ϸ⾡䫭䇃⥛˅Ǆ݊ϔᰃৃ⫼᭄
᥂ḋᴀϞ䆹‫؛‬䆒ⱘ䫭䇃⥛Ǆ݊Ѡᰃ೼ߚᏗЎ D ⱘᭈϾᅲ՟䲚ড়Ϟ䆹‫؛‬䆒ⱘ䫭䇃⥛ǄᅗӀߚ
߿㹿⿄Ўḋᴀ䫭䇃⥛੠ⳳᅲ䫭䇃⥛Ǆ
ᇍѢѢҢ X Ёᢑপⱘḋᴀ Sˈᶤ‫؛‬䆒݇Ѣ S ⱘḋ

ḋᴀ䫭䇃⥛˄sample error˅
ˈᰃ䆹‫؛‬䆒䫭
䇃ߚ㉏ⱘᅲ՟೼ S Ё᠔ऴ↨՟˖
ḋᴀ䫭䇃⥛˄ᷛ䆄ЎerrorS(h)˅Ў˖
定义： ‫؛‬䆒h݇ѢⳂᷛߑ᭄f੠᭄᥂ḋᴀSⱘḋ
1
errorS (h) { ¦ G ( f ( x), h( x))
n xS
݊Ё n Ў S Ёḋ՟ⱘ᭄䞣ˈ㗠 G ( f ( x ), h( x )) ೼ f ( x ) z h( x ) ᯊЎ 1ˈ৺߭Ў 0Ǆ
ⳳᅲ䫭䇃⥛(true error)ᰃᇍѢᣝ D ߚᏗ䱣ᴎᢑপⱘᅲ՟ˈ䆹‫؛‬䆒ᇍᅗ䫭䇃ߚ㉏ⱘὖ⥛Ǆ
ⳳᅲ䫭䇃⥛˄⬅errorD(h)㸼⼎˅ˈЎh䇃ߚ㉏ᣝD
定义： ‫؛‬䆒h݇ѢⳂᷛߑ᭄f੠ߚᏗDⱘⳳ
ߚᏗ䱣ᴎᢑপᅲ՟ⱘὖ⥛˖
errorD ( h) { Pr [ f ( x ) z h( x )]
xD
䖭䞠ˈ䆄ো Pr 㸼⼎ὖ⥛೼ᅲ՟ߚᏗ D Ϟ䅵ㅫǄ

xD
៥Ӏ䗮ᐌᛇⶹ䘧ⱘᰃ‫؛‬䆒ⱘⳳᅲ䫭䇃⥛errorD(h)ˈ಴Ў䖭ᰃ೼ߚ㉏᳾ᴹḋ՟ᯊߎ⦄䫭䇃
ⱘৃ㛑ᗻǄ✊㗠៥Ӏ᠔㛑⌟䞣ⱘাᰃḋᴀ䫭䇃⥛errorS(h)ˈᅗ᠔㽕∖ⱘ᭄᥂ḋᴀSᰃ៥Ӏ᠔
ᢹ᳝ⱘǄᴀ㡖᠔㽕㗗㰥ⱘЏ㽕䯂乬ህᰃĀerrorS(h)೼ԩ⾡⿟ᑺϞᦤկњᇍerrorD(h) ⱘԄ
䅵˛āǄ
5.2.2 ⿱ᮙ‫ٽٲ‬䇴Ⲻ㖤‫ؗ‬॰䰪
Ў㾷‫އ‬ĀerrorS(h)೼ԩ⾡⿟ᑺϞᦤկњᇍerrorD(h) ⱘԄ䅵āⱘ䯂乬ˈ‫ܜ‬㗗㰥hЎ⾏ᬷؐ
‫؛‬䆒ⱘᚙ‫މ‬Ǆ‫݋‬ԧഄ䇈ˈ↨བ㽕෎Ѣᶤ⾏ᬷؐ‫؛‬䆒h೼ḋᴀSϞ㾖ᆳࠄⱘḋᴀ䫭䇃⥛ˈԄ䅵ᅗ
ⱘⳳᅲ䫭䇃⥛ˈ݊Ё˖
z ḋᴀ S ࣙ৿ n Ͼḋ՟ˈᅗӀⱘᢑপᣝ✻ὖ⥛ߚᏗ Dˈᢑপ䖛⿟ᰃⳌѦ⣀ゟⱘˈᑊ
Ϩϡձ䌪Ѣ h
z nı30
z ‫؛‬䆒h೼䖭nϾḋ՟Ϟ⢃њrϾ䫭䇃˄errorS(h)=r/n˅
Ꮖⶹ䖭ѯᴵӊˈ㒳䅵⧚䆎ৃ㒭ߎҹϟᮁ㿔˖
1.≵᳝݊Ҫֵᙃⱘ䆱ˈerrorD(h)᳔ৃ㛑ⱘؐЎerrorS(h)
2.᳝໻㑺 95ˁⱘৃ㛑ᗻˈⳳᅲ䫭䇃⥛errorD(h)໘Ѣϟ䴶ⱘऎ䯈‫˖ݙ‬
errorS (h)(1 errorS (h))

errorS (h) r 1.96
n
В՟䇈ᯢˈ‫؛‬བ᭄᥂ḋᴀSࣙ৿n=40 Ͼḋ՟ˈᑊϨ‫؛‬䆒h೼䖭ѯ᭄᥂Ϟѻ⫳њr=12 Ͼ䫭
䇃Ǆ䖭ḋˈḋᴀ䫭䇃⥛ЎerrorS(h)=12/40˙0.3Ǆབᵰ≵᳝᳈໮ⱘֵᙃˈᇍⳳᅲ䫭䇃⥛errorD(h)
ⱘ᳔དⱘԄ䅵ेЎḋᴀ䫭䇃⥛ 0.3Ǆ✊㗠៥Ӏϡ㛑ᳳᳯ䖭ᰃᇍⳳᅲ䫭䇃⥛ⱘᅠ㕢Ԅ䅵Ǆབᵰ
঺໪᧰䲚 40 Ͼ䱣ᴎᢑপⱘḋ՟S´ˈḋᴀ䫭䇃⥛errorS´(h)ᇚϢॳᴹⱘerrorS(h)ᄬ೼ѯ䆌ϡৠǄ
䖭⾡ϡৠᰃ⬅S´੠S㒘៤Ϟⱘ䱣ᴎᏂᓖ᠔ѻ⫳ⱘǄᅲ䰙Ϟˈབᵰϡᮁ䞡໡䖭ϔᅲ偠ˈ↣⃵ᢑ
পϔϾࣙ৿ 40 ḋ՟ⱘḋᴀSiˈᇚӮথ⦄㑺 95%ⱘᅲ偠Ё䅵ㅫ᠔ᕫⱘऎ䯈ࣙ৿ⳳᅲ䫭䇃⥛Ǆ
಴ℸˈ៥Ӏᇚℸऎ䯈⿄ЎerrorD(h)ⱘ 95%㕂ֵऎ䯈Ԅ䅵Ǆ೼ᴀ՟Ёˈr=12 ੠n=40ˈḍ᥂Ϟᓣˈ
95%㕂ֵऎ䯈Ў 0.30f(1.96h0.07)˙0.30f0.14Ǆ
Ϟ䴶ⱘ 95ˁ㕂ֵऎ䯈㸼䖒ᓣৃ᥼ᑓࠄϔ㠀ᚙᔶҹ䅵ㅫӏᛣ㕂ֵᑺǄᐌ᭄ 1.96 ᰃ⬅ 95%

䖭ϔ㕂ֵᑺ⹂ᅮⱘǄᅮНzNЎ䅵ㅫN%㕂ֵऎ䯈ᯊⱘᐌ᭄Ǆ䅵ㅫerrorD(h)ⱘN%㕂ֵऎ䯈ⱘϔ
㠀㸼䖒ᓣЎ˖

errorS (h) r z N (5.1)
n
݊ЁzNⱘؐձ䌪Ѣ᠔䳔ⱘ㕂ֵᑺˈখ㾕㸼 5-1 ЁⱘপؐǄ
㸼 5-1 ঠջⱘNˁ㕂ֵऎ䯈ⱘzNؐ
㕂ֵᑺ N% 50% 68% 80% 90% 95% 98% 99%

ᐌ䞣zN 0.67 1.00 1.28 1.64 1.96 2.33 2.58
಴ℸˈℷབerrorD(h)ⱘ 95ˁ㕂ֵऎ䯈Ў 0.30f(1.96噝 0.07)˄݊Ёr=12ˈn=40˅ˈৃҹ∖

ᕫৠḋᚙ‫މ‬ϟ 68%㕂ֵऎ䯈Ў 0.30f˄1.0噝 0.07˅ǄҢⳈ㾝Ϟ៥Ӏгৃҹⳟߎ 68ˁ㕂ֵऎ䯈
㽕ᇣѢ 95ˁ㕂ֵऎ䯈ˈ಴Ў៥Ӏ‫ޣ‬ᇣњ㽕∖errorD(h)㨑ܹℸऎ䯈ⱘὖ⥛Ǆ
ㄝᓣ 5.1 ᦣ䗄њЎњ೼errorS(h)෎⸔ϞԄ䅵errorD(h)ˈབԩ䅵ㅫ㕂ֵऎ䯈˄े䇃Ꮒ䮼䰤˅Ǆ
䖭ϔ㸼䖒ᓣᯊা㛑ᑨ⫼Ѣ⾏ᬷؐ‫؛‬䆒Ǆᅗ‫؛‬ᅮḋᴀSᢑপⱘߚᏗϢᇚᴹⱘ᭄᥂ᢑপⱘߚᏗⳌ
ৠˈᑊϨ‫؛‬ᅮ᭄᥂ϡձ䌪Ѣ᠔Ẕ偠ⱘ‫؛‬䆒Ǆ䖬᳝ˈ䆹㸼䖒ᓣাᦤկњ䖥Ԑⱘ㕂ֵऎ䯈ˈϡ䖛
䖭ϔ䖥Ԑ೼㟇ᇥࣙ৿ 30 Ͼḋ՟ᑊϨerrorS(h)ϡ໾䴴䖥 0 ៪ 1 ᯊᕜ᥹䖥ⳳᅲᚙ‫މ‬Ǆ߸ᮁ䖭⾡
䖥Ԑᰃ৺᥹䖥ⳳᅲˈ᳈㊒⹂ⱘ㾘߭Ў˖
n errorS (h)(1 errorS (h)) t 5
Ϟ䴶៥Ӏὖ䗄њ䅵ㅫ⾏ᬷؐ‫؛‬䆒ⱘ㕂ֵऎ䯈ⱘ䖛⿟ˈϟϔ㡖ᇚ㒭ߎ䖭ϔ䖛⿟ⱘ㒳䅵ᄺ෎
⸔Ǆ
5.3 䟽ṭ⨼䇰ะ⹶
ᴀ㡖ҟ㒡њ㒳䅵ᄺ੠䞛ḋ⧚䆎ⱘ޴Ͼ෎ᴀὖᗉˈࣙᣀὖ⥛ߚᏗǃᳳᳯؐǃᮍᏂǃѠ乍ߚ
Ꮧ੠ℷᗕߚᏗǃҹঞঠջ੠ऩջऎ䯈ǄᇍѢ䖭ѯὖᗉⱘ෎ᴀњ㾷ᇚ᳝ࡽѢ⧚㾷‫؛‬䆒䆘Ԅ੠ㅫ
⊩䆘ԄǄ᳈Ў䞡㽕ⱘˈᅗӀᦤկњϔ⾡䞡㽕ⱘὖᗉḚᶊˈҹ֓Ѣ⧚㾷Ⳍ݇ⱘᴎ఼ᄺд䯂乬˄བ
䖛ᑺᢳড়䯂乬˅ҹঞ⧚㾷೼៤ࡳⱘ⊯࣪੠䆁㒗ḋ՟᭄ⳂП䯈ⱘ݇㋏ǄᏆ㒣❳ᙝ䖭ѯὖᗉⱘ䇏
㗙ৃҹ䏇䖛ᴀ㡖Ǆ݊Ёҟ㒡ⱘ݇䬂ὖᗉ೼㸼 5-2 Ё߫ߎǄ
㸼 5-2 㒳䅵ᄺЁⱘ෎ᴀᅮН੠ὖᗉ
䱣ᴎব䞣(random variable)ৃ㹿ⳟ԰ᰃ᳝ὖ⥛䕧ߎⱘϔϾᅲ偠ǄᅗⱘؐЎᅲ偠ⱘ䕧ߎ㒧ᵰǄ
ᶤ䱣ᴎব䞣ⱘὖ⥛ߚᏗ(probability distribution)ᣛᅮњYপؐЎӏϔৃ㛑ⱘؐyiⱘৃ㛑ᗻPr(Y=yi)
䱣ᴎব䞣Yⱘᳳᳯؐ(expected value)៪ഛؐ(mean)Ў ¦y i i Pr(Y y i ) Ǆ䗮ᐌ⫼ヺোYᴹ㸼⼎

E[Y]Ǆ
䱣ᴎব䞣ⱘᮍᏂ(Variance)ЎVar(Y)=E[(Y-Y)2]Ǆᅗᦣ䗄њY݇Ѣ݊ഛؐߚᏗⱘᆑᑺ៪ߚᬷᑺǄ
Yⱘᷛ‫ޚ‬Ꮒ(Standard deviation)Ў Var (Y ) Ǆ䗮ᐌ⫼ヺো³Yᴹҷ㸼Ǆ

Ѡ乍ߚᏗ(Binomial distribution)ᰃ೼⹀Ꮥᡩᦋ䯂乬Ёˈ㢹ߎ⦄ℷ䴶ⱘὖ⥛Ў pˈ䙷М೼ n Ͼ⣀ゟ
ⱘᅲ偠Ёߎ⦄ r ⃵ℷ䴶ⱘߚᏗᚙ‫މ‬Ǆ
ℷᗕߚᏗ(Normal distribution)ᰃϔ䩳ᔶⱘὖ⥛ߚᏗˈᅗ೼䆌໮㞾✊⦄䈵Ё䛑Ӯߎ⦄Ǆ
Ёᖗᵕ䰤ᅮ⧚(Central Limit Theorem)䇈ᯢ⣀ゟৠߚᏗⱘ䱣ᴎব䞣೼໻䞣ᅲ偠Ёⱘ੠䙉ᕾℷᗕߚ
ᏗǄ
Ԅ䅵䞣(estimator)Ўϔ䱣ᴎব䞣 Yˈᅗ⫼ᴹԄ䅵ϔ෎‫ޚ‬ᘏԧⱘᶤϔখ᭄ pǄ
P ⱘԄ䅵䞣 Y ⱘԄ䅵‫أ‬Ꮒ(estimation bias)Ў(E[Y]ˉp)Ǆ᮴‫أ‬Ԅ䅵䞣ᰃᣛ䆹‫أ‬ᏂЎ 0Ǆ
N%㕂ֵऎ䯈(confidence interval)⫼ѢԄ䅵খ᭄ pˈ䆹ऎ䯈ҹ N%ⱘὖ⥛ࣙ৿ pǄ
5.3.1 䭏䈥⦽զ䇗ૂӂ亯∊ׁզ䇗
೼ḋᴀ䫭䇃⥛੠ⳳᅲ䫭䇃⥛П䯈ⱘᏂᓖϢ᭄᥂ḋᴀ໻ᇣⱘձ䌪݇㋏བԩ˛䖭ϔ䯂乬೼
㒳䅵ᄺЁᏆ䗣ᕏⷨおǄᅗৃ㸼䗄Ў˖㒭ᅮϔᘏԧЁ䱣ᴎᢑপⱘ䚼ߚḋᴀⱘ㾖ᆳ乥⥛ˈԄ䅵ᭈ
Ͼᘏԧⱘὖ⥛Ǆ೼䖭䞠ˈ៥Ӏᛳ݈䍷ⱘ㾖ᆳ䞣Ў h ᰃ৺䇃ߚ㉏ḋ՟Ǆ
㾷‫އ‬䆹䯂乬佪‫ܜ‬㽕⊼ᛣࠄˈ⌟䞣ḋᴀ䫭䇃⥛ⳌᔧѢ೼԰ϔϾ᳝䱣ᴎ䕧ߎⱘᅲ偠Ǆ៥Ӏ‫ܜ‬
ҢߚᏗDЁ䱣ᴎᢑপߎnϾ⣀ゟⱘᅲ՟ˈᔶ៤ḋᴀSˈ✊ৢ⌟䞣ḋᴀ䫭䇃⥛errorS(h)ˈབࠡϔ
㡖᠔䗄ˈབᵰᇚᅲ偠䞡໡໮⃵ˈ↣⃵ᢑপ໻ᇣЎnⱘϡৠⱘḋᴀSiˈᇚৃҹᕫࠄϡৠⱘerrorSi(h)
ⱘؐˈᅗপ‫އ‬ѢϡৠSiⱘ㒘៤Ёⱘ䱣ᴎᏂᓖǄ䖭⾡ᚙ‫މ‬ϟˈ㄀iϾ䖭ḋⱘᅲ偠ⱘ䕧ߎerrorSi(h)
㹿⿄Ўϔ䱣ᴎব䞣˄random variable˅
Ǆϔ㠀ᚙ‫މ‬ϟˈৃҹᇚ䱣ᴎব䞣ⳟ៤ϔϾ᳝䱣ᴎ䕧ߎⱘ
ᅲ偠Ǆ䱣ᴎব䞣ؐेЎ䱣ᴎᅲ偠ⱘ㾖ᆳ䕧ߎǄ
䆒ᛇ㽕䖤㸠kϾ䖭ḋⱘ䱣ᴎᅲ偠ˈ⌟䞣䱣ᴎব䞣errorS1(h) ˈerrorS2(h)ˈĂĂˈ errorSk(h)Ǆ

✊ৢ៥Ӏҹ೒㸼ⱘᔶᓣᰒ⼎ߎ㾖ᆳࠄⱘ↣Ͼ䫭䇃⥛ؐⱘ乥⥛Ǆᔧkϡᮁ๲䭓ˈ䆹೒㸼ᇚਜ⦄
བ㸼 5-3 䙷ḋⱘߚᏗǄ䆹㸼ᦣ䗄ⱘὖ⥛ߚᏗ⿄ЎѠ乍ߚᏗ˄Binomial distribution˅Ǆ
㸼 5-3 Ѡ乍ߚᏗ
Binomial distribution for n=40, p=0.3: n=40ˈp=0.3 ᯊⱘѠ乍ߚᏗ
ϔϾѠ乍ߚᏗ˄Binomial distribution˅㒭ߎњᔧऩϾ⹀Ꮥᡩᦋߎ⦄ℷ䴶ⱘὖ⥛Ў p ᯊˈ೼ n Ͼ⣀ゟ⹀Ꮥ

ᡩᦋⱘḋᴀЁ㾖ᆳࠄ r ⃵ℷ䴶ⱘὖ⥛Ǆᅗ⬅ҹϟⱘὖ⥛ߑ᭄ᅮН˖
n!
P(r ) p r (1 p) n r
r!(n r )!
བᵰ䱣ᴎব䞣 X 䙉ᕾѠ乍ߚᏗˈ߭˖
X পؐЎ r ⱘὖ⥛ Pr(X=r)⬅ P(r)㒭ߎǄ
X ⱘᳳᳯؐ៪ഛؐ E[X]Ў˖
E[X]=np
X ⱘᮍᏂ Var(X)Ў
Var(X)=np(1-p)
Xⱘᷛ‫ޚ‬ᏂЎ³XЎ˖
VX np(1 p)
ᇍѢ䎇໳໻ⱘ n ؐˈѠ乍ߚᏗᕜ᥹䖥Ѣ᳝ৠḋഛؐ੠ᮍᏂⱘℷᗕߚᏗ˄㾕㸼 5-4˅Ǆ໮᭄㒳䅵ᄺᆊᓎ䆂
া೼ np(1-p)ı5 ᯊՓ⫼ℷᗕߚᏗᴹ䖥ԐѠ乍ߚᏗǄ
5.3.2 ӂ亯࠼ᐹ
Ў䕗དഄ⧚㾷Ѡ乍ߚᏗˈ㗗㰥ҹϟⱘ䯂乬Ǆ᳝ϔ⺼ᤳᑊᔃ᳆њⱘ⹀Ꮥˈ㽕Ԅ䅵೼ᡯ⹀Ꮥ
ᯊߎ⦄ℷ䴶ⱘὖ⥛ǄҸℸ᳾ⶹὖ⥛Ў pǄᡩᦋ䆹⹀Ꮥ n ⃵ᑊ䅵ㅫߎ⦄ℷ䴶ⱘ⃵᭄ rǄᇍѢ p
ⱘϔড়⧚ⱘԄ䅵Ў r/nǄ⊼ᛣˈབᵰ䞡ᮄ䖯㸠ϔ⃵䆹ᅲ偠ˈ⫳៤ϔϾᮄⱘ n ⃵ᡯ⹀Ꮥⱘ䲚ড়ˈ
݊ߎ⦄ℷ䴶⃵᭄ r ᇚϢ㄀ϔ⃵ᅲ偠᳝⿡䆌ϡৠˈҢ㗠ᕫࠄᇍ p ⱘ঺ϔϾԄ䅵ǄѠ乍ߚᏗᦣ䗄
ⱘᰃᇍӏϔৃ㛑ⱘ r ؐ˄Ң 0 ࠄ n˅ˈ䖭Ͼℷ䴶ὖ⥛Ў p ⱘ⹀Ꮥᡯᦋ n ⃵ᙄདߎ⦄ r ⃵ℷ䴶ⱘ
ὖ⥛Ǆ
᳝䍷ⱘᰃˈҢᡯᦋ⹀Ꮥⱘ䱣ᴎḋᴀЁԄ䅵pˈϢ೼ᅲ՟ⱘ䱣ᴎḋᴀϞ⌟䆩hҹԄ䅵errorD(h)
ᰃⳌৠⱘ䯂乬Ǆϔ⃵⹀ᏕᡯᦋᇍᑨѢҢDЁᢑপϔϾᅲ՟ᑊ⌟䆩ᅗᰃ৺㹿h䇃ߚ㉏Ǆϔ⃵䱣
ᴎᡯᦋߎ⦄ℷ䴶ⱘὖ⥛pᇍᑨѢ䱣ᴎᢑপⱘᅲ՟㹿䇃ߚ㉏ⱘὖ⥛˄ेpᇍᑨerrorD(h)˅ Ǆn⃵ᡯ
ᦋⱘḋᴀ㾖ᆳࠄr⃵ℷ䴶ˈᇍᑨnϾᢑপⱘᅲ՟㹿䇃ߚ㉏ⱘ᭄ⳂǄ಴ℸr/nᇍᑨerrorS(h)ǄԄ䅵
pⱘ䯂乬ㄝᬜѢԄ䅵errorD(h)ǄѠ乍ߚᏗ㒭ߎњϔϾϔ㠀ᔶᓣⱘὖ⥛ߚᏗˈ᮴䆎⫼Ѣ㸼⼎n⃵
⹀Ꮥߎ⦄ℷ䴶ⱘ⃵᭄䖬ᰃ೼nϾḋ՟Ё‫؛‬䆒ߎ䫭ⱘ⃵᭄ǄѠ乍ߚᏗⱘ䆺㒚ᔶᓣձ䌪Ѣḋᴀ໻
ᇣnҹঞὖ⥛p៪errorD(h)Ǆ
ϔ㠀ᴹ䇈ᑨ⫼Ѡ乍ߚᏗⱘᴵӊࣙᣀ˖
1.᳝ϔ෎‫ޚ‬ᅲ偠˄བᡩᦋ⹀Ꮥ˅ˈ݊䕧ߎৃ㹿ᦣ䗄Ўϔ䱣ᴎব䞣 YǄ䱣ᴎব䞣 Y ᳝ϸ⾡

পؐ˄བ Y˙1 Ўℷ䴶ˈY˙0 ড䴶˅Ǆ
2.೼ᅲ偠ⱘӏϔ⃵ᇱ䆩Ё Y=1 ⱘὖ⥛Ўᐌ᭄ pǄᅗϢ݊Ҫⱘᅲ偠ᇱ䆩᮴݇Ǆ಴ℸ Y=0 ⱘ

ὖ⥛Ў 1-pǄϔ㠀 p Ў乘‫ܜ‬᳾ⶹⱘˈ䴶Јⱘ䯂乬ህ೼ѢԄ䅵ᅗǄ
3.෎‫ޚ‬ᅲ偠ⱘn⃵⣀ゟᇱ䆩ᣝᑣ߫ᠻ㸠ˈ⫳៤ϔϾ⣀ゟৠߚᏗⱘ䱣ᴎব䞣ᑣ߫Y1ˈY2ˈĂĂ
YnǄҸRҷ㸼n⃵䆩偠Ёߎ⦄Yi˙1 ⱘ⃵᭄˖
n
R { ¦ Yi
i 1
4.䱣ᴎব䞣 R প⡍ᅮؐ r ⱘὖ⥛˄བ㾖ᆳࠄ r ⃵ℷ䴶ⱘὖ⥛˅⬅Ѡ乍ߚᏗ㒭ߎ˖
n!
Pr( R r) p r (1 p) n r ˄5.2˅
r!(n r )!
ℸὖ⥛ߚᏗⱘϔϾ೒㸼೼㸼 5-3 Ё㒭ߎǄ
Ѡ乍ߚᏗࠏ⬏њ n ⃵⹀Ꮥᡩᦋߎ⦄ r ⃵ℷ䴶ⱘὖ⥛ˈгࠏ⬏њࣙ৿ n Ͼ䱣ᴎḋ՟ⱘ᭄᥂

ḋᴀߎ⦄ r ⃵䇃ߚ㉏䫭䇃ⱘὖ⥛Ǆ
5.3.3 ൽ‫ૂٲ‬ᯯᐤ
䱣ᴎব䞣ⱘϸϾ᳔ᐌ⫼ࠄⱘሲᗻЎ݊ᳳᳯؐ˄г⿄Ўഛؐ˅੠ᮍᏂǄᳳᳯؐᰃ䞡໡䞛ḋ
䱣ᴎব䞣ᕫࠄⱘؐⱘᑇഛǄ᳈㊒⹂ⱘᅮНབϟ˖
定义：㗗㰥䱣ᴎব䞣Yৃ㛑ⱘপؐЎy1…ynˈYⱘᳳ
ᳳᳯؐ(expected value)E(Y)Ў˖
n
E[Y ] { ¦ y i Pr(Y yi ) ˄5.3˅
i 1
՟བˈབᵰ Y পؐ 1 ⱘὖ⥛Ў 0.7ˈপؐ 2 ⱘὖ⥛ 0.3ˈ䙷МᳳᳯؐЎ˄1g0.7ˇ2g0.3

˙1.3˅Ǆབᵰ䱣ᴎব䞣 Y ᳡ҢѠ乍ߚᏗˈ䙷Мৃᕫ˖
E[Y ] np ˄5.4˅
݊Ё n ੠ p Ўᓣ 5.2 ЁᅮНⱘѠ乍ߚᏗⱘখ᭄Ǆ
঺ϔ䞡㽕ሲᗻᮍ
ᮍᏂᦣ䗄ⱘᰃὖ⥛ߚᏗⱘᆑᑺ៪ᬷᑺˈेᅗᦣ䗄њ䱣ᴎব䞣Ϣ݊ഛؐП䯈
ⱘᏂ᳝໮໻Ǆ
定义：䱣ᴎব䞣 Y ⱘᮍ
ᮍᏂ˄variance˅Var[Y]Ў˖
Var[Y ] { E[(Y E[Y ]) 2 ] ˄5.5˅
ᮍᏂᦣ䗄ⱘᰃҢYⱘϔϾ㾖ᆳএԄ䅵݊ഛؐE(Y)ⱘ䇃ᏂᑇᮍⱘᳳᳯǄᮍᏂⱘᑇᮍḍ㹿⿄
ЎYⱘᷛ‫ޚ‬Ꮒˈ䆄Ў³YǄ
ᷛ‫ޚ‬Ꮒ˄standard deviation˅³YЎ˖
定义：䱣ᴎব䞣Yⱘᷛ
V Y { E[(Y E[Y ]) 2 ] ˄5.6˅
㢹䱣ᴎব䞣 Y ᳡ҢѠ乍ߚᏗˈ߭ᮍᏂ੠ᷛ‫ޚ‬Ꮒߚ߿Ў˖
Var[Y ] { np (1 p )
V Y { np(1 p) ˄5.7˅
5.3.4 զ䇗䠅θ‫څ‬ᐤૂᯯᐤ
៥ӀᏆᕫߎ䱣ᴎব䞣errorS(h)᳡ҢѠ乍ߚᏗˈ⦄೼ಲࠄࠡ䴶ⱘ䯂乬˖errorS(h)੠ⳳᅲ䫭
䇃⥛errorD(h)П䯈ৃ㛑ⱘᏂᓖᰃ໮ᇥ˛
⫼ᓣ 5.2 ЁѠ乍ߚᏗⱘᅮНᴹᦣ䗄errorS(h)੠errorD(h)ˈৃᕫ˖
r
errorS (h)
n
errorD (h) p
݊Ё n Ўḋᴀ S Ёᅲ՟᭄ˈr ᰃ S Ё㹿 h 䇃ߚ㉏ⱘᅲ՟᭄ˈp ЎҢ D Ёᢑপϔᅲ՟㹿䇃

ߚ㉏ⱘὖ⥛Ǆ
㒳䅵ᄺЁᇚerrorS(h)⿄Ўⳳᅲ䫭䇃⥛errorD(h)ⱘϔϾԄ䅵䞣 (estimator)Ǆϔ㠀ഄˈԄ䅵
䞣ᰃ⫼ᴹԄ䅵ᶤ෎‫ޚ‬ᘏԧⱘᶤϔখ᭄ⱘ䱣ᴎব䞣ǄᇍѢԄ䅵䞣ˈᰒ✊᳔݇ᖗⱘᰃᅗᑇഛᴹ䇈
ᰃ৺㛑ѻ⫳ℷ⹂Ԅ䅵Ǆϟ䴶ᅮНԄ䅵‫أ‬Ꮒ(estimation bias)ЎԄ䅵䞣ⱘᳳᳯؐৠⳳᅲখ᭄ؐП
䯈ⱘᏂᓖǄ
定义：䩜ᇍӏᛣখ᭄ p ⱘԄ䅵䞣 Y ⱘԄ

Ԅ䅵‫أ‬ᏂЎ˖
E[Y]-p
བᵰԄ䅵‫أ‬ᏂЎ 0ˈ៥Ӏ⿄ Y Ў p ⱘ᮴‫أ‬Ԅ䅵䞣˄unbiased estimator˅Ǆ⊼ᛣˈ೼ℸᚙ‫މ‬

ϟ⬅໮⃵䞡໡ᅲ偠⫳៤ⱘ Y ⱘ໮Ͼ䱣ᴎؐⱘᑇഛ˄े E[Y]˅ᇚᬊᬯѢ pǄ
errorS(h)ᰃ৺ЎerrorD(h)ⱘϔϾ᮴‫أ‬Ԅ䅵䞣˛⹂ᅲབℸˈ಴ЎᇍѢѠ乍ߚᏗˈrⱘᳳᳯؐ
Ўnp˄ㄝᓣ[5.4]˅Ǆ⬅ℸˈᑊϨ಴ЎnЎϔᐌ᭄ˈ䙷Мr/nⱘᳳᳯؐЎpǄ
ᇍԄ䅵‫أ‬Ꮒ䖬䳔㽕԰ϸ⚍䇈ᯢǄ佪‫ˈܜ‬೼ᴀゴᓔྟ៥Ӏᦤࠄˈ೼䆁㒗ḋ՟Ϟ⌟䆩‫؛‬䆒ᕫ
ࠄⱘᇍ‫؛‬䆒䫭䇃⥛ⱘԄ䅵‫أ‬ѢФ㾖࣪ˈ᠔ᣛⱘℷᰃԄ䅵‫أ‬ᏂǄ㽕ՓerrorS(h)ᇍerrorD(h)᮴‫أ‬
Ԅ䅵ˈ‫؛‬䆒h੠ḋᴀSᖙ乏⣀ゟ䗝পǄ㄀ѠˈԄ䅵‫أ‬Ꮒ(estimation bias)䖭ϔὖᗉϡ㛑Ϣ㄀Ѡゴ
ҟ㒡ⱘᄺд఼ⱘᔦ㒇‫أ‬㕂(inductive bias)Ⳍ⏋⎚ǄԄ䅵‫أ‬ᏂЎϔ᭄ᄫ䞣ˈ㗠ᔦ㒇‫أ‬㕂Ўᮁ㿔
䲚ড়Ǆ
Ԅ䅵䞣ⱘ঺ϔ䞡㽕ሲᗻЎ݊ᮍᏂǄ㒭ᅮ໮Ͼ᮴‫أ‬Ԅ䅵䞣ˈⳈ㾖Ϟᑨ䗝প݊ЁᮍᏂ᳔ᇣⱘǄ
⬅ᮍᏂⱘᅮНˈ᠔䗝ᢽⱘᑨЎখ᭄ؐ੠Ԅ䅵ؐП䯈ᳳᳯᑇᮍ䇃Ꮒ᳔ᇣⱘǄ
‫؛‬བ೼⌟䆩ϔ‫؛‬䆒ᯊˈᅗᇍn˙40 Ͼ䱣ᴎḋ՟ⱘḋᴀѻ⫳r˙12 Ͼ䫭䇃ˈ䙷МᇍerrorD(h)

ⱘ᮴‫أ‬Ԅ䅵ЎerrorS(h)˙r/n˙0.3ǄԄ䅵Ёѻ⫳ⱘᮍᏂᅠܼᴹ⑤ѢrЁⱘᮍᏂˈ಴ЎnЎϔᐌ᭄Ǆ
⬅ѢrᰃѠ乍ߚᏗˈᅗⱘᮍᏂ⬅ᓣ˄5.7˅Ўnp(1-p)Ǆ✊㗠p᳾ⶹˈ៥Ӏৃҹ⫼Ԅ䅵䞣r/nᴹҷ
᳓pǄ⬅ℸᕫߎrⱘⱘԄ䅵ᮍᏂЎ 40g0.3(1-0.3)˙8.4ˈ៪Ⳍᑨⱘᷛ‫ޚ‬Ꮒ 8.4 Ĭ2.9Ǆ䖭㸼⼎
errorS(h)˙r/nЁⱘᷛ‫ޚ‬Ꮒ㑺Ў 2.9/40˙0.07Ǆὖ㗠㿔Пˈ㾖ᆳࠄⱘerrorS(h)Ў 0.3ˈᷛ‫ޚ‬Ꮒ㑺
Ў 0.07Ǆ˄㾕д乬 5.1˅Ǆ
ϔ㠀ᴹ䇈ˈ㢹೼nϾ䱣ᴎ䗝পⱘḋᴀЁ᳝rϾ䫭䇃ˈerrorS (h)ⱘᷛ‫ޚ‬ᏂЎ˖
Vr p(1 p)
V error ( h ) ˄5.8˅
S
n n
ᅗ㑺ㄝѢ⫼r/n˙errorS (h)ᴹҷ᳓p˖

V error ( h ) | ˄5.9˅
S
n
5.3.5 㖤‫ؗ‬॰䰪
䗮ᐌᦣ䗄ᶤԄ䅵ⱘϡ⹂ᅮᗻⱘᮍ⊩ᰃՓ⫼ϔ㕂ֵऎ䯈ˈⳳᅲⱘؐҹϔᅮⱘὖ⥛㨑ܹ䆹ऎ
䯈ЁǄ䖭ḋⱘԄ䅵⿄Ў㕂
㕂ֵऎ䯈˄confidence interval˅Ԅ䅵Ǆ
㕂ֵऎ䯈ᰃϔϾҹ Nˁⱘὖ⥛ࣙ৿ p ⱘऎ䯈Ǆ

定义： ᶤϾখ᭄ p ⱘ Nˁ㕂
՟བˈབᵰ೼n˙40 Ͼ⣀ゟᢑপⱘḋ՟ⱘḋᴀЁ᳝r˙12 Ͼ䫭䇃ˈৃҹ⿄ऎ䯈 0.3f0.14

᳝ 95ˁⱘৃ㛑ᗻࣙ৿ⳳᅲ䫭䇃⥛errorD(h)Ǆ
བԩ㦋ᕫerrorD(h)ⱘ㕂ֵऎ䯈˛ㄨḜ೼ѢԄ䅵䞣errorS(h)᳡ҢѠ乍ߚᏗǄ䖭ϔߚᏗⱘഛ
ؐЎerrorD(h)ˈᷛ‫ޚ‬Ꮒৃ⬅ᓣ 5.9 䅵ㅫǄ಴ℸˈЎ䅵ㅫ 95ˁ㕂ֵऎ䯈ˈা䳔㽕ᡒࠄϔϾҹഛ
ؐerrorD(h)ЎЁᖗⱘऎ䯈ˈᅗⱘᆑᑺ䎇ҹࣙ৿䆹ߚᏗϟܼ䚼ὖ⥛ⱘ 95ˁǄㄝӋഄˈᅗᣛᅮњ
errorD(h)᳝ 95ˁⱘᴎӮ㨑ܹerrorS(h)਼ೈⱘᶤऎ䯈ⱘ໻ᇣǄ
ᇍѢ㒭ᅮⱘ N ؐˈབԩ䅵ㅫऎ䯈໻ᇣҹՓ݊ࣙ৿ Nˁⱘὖ⥛䋼䞣˛ᇍѢѠ乍ߚᏗᴹ䇈

䖭ϔ䅵ㅫकߚ⚺⧤Ǆ✊㗠໮᭄ᚙ‫މ‬ϟৃҹ䖯㸠ϔ䖥ԐˈՓ䅵ㅫ䖛⿟᳈ᆍᯧǄ䖭෎Ѣབϟџᅲ˖
ेᇍѢ䎇໳໻ⱘḋᴀˈѠ乍ߚᏗৃҹᕜདഄ⬅ℷᗕߚᏗᴹ䖥ԐǄℷᗕߚᏗ˄೼㸼 5-4 Ёὖ䗄˅
ᰃ㒳䅵ᄺЁⷨおᕫ᳔䗣ᕏⱘὖ⥛ߚᏗПϔǄབ㸼 5-4 ᠔⼎ˈℷᗕߚᏗᰃϔ䩳ᔶߚᏗˈ⬅݊ഛ
ؐ੠ᷛ‫ޚ‬Ꮒ³ᅠܼᅮНǄᇍѢ໻ⱘ nˈѠ乍ߚᏗ䴲ᐌ䖥ԐѢϔϾৠḋഛؐ੠ᮍᏂⱘℷᗕߚ
ᏗǄ
П᠔ҹՓ⫼ℷᗕߚᏗᴹҷ᳓ˈϔϾॳ಴ᰃ໮᭄㒳䅵খ㗗䛑߫㸼㒭ߎњℷᗕߚᏗϟࣙ৿N
ˁⱘὖ⥛䋼䞣ⱘഛ਼ؐೈⱘऎ䯈ⱘ໻ᇣǄ䖭ህᰃ䅵ㅫNˁ㕂ֵऎ䯈᠔䳔ⱘֵᙃǄᅲ䰙Ϟ㸼 5-1
ℷᰃ䖭ḋϔϾ㸼Ǆ㸼 5-1 Ё㒭ᅮⱘᐌ᭄zNᅮНⱘᰃ೼䩳ᔶℷᗕߚᏗϟˈࣙ৿Nˁὖ⥛䋼䞣ⱘ
ഛ਼ؐೈⱘ᳔ᇣऎ䯈ⱘᆑᑺǄ᳈㊒⹂ഄ䇈ˈzNҹᷛ‫ޚ‬Ꮒ㒭ᅮњऎ䯈ⱘञᆑᑺ˄े೼ӏϔᮍ৥
䎱ഛؐⱘ䎱⾏˅ˈ೒ 5-1(a)㒭ߎњ䩜ᇍz80ⱘϔϾऎ䯈Ǆ
㸼 5-4 ℷᗕ៪催ᮃߚᏗ
Normal distribution with mean 0, standard deviation 1˖ഛؐЎ 0ˈᷛ‫ޚ‬ᏂЎ 1 ⱘℷᗕߚᏗ
ϔϾℷᗕߚᏗ˄г㹿⿄Ў催ᮃߚᏗ˅ᰃϔ䩳ൟߚᏗˈᅗᅮНЎϟ䴶ⱘὖ⥛ᆚᑺߑ᭄˖
2
1 § xP ·
1 ¨ ¸
2© V ¹
p ( x) e
2SV 2
ϔϾℷᗕߚᏗ⬅Ϟ䴶݀ᓣЁⱘϸϾখ᭄ᅠܼ⹂ᅮ˖੠³Ǆ
བᵰ䱣ᴎব䞣 X 䙉ᕾℷᗕߚᏗˈ߭˖
X 㨑ܹࠄ(a, b)ⱘὖ⥛Ў˖
b
³ p( x)dx
a
X ⱘᳳᳯؐ៪ഛؐ E[X]Ў˖
E[X]=
X ⱘᮍᏂ Var(X)Ў
Var(X)=³2
Xⱘᷛ‫ޚ‬ᏂЎ³XЎ˖
³X=³
Ёᖗᵕ䰤ᅮ⧚˄5.4.1 㡖˅䇈ᯢ໻䞣⣀ゟৠߚᏗⱘ䱣ᴎব䞣ⱘ੠䙉ᕾⱘߚᏗ䖥ԐЎℷᗕߚᏗǄ
೒ 5-1 ϔϾഛؐЎ 0ˈᷛ‫ޚ‬ᏂЎ 1 ⱘℷᗕߚᏗǄ

(a)೼ 80%㕂ֵᑺϟˈ䱣ᴎব䞣ؐԡѢঠջऎ䯈[-1.28, 1.28]П䯈Ǆ⊼ᛣz80˙1.28Ǆ᳝ 10%㕂ֵᑺ݊㨑
ܹऎ䯈Ꮊջˈ10%㨑ܹऎ䯈েջǄ(b)೼ 90%㕂ֵᑺϟˈ䱣ᴎব䞣ԡѢऩջऎ䯈[-Ğ, 1.28]ϞǄ
ὖ㗠㿔Пˈབᵰ䱣ᴎব䞣 Y ᳡ҢℷᗕߚᏗˈഛؐЎˈᷛ‫ޚ‬ᏂЎ³ˈ䙷М Y ⱘӏϔ㾖

ᆳؐ y ᳝ N%ⱘᴎӮ㨑ܹϟ䴶ⱘऎ䯈˖
fzN³ ˄5.10˅
ⳌԐഄˈഛ᳝ؐ N%ⱘᴎӮ㨑ܹϟ䴶ⱘऎ䯈˖
yfzN³ ˄5.11˅
ᕜᆍᯧᇚℸ㒧䆎੠ࠡ䴶ⱘ㒧䆎㒧ড়䍋ᴹ᥼ᇐᓣ 5.1 ⱘ⾏ᬷؐ‫؛‬䆒ⱘNˁ㕂ֵऎ䯈ⱘϔ㠀

㸼䖒ᓣǄ佪‫⬅ˈܜ‬ѢerrorS(h)䙉ҢѠ乍ߚᏗˈ݊ഛؐЎerrorD(h)ˈᷛ‫ޚ‬Ꮒབᓣ 5.9 ᠔⼎Ǆ݊⃵ˈ
៥Ӏⶹ䘧ᇍѢ䎇໳໻ⱘḋᴀ໻ᇣnˈѠ乍ߚᏗ䴲ᐌ䖥ԐѢℷᗕߚᏗǄ㄀ϝˈᓣ 5.11 ਞ䆝៥Ӏ
བԩЎԄ䅵ℷᗕߚᏗⱘഛؐ∖ߎNˁ㕂ֵऎ䯈Ǆ಴ℸˈᇚerrorS(h)ⱘഛؐ੠ᷛ‫ޚ‬Ꮒҷܹࠄᓣ
5.11 Ёᇚᕫࠄᓣ 5.1 Ёᇍ⾏ᬷؐ‫؛‬䆒ⱘNˁ㕂ֵऎ䯈Ў˖

errorS (h) r z N
n
ಲᖚϔϟˈ೼㸼䖒ᓣⱘ᥼ᇐЁ᳝ϸϾ䖥Ԑ࣪˖
1.Ԅ䅵errorS(h)ⱘᷛ‫ޚ‬Ꮒ³ᯊˈ៥ӀᇚerrorD(h)䖥ԐЎerrorS(h)˄ेҢᓣ 5.8 ࠄᓣ 5.9 ⱘ᥼

ᇐ˅Ǆ
2.Ѡ乍ߚᏗ⬅ℷᗕߚᏗ䖥ԐǄ
㒳䅵ᄺЁⱘϔ㠀㾘߭㸼ᯢˈ䖭ϸϾ䖥Ԑ೼ nı30 ៪ np(1-p)ı5 ᯊᎹ԰ᕫᕜདǄᇍѢ䕗ᇣ

ⱘ n ؐˈ᳔དՓ⫼߫㸼ⱘᔶᓣ㒭ߎѠ乍ߚᏗⱘ‫݋‬ԧؐǄ
5.3.6 ਂ‫ૂם‬ঋ‫ם‬䗯⮂
Ϟ䗄ⱘ㕂ֵऎ䯈ᰃঠջⱘˈेᅗ㾘ᅮњԄ䅵䞣ⱘϞ⬠੠ϟ⬠Ǆ೼ᶤѯᚙ‫މ‬ϟˈৃ㛑㽕⫼
ࠄऩջ䖍⬠Ǆ՟བˈᦤߎ䯂乬ĀerrorD(h)㟇໮ЎUⱘὖ⥛āǄ೼া㽕䰤ᅮhⱘ᳔໻䫭䇃⥛ˈ㗠
ϡ೼Тⳳᅲ䫭䇃⥛ᰃ৺ᇣѢԄ䅵䫭䇃⥛ᯊˈᕜ㞾✊Ӯᦤߎ䖭⾡䯂乬Ǆ
া㽕ᇍϞ䗄ⱘ䖛⿟԰ϔᇣⱘׂᬍህৃ䅵ㅫऩջ䫭䇃⥛䖍⬠Ǆᅗ᠔෎ѢⱘџᅲЎℷᗕߚᏗ
݇Ѣ݊ഛؐᇍ⿄Ǆ಴ℸˈӏᛣℷᗕߚᏗϞⱘঠջ㕂ֵऎ䯈㛑໳䕀ᤶЎⳌᑨⱘऩջऎ䯈ˈ㕂ֵ
ᑺЎॳᴹⱘϸ‫ס‬Ǆ˄㾕೒ 5-1(b)˅Ǆᤶ㿔Пˈ⬅ϔϾ᳝ϟ⬠ L ੠Ϟ⬠ U ⱘ 100(1-¢)%㕂ֵऎ
䯈ˈৃᕫࠄϔϾϟ⬠Ў L ᮴Ϟ⬠ⱘ 100(1-¢/2)%㕂ֵऎ䯈ˈৠᯊгৃᕫߎϔϾ᳝Ϟ⬠ U ᮴
ϟ⬠ⱘ 100(1-¢/2)%㕂ֵऎ䯈Ǆ䖭䞠¢ᇍᑨѢⳳᅲؐ㨑೼ᣛᅮऎ䯈໪ⱘὖ⥛Ǆᤶহ䆱䇈ˈ¢
ᰃⳳᅲؐ㨑ܹ೒ 5-1(a)Ё᮴䰈ᕅ䚼ߚⱘὖ⥛ˈ¢/2 ᰃ㨑ܹ೒ 5-1˄b˅ⱘ᮴䰈ᕅ䚼ߚⱘὖ⥛Ǆ
Ў䇈ᯢ䖭ϔ⚍ˈ‫⃵ݡ‬㗗㰥hѻ⫳r=12 Ͼ䫭䇃Ϩḋᴀ໻ᇣn=40 ⱘ䖭Ͼ՟ᄤǄབϞ᠔䗄ˈᅗ

ᇐ㟈ϔϾঠջⱘ 95ˁ㕂ֵऎ䯈 0.3f0.14Ǆ݊Ё 100(1-¢)=95ˁˈ᠔ҹ¢˙0.05Ǆ಴ℸˈᑨ
⫼ҹϞ㾘߭ˈৃᕫ᳝ 100(1-¢/2)˙97.5ˁⱘ㕂ֵᑺerrorD(h)᳔໮Ў 0.30ˇ0.14˙0.44ˈ㗠ϡㅵ
errorD(h)ⱘϟ⬠Ǆ಴ℸ೼errorD(h)Ϟⱘऩջ䫭䇃⥛䖍⬠↨Ⳍᑨⱘঠջ䖍⬠᳝ঠ‫ⱘס‬㕂ֵᑺ˄㾕
д乬 5.3˅Ǆ
5.4 ᧞ሲ㖤‫ؗ‬॰䰪Ⲻж㡢ᯯ⌋
ࠡϔ㡖ҟ㒡ⱘᰃ䩜ᇍϔ⡍ᅮᚙ‫᥼މ‬ᇐ㕂ֵऎ䯈Ԅ䅵˖෎Ѣ⣀ゟᢑপⱘnḋᴀˈԄ䅵⾏ᬷ
ؐ‫؛‬䆒ⱘerrorD(h)Ǆᴀ㡖ҟ㒡ⱘᮍ⊩ᰃ೼䆌໮Ԅ䅵䯂乬Ё⫼ࠄⱘ䗮⫼ⱘᮍ⊩Ǆ⹂ߛഄ䆆ˈ៥
Ӏৃҹᇚℸⳟ԰ᰃ෎Ѣ໻ᇣЎnⱘ䱣ᴎᢑপḋᴀⱘഛؐˈᴹԄ䅵ᘏԧഛؐⱘ䯂乬Ǆ
䗮⫼ⱘ䖛⿟ࣙ৿ҹϟℹ偸˖
1.⹂ᅮ෎‫ޚ‬ᘏԧЁ㽕Ԅ䅵ⱘখ᭄pˈ՟བerrorD(h)Ǆ
2.ᅮНϔϾԄ䅵䞣Y˄བerrorS(h)˅ᅗⱘ䗝ᢽᑨЎ᳔ᇣᮍᏂⱘ᮴‫أ‬Ԅ䅵䞣Ǆ
3.⹂ᅮԄ䅵䞣᠔᳡Ңⱘὖ⥛ߚᏗDYˈࣙᣀ݊ഛؐ੠ᮍᏂǄ
4.⹂ᅮNˁ㕂ֵऎ䯈ˈ䗮䖛ᇏᡒ䯜ؐL੠UҹՓ䖭ϾᣝDYߚᏗⱘ䱣ᴎব䞣᳝NˁᴎӮ㨑ܹL
੠UП䯈Ǆ
ৢ䴶ⱘ޴㡖ᇚᑨ⫼䆹䗮⫼ⱘᮍ⊩ࠄ݊Ҫ޴⾡ᴎ఼ᄺдЁᐌ㾕ⱘԄ䅵䯂乬Ǆ佪‫ܜ‬៥Ӏ䳔㽕
䅼䆎Ԅ䅵⧚䆎ⱘϔϾ෎ᴀ៤ᵰˈ⿄ЎЁᖗᵕ䰤ᅮ⧚(Central Limit Theorem)Ǆ
5.4.1 ѣᗹᶷ䲆ᇐ⨼
Ёᖗᵕ䰤ᅮ⧚ᰃㅔ࣪㕂ֵऎ䯈ⱘ᥼ᇐⱘϔϾ෎ᴀḍ᥂Ǆ㗗㰥བϟⱘϔ㠀Ḛᶊ˖೼nϾ⣀
ゟᢑপⱘϨ᳡Ңৠḋὖ⥛ߚᏗⱘ䱣ᴎব䞣Y1ĂYnЁ㾖ᆳᅲ偠ؐ˄བৠϔ⹀Ꮥⱘn⃵ᡯᦋ˅ǄҸ
ҷ㸼↣ϔব䞣Yi᳡Ңⱘ᳾ⶹߚᏗⱘഛؐˈᑊҸ³ҷ㸼ᷛ‫ޚ‬ᏂǄ⿄䖭ѯব䞣YiЎ⣀ ⣀ゟৠߚᏗ
(independent, identically distributed)䱣ᴎব䞣ˈ಴ЎᅗӀᦣ䗄ⱘᰃ৘㞾⣀ゟᑊϨ᳡Ңৠḋὖ⥛
1 n
ߚᏗⱘᅲ偠ǄЎԄ䅵Yi᳡ҢⱘߚᏗⱘഛؐˈ៥Ӏ䅵ㅫḋᴀⱘؐ Yn { ¦ Yi ˄བn⃵ᡩᦋ
n i1
ǄЁᖗᵕ䰤ᅮ⧚䇈ᯢ೼nėĞᯊ Yn ᠔᳡Ңⱘὖ⥛ߚᏗЎϔℷᗕߚᏗˈ
⹀ᏕЁߎ⦄ℷ䴶ⱘ↨՟˅
V
㗠ϡ䆎Yiᴀ䑿᳡ҢҔМḋⱘߚᏗǄ᳈䖯ϔℹˈYn ᳡ҢⱘߚᏗഛؐЎ㗠Ϩᷛ‫ޚ‬ᏂЎ ˈ㊒
n
⹂ⱘᅮНབϟ˖
ᅮ⧚ 5.1˖Ёᖗᵕ䰤ᅮ⧚Ǆ㗗㰥⣀ゟৠߚᏗⱘ䱣ᴎব䞣Y1…Ynⱘ䲚ড়ˈᅗӀ᳡Ңϔӏᛣⱘὖ⥛ߚ
1 n
ᏗˈഛؐЎˈ᳝䰤ᮍᏂ³2ǄᅮНḋᴀഛؐ Yn { ¦ Yi
n i1
Ǆ߭ᔧnėĞᯊϟ䴶ⱘᓣᄤ
Yn P
V
n
᳡ҢϔℷᗕߚᏗˈഛؐЎ 0 Ϩᷛ‫ޚ‬ᏂЎ 1Ǆ
䖭ϔ㒧䆎䴲ᐌҸҎৗ᚞ˈ಴Ўᅗ䇈ᯢ೼ϡⶹ䘧⣀ゟⱘYi᠔᳡Ңⱘ෎‫ߚޚ‬Ꮧⱘᚙ‫މ‬ϟˈ៥
Ӏৃҹᕫⶹḋᴀഛؐ Y ⱘߚᏗᔶᓣǄ᳈䖯ϔℹˈЁᖗᵕ䰤ᅮ⧚䇈ᯢњᗢḋՓ⫼ Y ⱘഛؐ੠
ᮍᏂᴹ⹂ᅮऩ⣀ⱘYiⱘഛؐ੠ᮍᏂǄ
Ёᖗᵕ䰤ᅮ⧚ᰃϔϾ䴲ᐌ᳝⫼ⱘ㒧䆎ˈ಴Ўᅗ㸼⼎ӏᛣḋᴀഛؐⱘԄ䅵䞣˄བerrorS(h)
Ўഛؐ䫭䇃⥛˅᳡ҢⱘߚᏗ೼n䎇໳໻ᯊৃ䖥ԐЎℷᗕߚᏗǄབᵰ䖬ⶹ䘧䖭ϔ䖥Ԑⱘℷᗕߚ
ᏗⱘᮍᏂˈህৃ⫼ᓣ 5.11 ᴹ䅵ㅫ㕂ֵऎ䯈ǄϔϾ䗮ᐌⱘ㾘߭ᰃ೼nı30 ᯊৃՓ⫼䖭ϔ䖥ԐǄ
ࠡ䴶ⱘゴ㡖៥ӀℷᰃՓ⫼њℷᗕߚᏗᴹ䖥Ԑഄᦣ䗄errorS(h)᳡ҢⱘѠ乍ߚᏗǄ
5.5 њ‫ٽ‬䇴䭏䈥⦽䰪Ⲻᐤᔸ
⦄㗗㰥ᇍᶤ⾏ᬷⳂᷛߑ᭄᳝ϸϾ‫؛‬䆒h1੠h2Ǆ‫؛‬䆒h1೼ϔᢹ᳝n1Ͼ⣀ゟᢑপḋ՟ⱘḋᴀ
S1Ϟ⌟䆩ˈϨh2೼n2Ͼৠḋᢑপⱘḋ՟ⱘḋᴀS2Ϟ⌟䆩Ǆ‫؛‬ᅮ㽕Ԅ䅵䖭ϸϾ‫؛‬䆒ⱘⳳᅲ䫭䇃
⥛䯈ⱘᏂᓖ˖
d { errorD (h1 ) errorD (h2 )
ৃՓ⫼㡖 5.4 Ёᦣ䗄ⱘಯℹ偸ᴹ᥼ᇐ d ⱘ㕂ֵऎ䯈Ԅ䅵Ǆ೼⹂ᅮ d ЎᕙԄ䅵ⱘখ᭄ৢˈ

ϟ䴶㽕ᅮНϔԄ䅵䞣Ǆᕜᰒ✊ˈ䖭䞠ৃ䗝ᢽḋᴀ䫭䇃⥛П䯈ⱘᏂᓖ԰ЎԄ䅵䞣ˈᷛ䆄Ў d̂ ˖
dˆ { errorS1 (h1 ) errorS 2 (h2 )
೼ℸϡࡴ䆕ᯢഄՓ⫼ d̂ ेЎ d ⱘ᮴‫أ‬Ԅ䅵䞣ˈे E[dˆ ] dǄ
䱣ᴎব䞣 dˆ ᳡Ңⱘὖ⥛ߚᏗᰃҔМ˛Ңࠡ䴶ⱘゴ㡖Ёˈ៥Ӏⶹ䘧ᇍѢ䕗໻ⱘn1੠n2˄↨
བ䛑ı30˅ˈerrorS1(h1)੠errorS2(h2)䛑䖥Ԑ䙉ҢℷᗕߚᏗǄ⬅ѢϸℷᗕߚᏗⱘᏂҡЎϔℷᗕߚ
Ꮧˈ಴ℸ dˆ г䖥Ԑ䙉ҢℷᗕߚᏗˈഛؐЎdǄৠᯊˈৃᕫ䆹ߚᏗⱘᮍᏂЎerrorS1(h1)੠errorS2(h2)
ⱘᮍᏂⱘ੠ǄՓ⫼ᓣ 5.9 㦋ᕫ䖭ϸϾߚᏗⱘ䖥ԐᮍᏂˈ᳝˖
errorS1 (h1 )(1 errorS1 (h1 )) errorS 2 (h2 )(1 errorS 2 (h2 ))
V d2ˆ | ˄5.12˅
n1 n2
⦄೼Ꮖ⹂ᅮњԄ䅵䞣 d̂ ᠔᳡Ңⱘὖ⥛ߚᏗˈᕜᆍᯧᇐߎ㕂ֵऎ䯈ҹ䇈ᯢՓ⫼ d̂ ᴹԄ䅵d

ⱘৃ㛑䇃ᏂǄ䱣ᴎব䞣 dˆ ᳡ҢഛؐdᮍᏂ³2ⱘℷᗕߚᏗˈ݊Nˁ㕂ֵऎ䯈Ԅ䅵Ў dˆ r z N V Ǆ
Փ⫼Ϟ䴶㒭ߎⱘᮍᏂ V d̂ ⱘ䖥Ԑؐˈdⱘ䖥ԐⱘNˁ㕂ֵऎ䯈Ԅ䅵Ў˖
2
errorS1 (h1 )(1 errorS1 (h1 )) errorS 2 (h2 )(1 errorS 2 (h2 ))
dˆ r z N (5.13)
n1 n2
݊ЁzNᰃ㸼 5-1 Ёᦣ䗄ⱘᐌ᭄ǄϞᓣ㒭ߎњϔ㠀ⱘঠջ㕂ֵऎ䯈ˈҹԄ䅵ϸϾ‫؛‬䆒䫭䇃

⥛П䯈ⱘᏂᓖǄ᳝ᯊৃ㛑䳔㽕ᶤϔ㕂ֵᑺϟⱘऩջⱘ䖍⬠üü㽕М⬠ᅮ᳔໻ৃ㛑Ꮒᓖˈ㽕М
Ў᳔ᇣⱘǄऩջ㕂ֵऎ䯈ৃҹ⫼ 5.3.6 㡖Ёᦣ䗄ⱘᮍ⊩ᴹׂᬍϞᓣ㗠ᕫࠄǄ
㱑✊Ϟ䴶ⱘߚᵤ㗗㰥ࠄњh1੠h2೼ⳌѦ⣀ゟⱘ᭄᥂ḋᴀϞ⌟䆩ˈ᳈䗮ᐌⱘᚙ‫މ‬ᰃ೼ϔϾ
ḋᴀS˄Sҡ✊⣀ゟѢh1੠h2˅Ǆ䖭ḋˈ d̂ 㹿䞡ᮄᅮНЎ˖
dˆ { errorS (h1 ) errorS (h2 )
ᔧՓ⫼Sᴹҷ᳓S1੠S2ᯊˈᮄⱘ d̂ ЁⱘᮍᏂ䗮ᐌᇣѢ 5.12 ᓣ㒭ߎⱘᮍᏂǄ䖭ᰃ಴ЎˈՓ

⫼ऩϾⱘḋᴀS⍜䰸њ⬅S1੠S2ⱘ㒘ড়ᏺᴹⱘ䱣ᴎᏂᓖǄ䖭ḋˈ⬅ᓣ 5.13 㒭ߎⱘ㕂ֵऎ䯈ϔ
㠀䇈ᴹӮ䖛ѢֱᅜˈԚҡ✊ᰃℷ⹂ⱘǄ
5.5.1 ‫ٽ‬䇴Ỷ僂
᳝ᯊ៥Ӏᛳ݈䍷ⱘᰃᶤ⡍ᅮⱘ⣰ᛇℷ⹂ⱘὖ⥛ˈ㗠ϡᰃᇍᶤখ᭄ⱘऎ䯈Ԅ䅵Ǆ↨བϟ䴶
ⱘ䯂乬ĀerrorD(h1)>errorD(h2)ⱘৃ㛑ᗻ᳝໮໻˛āǄҡՓ⫼ࠡϔ㡖ⱘᴵӊ䆒ᅮˈ‫؛‬ᅮ㽕⌟䞣h1
੠h2ⱘḋᴀ䫭䇃⥛ˈՓ⫼໻ᇣЎ 100 ⱘ⣀ゟḋᴀS1੠S2ˈᑊϨⶹ䘧errorS1(h1)˙0.30 ϨerrorS2(h2)
˙0.20ˈ಴ℸᏂᓖ d̂ Ў 0.10Ǆᔧ✊ˈ⬅Ѣ᭄᥂ḋᴀⱘ䱣ᴎᗻˈेՓerrorD(h1)İ errorD(h2)ˈ
ҡ᳝ৃ㛑ᕫࠄ䖭ḋⱘᏂᓖǄ೼䖭䞠ˈ㒭ᅮḋᴀ䫭䇃⥛ d̂ ˙0.10ˈerrorD(h1)>errorD (h2)ⱘὖ⥛
ᰃ໮ᇥ˛ㄝӋഄˈབԩ䅵ㅫ೼ d̂ ˙0.10 ᯊd>0 ⱘὖ⥛˛
⊼ᛣὖ⥛ Pr(d˚0)ㄝѢ dˆ 䖛催Ԅ䅵 d ϡ໮Ѣ 0.1 ⱘὖ⥛Ǆгेˈ䖭Ͼὖ⥛Ў d̂ 㨑ܹऩջ

ऎ䯈 d̂ <d+0.10 ⱘὖ⥛Ǆ⬅Ѣ d ᰃ d̂ ᠔᳡ҢߚᏗⱘഛؐˈϞᓣㄝӋѢ dˆ P dˆ 0.10 Ǆ
ὖᣀഄ䇈ˈὖ⥛ Pr(d˚0)ㄝѢ d̂ 㨑ܹऩջऎ䯈 dˆ P dˆ 0.10 ⱘὖ⥛Ǆ⬅Ѣࠡϔ㡖៥Ӏ
Ꮖ䅵ㅫߎ d̂ ⱘ໻㟈ߚᏗˈህৃҹ䗮䖛 d̂ ߚᏗ೼䆹ऎ䯈ⱘὖ⥛䋼䞣ᴹ⹂ᅮ d̂ 㨑ܹ䖭Ͼऩջऎ

䯈ⱘὖ⥛Ǆ
佪‫ܜ‬ᇚऎ䯈 dˆ P dˆ 0.10 㸼⼎Ў‫ܕ‬䆌Ңഛؐ‫ⱘ⾏أ‬ᷛ‫ޚ‬Ꮒ᭄ǄՓ⫼ᓣ 5.12 ৃᕫ

V dˆ | 0.061 ˈ᠔ҹ䖭ϔऎ䯈ৃ䖥Ԑ㸼⼎Ў
dˆ P dˆ 1.64V dˆ
ϢℸℷᗕߚᏗⱘऩջऎ䯈Ⳍ݇㘨ⱘ㕂ֵᑺᰃ໮ᇥ˛ᶹ㸼 5-1ˈৃᕫഛ਼ؐೈⱘ 1.64 ᷛ‫ޚ‬

Ꮒᇍᑨ㕂ֵᑺ 90ˁⱘঠջऎ䯈Ǆ಴ℸ䖭Ͼऩջऎ䯈‫݋‬᳝㕂ֵᑺ 95ˁǄ
಴ℸˈ㒭ᅮ㾖ᆳࠄⱘ d̂ ˙0.1ˈerrorD(h1)>errorD(h2)ⱘὖ⥛㑺Ў 0.95Ǆḍ᥂㒳䅵ᄺⱘᴃ䇁ˈ

ৃ㸼䗄Ў˖᥹ফ˄accept˅ĀerrorD(h1)>errorD(h2)ā䖭ϔ‫؛‬䆒ˈ㕂ֵᑺЎ 0.95Ǆᤶϔ⾡䇈⊩ˈ
ৃ㸼䗄Ў៥Ӏᢦ㒱˄reject˅ᇍゟ‫؛‬䆒˄ᐌ⿄Ў䳊‫؛‬䆒˅ˈҹ(1-0.95)=0.05 ⱘᬜᑺ˄significance
level˅Ǆ
5.6 ᆜҖ㇍⌋∊䖹
᳝ᯊ៥Ӏ᳈ᛳ݈䍷Ѣ↨䕗ϸϾᄺдㅫ⊩LA੠LBⱘᗻ㛑ˈ㗠ϡᰃϸϾ⡍ᅮ‫؛‬䆒Ǆᗢḋ䖥Ԑ
ഄẔ偠໮Ͼᄺдㅫ⊩ˈབԩ⹂ᅮϸϾㅫ⊩П䯈ⱘᏂᓖ೼㒳䅵Ϟᰃ᳝ᛣНⱘ˛㱑✊ˈ೼ᴎ఼ᄺ
дⷨお乚ඳˈᇍѢ↨䕗ⱘᮍ⊩ҡ೼ѝ䆎Ёˈϡ䖛䖭䞠ҟ㒡њϔϾড়⧚ⱘ䗨ᕘǄ݇Ѣϡৠᮍ⊩
ⱘ䅼䆎㾕Dietterich˄1996˅
Ǆ
ᓔྟ‫ܜ‬ᣛᅮ㽕Ԅ䅵ⱘখ᭄Ǆ‫؛‬ᅮ᳝LA੠LBϸϾㅫ⊩ˈ㽕⹂ᅮЎњᄺдϔ⡍ᅮⳂᷛߑ᭄f
ᑇഛᴹ䇈䙷Ͼㅫ⊩᳈ӬǄᅮНĀᑇഛāⱘϔ⾡ড়⧚ᮍ⊩ᰃˈҢϔ෎‫ޚ‬ᅲ՟ߚᏗDЁᢑপ᠔᳝
ࣙ৿nϾḋ՟ⱘ䆁㒗䲚ড়ˈ೼᠔᳝䖭ḋⱘ䲚ড়Ё⌟䞣ϸϾㅫ⊩ⱘᑇഛᗻ㛑Ǆᤶহ䆱䇈ˈ䳔㽕
Ԅ䅵‫؛‬䆒䫭䇃⥛П䯈Ꮒᓖⱘᳳᳯؐ˖
E [errorD ( L A ( S )) errorD ( L B ( S ))] (5.14)

S D
݊ЁL(S)ҷ㸼㒭ᅮ䆁㒗ḋᴀSᯊᄺдㅫ⊩L䕧ߎⱘ‫؛‬䆒ˈϟᷛSD㸼⼎ᳳᳯؐᰃ೼෎‫ߚޚ‬
ᏗDЁᢑপⱘḋᴀSϞ䅵ㅫǄϞ䗄㸼䖒ᓣᦣ䗄ⱘᰃᄺдㅫ⊩LA੠LBⱘᏂⱘᳳᳯؐǄ
೼ᅲ䰙ⱘᄺдㅫ⊩↨䕗Ёˈ៥Ӏা᳝ϔϾ᳝䰤ⱘḋᴀD0Ǆ೼䖭⾡ᚙ‫މ‬ϟˈᕜᰒ✊ˈ㽕
Ԅ䅵Ϟ䗄ⱘ䞣䳔㽕ᇚD0ߚࡆ៤䆁㒗䲚ড়S0੠ϡⳌѸⱘ⌟䆩䲚ড়T0Ǆ䆁㒗᭄᥂ৃҹ⫼ᴹ᮶䆁㒗
LAজ䆁㒗LBˈ㗠⌟䆩᭄᥂߭⫼ᴹ↨䕗ϸϾᄺдࠄⱘ‫؛‬䆒ⱘ‫⹂ޚ‬ᑺˈгህᰃˈՓ⫼ϟᓣᴹ䅵ㅫ˖
errorT0 ( L A ( S 0 )) errorT0 ( LB ( S 0 )) ˄5.15˅
ϞᓣϢ 5.14 ᓣⱘ䅵ㅫ᳝ϸϾ݇䬂ⱘϡৠǄ佪‫ܜ‬៥ӀՓ⫼errorT0(h)ᴹ䖥Ԑ errorD(h)Ǆ㄀

Ѡˈ䫭䇃⥛ⱘᏂᓖ⌟䞣ᰃ೼ϔϾ䆁㒗䲚ড়S0Ϟ㗠ϡᰃ೼ҢߚᏗDЁᢑপⱘ᠔᳝ⱘḋᴀSϞ䅵ㅫ
ᳳᳯؐǄ
ᬍ䖯 5.15 ᓣⱘϔ⾡ᮍ⊩ᰃᇚ᭄᥂D0໮⃵ߚࡆЎϡⳌѸⱘ䆁㒗੠⌟䆩䲚ড়ˈ✊ৢ೼݊Ё
䅵ㅫ䫭䇃⥛ⱘᑇഛؐǄ䖭ϔ䖛⿟೼㸼 5-5 Ё߫ߎˈᅗ೼ϔৃ⫼᭄᥂ⱘ೎ᅮḋᴀD0ϞԄ䅵ϸϾ
ᄺдㅫ⊩䫭䇃⥛П䯈ⱘᏂᓖǄ䆹䖛⿟佪‫ܜ‬ᇚ᭄᥂ᢚߚЎkϾϡⳌѸⱘⳌㄝᄤ䲚ˈᄤ䲚໻ᇣ㟇
ᇥЎ 30Ǆ✊ৢ䆁㒗੠⌟䆩ㅫ⊩k⃵ˈ↣⃵Փ⫼݊ЁϔϾᄤ䲚԰Ў⌟䆩᭄᥂䲚ˈ݊Ҫk-1 Ͼᄤ
䲚Ў䆁㒗䲚ǄՓ⫼䖭⾡ࡲ⊩ˈᄺдㅫ⊩೼kϾ⣀ゟ⌟䆩䲚Ϟ⌟䆩ˈ㗠䫭䇃⥛ⱘᏂᓖⱘഛؐ G
԰ЎϸϾᄺдㅫ⊩䯈ᏂᓖⱘԄ䅵Ǆ
㸼 5-5 Ԅ䅵ϸᄺдㅫ⊩LA੠LB䫭䇃⥛Ꮒᓖⱘϔ⾡ᮍ⊩
䖥Ԑⱘ㕂ֵऎ䯈ᇚ೼ℷ᭛Ё㒭ߎǄ
1.ᇚৃ⫼᭄᥂D0ߚࡆ៤kϾⳌৠ໻ᇣⱘϡⳌѸᄤ䲚T1, T2, …, TkǄ݊໻ᇣ㟇ᇥЎ 30Ǆ

2.Ҹ i Ң 1 ࠄ k ᕾ⦃ˈ‫خ‬ϟ䴶ⱘ᪡԰˖
Փ⫼Ti԰Ў⌟䆩䲚ড়ˈ㗠࠽ԭⱘ᭄᥂԰Ў䆁㒗䲚ড়Si
S i m {D0 Ti }
hA m LA (S i )
hB m LB ( S i )
G i m errorT (hA ) errorT (hB )
i i
3.䖨ಲؐ G ˈ݊Ё
1 k
G { ¦G i
k i1
㸼 5-5 䖨ಲⱘ G ৃ㹿⫼԰ᇍᓣ 5.14 ᠔䳔㒧ᵰⱘϔϾԄ䅵Ǆ᳈ড়䗖ⱘ䇈⊩ᰃᡞ G ⳟ԰ϟ
ᓣⱘԄ䅵˖
E [errorD ( L A ( S )) errorD ( L B ( S ))] ˄5.16˅

S D0
k 1
݊ЁSҷ㸼ϔϾ໻ᇣЎ D0 ˈҢD0Ёϔ㟈ᢑপⱘ䱣ᴎḋᴀǄ೼䆹ᓣ੠ 5.14 Ёॳᴹ
k
ⱘ㸼䖒ᓣП䯈ˈᚳϔⱘᏂ߿೼Ѣ݊ᳳᳯؐⱘ䅵ㅫᰃ೼ৃ⫼᭄᥂ⱘᄤ䲚D0Ϟ䅵ㅫˈ㗠ϡᰃ೼
ҢᭈϾߚᏗDϞᢑপⱘᄤ䲚Ϟ䅵ㅫǄ
Ԅ䅵 5.16 ᓣⱘ䖥Ԑⱘ Nˁ㕂ֵऎ䯈ৃՓ⫼ G 㸼⼎Ў˖
G r t N ,k 1 sG ˄5.17˅
݊ЁtN, k-1ᰃϔᐌ᭄ˈ݊ᛣН㉏ԐѢࠡ䴶㕂ֵऎ䯈㸼䖒ᓣЁⱘzNˈ㗠 G s ҷ㸼ᇍ G ᠔᳡Ң

ⱘὖ⥛ߚᏗⱘᷛ‫ޚ‬ᏂⱘԄ䅵ˈ⹂ߛⱘ䆆ˈ sG ᅮНЎ˖
k
1
sG { ¦
k (k 1) i 1
(G i G ) 2 ˄5.18˅
⊼ᛣ 5.17 Ёⱘᐌ䞣tN, k-1᳝ϸϾϟᷛǄ㄀ϔϾҷ㸼᠔䳔ⱘ㕂ֵᑺˈབࠡ䴶ⱘᐌ᭄zNЁ䙷ḋǄ

㄀ѠϾখ᭄⿄Ў㞾⬅ᑺ˄degree of freedom˅ˈᐌ㹿䆄԰vˈᅗϢ⫳៤䱣ᴎব䞣 G ⱘؐᯊ⣀ゟ
ⱘ䱣ᴎџӊ᭄ⳂⳌ݇Ǆ೼ᔧࠡⱘᴵӊϟˈ㞾⬅ᑺ᭄ؐЎk-1Ǆখ᭄tⱘ޴⾡পؐ೼㸼 5-6 Ё߫
ߎǄ⊼ᛣᔧkėĞᯊˈtN, k-1ⱘؐ䍟Ѣᐌ᭄zNǄ
⊼ᛣ䖭䞠ᦣ䗄ⱘ↨䕗ᄺдㅫ⊩ⱘ䖛⿟㽕೼ৠḋⱘ⌟䆩䲚ড়Ϟ⌟䆩ϸϾ‫؛‬䆒Ǆ䖭Ϣ 5.5 㡖
Ёᦣ䗄ⱘ↨䕗ϸϾ⫼⣀ゟ⌟䆩䲚ড়䆘Ԅ䖛ⱘ‫؛‬䆒ϡৠǄՓ⫼Ⳍৠḋᴀᴹ⌟䆩‫؛‬䆒㹿⿄Ў䜡ᇍ
⌟䆩(paired test)Ǆ䜡ᇍ⌟䆩䗮ᐌӮѻ⫳᳈㋻ᆚⱘ㕂ֵऎ䯈Ǆ಴Ў೼䜡ᇍ⌟䆩ЁӏᛣⱘᏂᓖ䛑
ᴹ⑤Ѣ‫؛‬䆒П䯈ⱘᏂᓖǄⳌডˈ㢹‫؛‬䆒೼ߚᓔⱘ᭄᥂ḋᴀϞⱘ⌟䆩ˈϸϾḋᴀ䫭䇃⥛П䯈ⱘ
Ꮒᓖгৃ㛑䚼ߚᴹ⑤ѢϸϾḋᴀ㒘៤ϞⱘϡৠǄ
㸼 5-6 ঠջ㕂ֵऎ䯈tN,vⱘؐǄᔧvėĞᯊˈtN,v䍟䖥ѢzNǄ
㕂ֵᑺ N
90% 95% 98% 99%
v=2 2.92 4.30 6.96 9.92
v=5 2.02 2.57 3.36 4.03
v=10 1.81 2.23 2.76 3.17
v=20 1.72 2.09 2.53 2.84
v=30 1.70 2.04 2.46 2.75
v=120 1.66 1.98 2.36 2.62
v=Ğ 1.64 1.96 2.33 2.58
5.6.1 䞃ሯ t ⎁䈋
Ϟ䴶ᦣ䗄њ೼㒭ᅮ೎ᅮ᭄᥂䲚ᯊ↨䕗ϸϾᄺдㅫ⊩ⱘ䖛⿟Ǆᴀ㡖䅼䆎䖭ϔ䖛⿟ҹঞ 5.17
੠ 5.18 ᓣЁ㕂ֵऎ䯈ⱘ㒳䅵ᄺ䆎䆕Ǆབᵰ㄀ϔ⃵䯙䇏ˈৃҹ䏇䖛ᅗ㗠ϡ༅䖲㓁ᗻǄ
Ўњ⧚㾷 5.17 ᓣЁⱘ㕂ֵऎ䯈ˈ㗗㰥ҹϟⱘԄ䅵䯂乬˖

z 㒭ᅮϔ㋏߫⣀ゟৠߚᏗⱘ䱣ᴎব䞣Y1,Y2…Ykⱘ㾖ᆳؐǄ
z 㽕Ԅ䅵䖭ѯYi᠔᳡Ңⱘὖ⥛ߚᏗⱘഛؐǄ
z Փ⫼ⱘԄ䅵䞣Ўḋᴀഛؐ Y
1 k
Y { ¦ Yi
k i1
䖭ϔ෎Ѣḋᴀഛؐ Y Ԅ䅵ߚᏗഛؐⱘ䯂乬䴲ᐌᐌ㾕Ǆ՟བˈᅗ㽚Ⲫњᮽ‫⫼ܜ‬errorS(h)
ᴹԄ䅵errorD(h)ⱘ䯂乬Ǆ˄݊ЁˈYiЎ 0 ៪ 1 㸼⼎hᰃ৺ᇍϔऩ⣀ⱘSḋ՟ѻ⫳䇃ߚ㉏ˈ㗠errorD(h)
Ў෎‫ߚޚ‬ᏗⱘഛؐǄ˅⬅ᓣᄤ 5.17 ੠ 5.18 ᦣ䗄ⱘt⌟䆩ᑨ⫼Ѣ䆹䯂乬ⱘϔ⡍⅞ᚙᔶüüे
↣Ͼऩ⣀ⱘYi䙉ᕾℷᗕߚᏗǄ
⦄㗗㰥↨䕗ᄺдㅫ⊩ⱘ㸼 5-5 䖛⿟ⱘϔϾ⧚ᛇ࣪ᴵӊǄ‫؛‬ᅮϡᰃᢹ᳝೎ᅮḋᴀ᭄᥂D0ˈ

㗠ᰃҢ෎‫ޚ‬ᅲ՟ߚᏗЁᢑপᮄⱘ䆁㒗ḋ՟Ǆ೼䖭䞠៥Ӏׂᬍ㸼 5-5 Ёⱘ䖛⿟ˈՓ↣ϔ⃵ᕾ⦃
⫳៤ϔϾᮄⱘ䱣ᴎ䆁㒗䲚Si੠ᮄⱘ䱣ᴎ⌟䆩䲚Tiˈ⫳៤ᮍ⊩ᰃҢ෎‫ߚޚ‬ᏗЁᢑপ㗠ϡᰃҢ೎
ᅮḋᴀD0ЁᢑপǄ䖭ϔ⧚ᛇ࣪ᮍ⊩㛑ᕜདഄऍ䜡Ϟ䴶ⱘԄ䅵䯂乬Ǆ⡍߿ഄˈ䆹䖛⿟᠔⌟䞣
ⱘ¥i⦄ᇍᑨࠄ⣀ゟৠߚℹⱘ䱣ᴎব䞣YiǄ݊ߚᏗⱘഛؐᇍᑨϸᄺдㅫ⊩䫭䇃⥛ⱘᳳᳯᏂᓖ
Ǆḋᴀഛؐ Y Ў䖭ϔ⧚ᛇ࣪ᮍ⊩䅵ㅫߎⱘ G Ǆ⦄Ꮰᳯಲㄨ˖Ā G ᰃ৺㛑䕗དഄ
˄ेᓣ 5.14˅
Ԅ䅵āǄ
佪‫⊼ˈܜ‬ᛣࠄ⌟䆩䲚Tiⱘ໻ᇣ㟇ᇥࣙ৿ 30 Ͼḋ՟Ǆ಴ℸˈऩ⣀ⱘ¥iᇚ䖥Ԑ䙉ᕾℷᗕߚ
Ꮧ˄⬅Ёᖗᵕ䰤ᅮ⧚˅Ǆ಴ℸˈ៥Ӏ᳝ϔ⡍⅞ᴵӊेYi᳡Ң䖥ԐⱘℷᗕߚᏗǄৃҹᕫࠄˈϔ
㠀ഄˈᔧ↣ϾYi䙉ᕾℷᗕߚᏗᯊˈḋᴀഛؐ Y г䙉ᕾℷᗕߚᏗǄ⬅ℸˈৃҹ㗗㰥Փ⫼ࠡ䴶䅵
ㅫ㕂ֵऎ䯈ⱘ㸼䖒ᓣ˄ㄝᓣ 5.11˅ˈ݊ЁⱘԄ䅵䞣ℷᰃ䙉ᕾњℷᗕߚᏗǄ✊㗠ˈ䆹ㄝᓣ㽕∖
ⶹ䘧ߚᏗⱘᷛ‫ޚ‬ᏂˈԚ䖭䞠᳾ⶹǄ
t ⌟䆩ℷད⫼Ѣ䖭ḋⱘᚙᔶˈेԄ䅵ϔ㋏߫⣀ゟৠℷᗕߚᏗⱘ䱣ᴎব䞣ⱘḋᴀഛؐǄ೼
䖭䞠ˈৃՓ⫼ᓣ 5.17 ੠ 5.18 Ёⱘ㕂ֵऎ䯈ˈᅗৃ㹿䞡ᮄ㸼䗄Ў˖
P Y r t N ,k 1 sY
݊Ё sY ЎԄ䅵ⱘḋᴀഛؐⱘᷛ‫ޚ‬Ꮒ˖
k
1
sY { ¦ (Yi Y )2
k (k 1) i 1
㗠tN, k-1㉏ԐѢࠡ䴶ⱘzNⱘᐌ䞣Ǆᅲ䰙Ϟˈᐌ䞣tN, k-1ᦣ䗄ⱘᰃ⿄ЎtߚᏗⱘὖ⥛ߚᏗϟⱘऎ

ඳˈℷབᐌ᭄zNᦣ䗄њℷᗕߚᏗϟⱘऎඳǄtߚᏗᰃϔ㉏ԐѢℷᗕߚᏗⱘ䩳ᔶߚᏗˈԚ᳈ᆑ
Ϩ᳈ⷁˈҹড᯴⬅ѢՓ⫼ sY ᴹ䖥Ԑⳳᅲᷛ‫ޚ‬Ꮒ V Y ᯊᏺᴹⱘ᳈໻ⱘᮍᏂǄᔧk䍟䖥Ѣ᮴かᯊˈ
tߚᏗ䍟䖥ѢℷᗕߚᏗ˄಴ℸtN, k-1䍟䖥ѢzN˅Ǆ䖭೼Ⳉ㾝Ϟᰃℷ⹂ⱘˈ಴Ў៥ӀᏠᳯḋᴀ໻ᇣk
๲ࡴᯊ sY ᬊᬯࠄⳳᅲⱘᷛ‫ޚ‬Ꮒ V Y ˈϨ಴Ўᔧᷛ‫ޚ‬Ꮒ⹂ߛᏆⶹᯊৃՓ⫼zNǄ
5.6.2 ᇔ䱻Ⲻ㘹㲇
Ϟ䴶ⱘ䅼䆎䆕ᯢњ೼Փ⫼ḋᴀഛؐ Y ᴹԄ䅵ϔϾࣙ৿kϾ⣀ゟৠℷᗕߚᏗⱘ䱣ᴎব䞣
ⱘḋᴀഛؐᯊˈՓ⫼ᓣ 5.17 ᴹԄ䅵㕂ֵऎ䯈Ǆ䖭ऍ䜡њ៥Ӏⱘ⧚ᛇⱘᴵӊˈे‫؛‬ᅮᇍѢⳂ
ᷛߑ᭄ⱘḋ՟ৃ䖯㸠᮴䰤ᄬপǄ೼ᅲ䰙Ёˈ㢹᭄᥂䲚D0᳝䰤ˈϨㅫ⊩Փ⫼㸼 5-5 ᦣ䗄ⱘᅲ䰙
ᮍ⊩ˈ䖭ϔ䆕ᯢᑊϡϹḐ䗖⫼Ǆᅲ䰙ⱘ䯂乬ᰃˈЎѻ⫳¥iা᳝䞡ᮄ䞛ḋD0ˈҹ঺໪ⱘᮍ⊩
ᡞᅗߚࡆЎ⌟䆩䲚੠䆁㒗䲚Ǆ¥iℸᯊⳌѦᑊϡ⣀ゟˈ಴ЎᅗӀ෎ѢҢ᳝䰤ᄤ䲚D0Ёᢑপⱘ
ⳌѦ䞡঴ⱘ䆁㒗ḋ՟ˈ㗠ϡᰃҢᭈϾߚᏗDЁᢑপǄ
ᔧা᳝ϔϾ᳝䰤ⱘ᭄᥂ḋᴀD0ৃ⫼ᯊˈ᳝޴⾡ᮍ⊩⫼ᴹ䞡䞛ḋD0Ǆ㸼 5-5 ᦣ䗄ⱘᰃk-fold

ᮍ⊩ˈ݊ЁD0㹿ߚЎkϾϡⳌѸⱘㄝ໻ᇣⱘᄤ䲚ˈ೼䖭⾡k-foldᮍ⊩ЁˈD0Ё↣ϔḋ՟䛑᳝ϔ
⃵⫼Ѣ⌟䆩ˈ㗠k-1 ⃵⫼Ѣ䆁㒗Ǆ঺ϔ⾡ᐌ⫼ⱘᮍ⊩ᰃҢD0Ё䱣ᴎᢑপ㟇ᇥ 30 Ͼḋ՟ⱘ䲚
ড়ˈ‫࠽⫼ݡ‬ԭⱘḋ՟ᴹ䆁㒗ˈ䞡໡䖭ϔ䖛⿟Ⳉࠄ䎇໳ⱘ⃵᭄Ǆ䖭⾡䱣ᴎᮍ⊩ⱘད໘ᰃ㛑໳䞡
໡᮴䰤⃵ˈҹ‫ޣ‬ᇣ㕂ֵऎ䯈ࠄ䳔㽕ⱘᆑᑺǄⳌডˈk-foldᮍ⊩ফ䰤Ѣḋ՟ⱘᘏ᭄ˈ䖭ᰃ಴Ў
↣Ͼḋ՟া᳝ϔ⃵⫼Ѣ⌟䆩ˈϨᏠᳯḋᴀ໻ᇣ㟇ᇥЎ 30Ǆ✊㗠ˈ䱣ᴎᮍ⊩ⱘ㔎⚍ᰃˈ⌟䆩
䲚ড়ϡ‫ݡ‬㛑ⳟ԰ᰃҢ෎‫ߚޚ‬ᏗD⣀ゟᢑপǄⳌডˈk-foldѸঝ偠䆕⫳៤ⱘ⌟䆩䲚ড়ᰃ⣀ゟⱘˈ
಴Ўϔᅲ՟া೼⌟䆩䲚ড়Ёߎ⦄ϔ⃵Ǆ
ὖᣀഄ䇈ˈ෎Ѣ᳝䰤᭄᥂ⱘᄺдㅫ⊩ⱘ↨䕗Ё≵᳝ϔϾऩ⣀ⱘᮍ⊩㛑⒵䎇៥ӀᏠᳯⱘ᠔
᳝㑺ᴳǄ᳝ᖙ㽕䆄ԣ㒳䅵ᄺ῵ൟ೼᭄᥂᳝䰤ᯊᕜᇥ㛑ᅠ㕢ഄऍ䜡ᄺдㅫ⊩偠䆕ⱘ᠔᳝㑺ᴳǄ
✊㗠ᅗӀ⹂ᅲᦤկњ䖥Ԑⱘ㕂ֵऎ䯈ˈ᳝ࡽѢ㾷䞞ᄺдㅫ⊩ⱘᅲ偠ᗻ↨䕗Ǆ
ᴀゴⱘ㽕⚍ࣙᣀ˖
z 㒳䅵⧚䆎ᦤկњϔϾ෎⸔ˈҢ㗠෎Ѣ೼᭄᥂ḋᴀSϞⱘ㾖ᆳ䫭䇃⥛errorS(h)ˈԄ
䅵‫؛‬䆒hⱘⳳᅲ䫭䇃⥛errorD(h)Ǆ՟བˈབᵰhЎϔ⾏ᬷؐ‫؛‬䆒ˈ㗠Ϩ᭄᥂ḋᴀࣙ
ᣀnı30 Ͼϡձ䌪hϨⳌѦ⣀ゟⱘḋ՟ᯊˈ䙷МerrorD(h)ⱘNˁ㕂ֵऎ䯈䖥ԐЎ˖
errorS (h) r z N
n
݊ЁzNⱘؐ⬅㸼 5-1 㒭ߎǄ
z ϔ㠀ഄˈԄ䅵㕂ֵऎ䯈ⱘ䯂乬ৃ䗮䖛⹂ᅮϔᕙԄ䅵ⱘখ᭄˄བerrorD (h)˅ҹঞⳌ
ᇍᑨⱘԄ䅵䞣˄errorS (h)˅ᴹᅠ៤Ǆ⬅ѢԄ䅵䞣ᰃϔ䱣ᴎব䞣˄བerrorS(h)ձ䌪
Ѣ䱣ᴎḋᴀS˅ˈᅗৃ⬅݊᳡Ңⱘὖ⥛ߚᏗᴹᦣ䗄Ǆ㕂ֵऎ䯈ⱘ䅵ㅫৃ䗮䖛⹂ᅮ
䆹ߚᏗϟࣙ৿᠔䳔ὖ⥛䋼䞣ⱘऎ䯈ᴹᦣ䗄Ǆ
z Ԅ䅵‫؛‬䆒㊒ᑺⱘϔ⾡䇃Ꮒৃ㛑ЎԄ䅵‫أ‬Ꮒ˄estimation bias˅ǄབᵰYЎᇍᶤখ᭄
pⱘԄ䅵䞣ˈYⱘԄ䅵‫أ‬ᏂЎYⱘᳳᳯؐ੠pП䯈ⱘᏂǄ՟བˈབᵰSᰃ⫼ᴹᔶ៤‫؛‬
䆒hⱘ䆁㒗᭄᥂ˈ߭errorS(h)㒭ߎњⳳᅲ䫭䇃⥛errorD(h)ⱘϔϾ‫أ‬ѢФ㾖࣪ⱘԄ
䅵Ǆ
z Ԅ䅵ѻ⫳䇃Ꮒⱘ㄀Ѡ⾡ॳ಴ᰃԄ䅵ᮍᏂ˄variance˅ǄेՓᇍѢ᮴‫أ‬Ԅ䅵ˈԄ䅵
䞣ⱘ㾖ᆳؐг᳝ৃ㛑೼৘ᅲ偠ЁϡৠǄԄ䅵䞣ߚᏗⱘᮍᏂ³2ᦣ䗄њ䆹Ԅ䅵Ϣⳳ
ᅲؐⱘϡৠ᳝໮໻Ǆ䆹ᮍᏂ೼᭄᥂ḋᴀ๲໻ᯊ䰡ԢǄ
z ↨䕗ϸᄺдㅫ⊩ᬜᵰⱘ䯂乬೼᭄᥂੠ᯊ䯈᮴䰤ᯊᰃϔϾⳌᇍᆍᯧⱘԄ䅵䯂乬ˈ
Ԛ೼䌘⑤᳝䰤ᯊ㽕ೄ䲒ᕫ໮Ǆᴀゴᦣ䗄ⱘϔ⾡䗨ᕘᰃ೼ৃ⫼᭄᥂ⱘϡৠᄤ䲚Ϟ
䖤㸠ᄺдㅫ⊩ˈ೼࠽ԭ᭄᥂Ϟ⌟䆩ᄺࠄⱘ‫؛‬䆒ˈ✊ৢᇚ䖭ѯᅲ偠ⱘ㒧ᵰᑇഛǄ
z 䖭䞠᠔㗗㰥ⱘ໮᭄ᚙ‫މ‬ϟˈ᥼ᇐ㕂ֵऎ䯈䳔㽕䖯㸠໮Ͼ‫؛‬ᅮ੠䖥ԐǄ՟བϞ䴶
ⱘerrorD (h)ⱘ㕂ֵऎ䯈䳔㽕ᇚѠ乍ߚᏗ䖥ԐЎℷᗕߚᏗ˗䖥Ԑ䅵ㅫߚᏗⱘᮍᏂ˗
ҹঞ‫؛‬ᅮᅲ՟Ңϔ೎ᅮϡবⱘὖ⥛ߚᏗЁ⫳៤Ǆ෎Ѣ䖭ѯ䖥Ԑᕫࠄⱘऎ䯈াᰃ
䖥Ԑ㕂ֵऎ䯈ˈԚᅗӀҡᦤկњ䆒䅵੠㾷䞞ᴎ఼ᄺдᅲ偠㒧ᵰⱘ᳝ᬜᣛᇐǄ
ᴀゴҟ㒡ⱘ݇䬂㒳䅵ᄺᅮН೼㸼 5-2 Ё߫ߎǄ
Փ⫼㒳䅵ⱘᮍ⊩ᴹԄ䅵੠⌟䆩‫؛‬䆒䖭ϔЏ乬᳝໻䞣ⱘ᭛⤂Ǆᴀゴাҟ㒡њ෎ᴀὖᗉˈ㒚
㡖ⱘ䯂乬ৃ೼䆌໮к㈡੠᭛ゴЁᡒࠄǄBillingsley et al.˄1986˅ᦤկњᇍ㒳䅵ᄺⱘϔϾᕜㅔ
ᯢⱘҟ㒡ˈ䆺ሑᦣ䗄њ䖭䞠᠔䅼䆎ⱘϔѯ䯂乬Ǆ݊Ҫ᭛⤂ࣙᣀ DeGroot˄1986˅˗ Casella &
Berger˄1990˅ǄDuda & Hart˄1973˅೼᭄ؐ῵ᓣ䆚߿乚ඳᦤߎњ䖭ѯ䯂乬ⱘ㾷‫އ‬Ǆ
ˈEtzioni & Etzioni˄1994˅

Segre et al.˄1991ˈ1996˅ ˈҹঞ Gordon & Segre˄1996˅䅼
䆎њ䆘Ԅᄺдㅫ⊩ⱘ㒳䅵ᛣН⌟䆩ˈㅫ⊩ⱘᗻ㛑ḍ᥂݊ᬍ䖯䅵ㅫᬜ⥛ⱘ㛑࡯ᴹ䆘⌟Ǆ
Geman et al.˄1992˅䅼䆎њ೼ৠᯊ᳔ᇣ࣪‫أ‬Ꮒ੠᳔ᇣ࣪ᮍᏂП䯈԰ߎⱘᡬЁǄ䖭ϔҢ
᳝䰤᭄᥂Ёᄺд੠↨䕗‫؛‬䆒ⱘЏ乬ҡ೼ѝ䆎ЁǄ՟བˈDietterich˄1996˅䅼䆎њ೼ϡৠⱘ䆁
㒗-⌟䆩᭄᥂ߚࡆϟՓ⫼䜡ᇍᏂᓖ t ⌟䆩ᏺᴹⱘ亢䰽Ǆ
д乬
5.1 ‫؛‬ᅮẔ偠ϔ‫؛‬䆒hˈᑊথ⦄೼ϔࣙ৿n=1000 Ͼ䱣ᴎᢑপḋ՟ⱘḋᴀSϞˈᅗߎ⦄r=300

Ͼ䫭䇃ǄerrorS(h)ⱘᷛ‫ޚ‬ᏂᰃҔМ˛ᇚℸ㒧ᵰϢ 5.3.4 㡖᳿ሒⱘ՟ᄤЁᷛ‫ޚ‬ᏂⳌ↨䕗Ӯᕫߎ
ҔМ㒧䆎˛
5.2 㗗㰥ᶤᏗᇨؐὖᗉЁᄺࠄⱘ‫؛‬䆒hǄᔧh೼ 100 Ͼḋ՟ⱘ䲚ড়Ϟ⌟䆩ᯊˈ᳝ 83 Ͼߚ㉏

ℷ⹂Ǆ䙷Мⳳᅲ䫭䇃⥛errorD(h)ⱘᷛ‫ޚ‬Ꮒ੠ 95ˁ㕂ֵऎ䯈ᰃ໮ᇥ˛
5.3 བᵰ‫؛‬䆒h೼n=65 ⱘ⣀ゟᢑপḋᴀϞߎ⦄r=10 Ͼ䫭䇃Ǆⳳᅲ䫭䇃⥛ⱘ 90ˁ㕂ֵऎ䯈

˄ঠջⱘ˅ᰃ໮ᇥǄ95ˁऩջ㕂ֵऎ䯈˄ेϔϾϞ⬠UˈՓᕫ᳝ 95ˁ㕂ֵᑺerrorD(h)İU˅
ᰃ໮ᇥ˛90ˁऩջऎ䯈ᰃ໮ᇥ˛
5.4 㽕⌟䆩ˉ‫؛‬䆒hˈ݊errorD(h)Ꮖⶹ೼ 0.2 ੠ 0.6 㣗ೈ‫ݙ‬Ǆ㽕ֱ䆕 95ˁঠջ㕂ֵऎ䯈ⱘ

ᆑᑺᇣѢ 0.1ˈ᳔ᇥᑨ᧰䲚ⱘḋ՟᭄ᰃ໮ᇥǄ
5.5 ᇍѢ೼ϡৠ᭄᥂ḋᴀϞ⌟䆩ⱘϸ‫؛‬䆒䫭䇃⥛ⱘᏂˈ㒭ߎ䅵ㅫऩջϞ⬠੠ऩջϟ⬠ⱘ
Nˁ㕂ֵऎ䯈ⱘ䗮⫼㸼䖒ᓣǄ
5.6 㾷䞞ЎҔМᓣ 5.17 㒭ߎⱘ㕂ֵऎ䯈Ԅ䅵ৃ⫼ѢԄ䅵ᓣ 5.16ˈ㗠ϡ㛑Ԅ䅵ᓣ 5.14Ǆ

ㅢ6ㄖ䍓ਬᯥᆜҖ
䋱৊ᮃ᥼⧚ᦤկњ᥼⧚ⱘϔ⾡ὖ⥛᠟↉Ǆᅗ෎Ѣབϟⱘ‫؛‬ᅮˈेᕙ㗗ᶹⱘ䞣䙉ᕾᶤὖ⥛
ߚᏗˈϨৃḍ᥂䖭ѯὖ⥛ঞᏆ㾖ᆳࠄⱘ᭄᥂䖯㸠᥼⧚ˈҹ԰ߎ᳔Ӭⱘ‫އ‬ㄪǄ䋱৊ᮃ᥼⧚ᇍᴎ
఼ᄺдकߚ䞡㽕ˈ಴ЎᅗЎ㸵䞣໮Ͼ‫؛‬䆒ⱘ㕂ֵᑺᦤկњᅮ䞣ⱘᮍ⊩Ǆ䋱৊ᮃ᥼⧚ЎⳈ᥹᪡
԰ὖ⥛ⱘᄺдㅫ⊩ᦤկњ෎⸔ˈ㗠ϨᅗгЎ݊Ҫㅫ⊩ⱘߚᵤᦤկњ⧚䆎ḚᶊǄ
6.1 ԁ㔃
䋱৊ᮃᄺдৠ៥Ӏⱘᴎ఼ᄺдⷨおⳌ݇ˈ᳝ϸϾॳ಴Ǆ佪‫ˈܜ‬䋱৊ᮃᄺдㅫ⊩㛑໳䅵ㅫ
ᰒᓣⱘ‫؛‬䆒ὖ⥛ˈབᴈ㋴䋱৊ᮃߚ㉏఼ˈᅗᰃ㾷‫Ⳍއ‬ᑨᄺд䯂乬ⱘ᳔᳝ᅲ䰙Ӌؐⱘᮍ⊩ПϔǄ
՟བˈMichie et al.˄1994˅䆺㒚ⷨお↨䕗њᴈ㋴䋱৊ᮃߚ㉏఼˄naïve Bayesian classifier˅੠
݊Ҫᄺдㅫ⊩ˈࣙᣀ‫އ‬ㄪᷥ੠⼲㒣㔥㒰ǄҪӀথ⦄ᴈ㋴䋱৊ᮃߚ㉏఼೼໮᭄ᚙ‫މ‬ϟϢ݊Ҫᄺ
дㅫ⊩ᗻ㛑Ⳍᔧˈ೼ᶤѯᚙ‫މ‬ϟ䖬ӬѢ݊Ҫㅫ⊩Ǆᴀゴᦣ䗄њᴈ㋴䋱৊ᮃߚ㉏఼ˈᑊᦤկњ
ϔϾ䆺㒚՟ᄤ˖ेᅗᑨ⫼Ѣ᭛ᴀ᭛ḷߚ㉏ⱘᄺд䯂乬˄བ⬉ᄤᮄ䯏ߚ㉏˅ǄᇍѢ䖭ḋⱘᄺд
ӏࡵˈᴈ㋴䋱৊ᮃߚ㉏ᰃ᳔᳝ᬜⱘㅫ⊩ПϔǄ
䋱৊ᮃᮍ⊩ᇍѢᴎ఼ᄺдⷨおⱘ䞡㽕ᗻ䖬ԧ⦄೼ˈᅗЎ⧚㾷໮᭄ᄺдㅫ⊩ᦤկњϔ⾡᳝
ᬜⱘ᠟↉ˈ㗠䖭ѯㅫ⊩ϡϔᅮⳈ᥹᪡԰ὖ⥛᭄᥂Ǆ՟བˈᴀゴߚᵤњ㄀ 2 ゴⱘ Find-S ੠‫׭‬
䗝⍜䰸ㅫ⊩ˈҹ߸ᮁ೼㒭ᅮ᭄᥂ᯊાϔϾㅫ⊩ᇚ䕧ߎ᳔᳝ৃ㛑ⱘ‫؛‬䆒Ǆ៥Ӏ䖬Փ⫼䋱৊ᮃߚ
ᵤ䆕ᯢњ⼲㒣㔥㒰ᄺдЁⱘϔϾ݇䬂ᗻⱘ䗝ᢽ˖े೼᧰㋶⼲㒣㔥㒰ぎ䯈ᯊˈ䗝ᢽՓ䇃Ꮒᑇᮍ
੠᳔ᇣ࣪ⱘ⼲㒣㔥㒰Ǆ៥Ӏ䖬᥼ᇐߎ঺ϔ⾡䇃Ꮒߑ᭄˖Ѹঝ➉Ǆᅗ೼ᄺд乘⌟ὖ⥛Ⳃᷛߑ᭄
ᯊ↨䇃Ꮒᑇᮍ੠᳈ড়䗖Ǆᴀゴ䖬⫼䋱৊ᮃⱘ᠟↉ߚᵤњ‫އ‬ㄪᷥⱘᔦ㒇‫أ‬㕂˄ेӬ䗝᳔ⷁⱘ‫އ‬
ㄪᷥ˅ˈᑊ㗗ᶹњᆚߛⳌ݇ⱘ᳔ᇣᦣ䗄䭓ᑺ˄Minimum Description Length˅ॳ߭Ǆᇍ䋱৊ᮃ
ᮍ⊩ⱘ෎ᴀњ㾷ˈᇍѢ⧚㾷੠ࠏ⬏ᴎ఼ᄺдЁ䆌໮ㅫ⊩ⱘ᪡԰ᕜ䞡㽕Ǆ
䋱৊ᮃᄺдᮍ⊩ⱘ⡍ᗻࣙᣀ˖
x 㾖ᆳࠄⱘ↣Ͼ䆁㒗ḋ՟ৃҹ๲䞣ᓣഄ䰡Ԣ៪छ催ᶤ‫؛‬䆒ⱘԄ䅵ὖ⥛Ǆ䖭ᦤկњ
ϔ⾡↨݊Ҫㅫ⊩᳈ড়⧚ⱘᄺд䗨ᕘǄ݊Ҫㅫ⊩Ӯ೼ᶤϾ‫؛‬䆒Ϣӏϔḋ՟ϡϔ㟈
ᯊᅠܼএᥝ䆹‫؛‬䆒Ǆ
x ‫ܜ‬偠ⶹ䆚ৃҹϢ㾖ᆳ᭄᥂ϔ䍋‫އ‬ᅮ‫؛‬䆒ⱘ᳔㒜ὖ⥛Ǆ೼䋱৊ᮃᄺдЁˈ‫ܜ‬偠ⶹ
䆚ⱘᔶᓣৃҹᰃ˄1˅↣Ͼ‫׭‬䗝‫؛‬䆒ⱘ‫ܜ‬偠ὖ⥛˄2˅↣Ͼৃ㛑‫؛‬䆒೼ৃ㾖ᆳ᭄
᥂Ϟⱘὖ⥛ߚᏗǄ
x 䋱৊ᮃᮍ⊩ৃ‫ܕ‬䆌‫؛‬䆒‫ߎخ‬ϡ⹂ᅮᗻⱘ乘⌟Ǆ˄↨བ䖭ḋⱘ‫؛‬䆒˖䖭ϔ㚎♢⮙
Ҏ᳝ 93%ⱘᴎӮᒋ໡˅Ǆ
x ᮄⱘᅲ՟ߚ㉏ৃ⬅໮Ͼ‫؛‬䆒ϔ䍋԰ߎ乘⌟ˈҹᅗӀⱘὖ⥛Ўᴗ䞡Ǆ
x ेՓ೼䋱৊ᮃᮍ⊩䅵ㅫ໡ᴖᑺ䕗催ᯊˈᅗӀҡৃ‫خ‬ЎϔϾ᳔Ӭⱘ‫އ‬ㄪⱘᷛ‫ޚ‬㸵
䞣݊Ҫᮍ⊩Ǆ
೼ᅲ䏉Ёᑨ⫼䋱৊ᮃᮍ⊩ⱘ䲒ᑺПϔ೼ѢˈᅗӀ䳔㽕ὖ⥛ⱘ߱ྟⶹ䆚Ǆᔧ䖭ὖ⥛乘‫ܜ‬᳾
ⶹᯊˈৃҹ෎Ѣ㚠᱃ⶹ䆚ǃ乘‫ޚܜ‬໛དⱘ᭄᥂ҹঞ݇Ѣ෎‫ߚޚ‬Ꮧⱘ‫؛‬ᅮᴹԄ䅵䖭ѯὖ⥛Ǆ঺
ϔᅲ䰙ೄ䲒೼Ѣˈϔ㠀ᚙ‫މ‬ϟ⹂ᅮ䋱৊ᮃ᳔Ӭ‫؛‬䆒ⱘ䅵ㅫҷӋ↨䕗໻˄ৠ‫׭‬䗝‫؛‬䆒ⱘ᭄䞣៤
㒓ᗻ݇㋏˅Ǆ೼ᶤѯ⡍ᅮᚙᔶϟˈ䖭⾡䅵ㅫҷӋৃҹ㹿໻໻䰡ԢǄ
ᴀゴ࠽ԭ䚼ߚⱘ㒘៤བϟǄ6.2 㡖ҟ㒡њ䋱৊ᮃ⧚䆎ˈᑊᅮНњᵕ໻Ԑ✊˄maximum
likelihood˅‫؛‬䆒੠ᵕ໻ৢ偠ὖ⥛‫؛‬䆒˄maximum a posteriori probability hypotheses˅Ǆ᥹ϟᴹ
ⱘಯ㡖ᇚℸὖ⥛Ḛᶊᑨ⫼Ѣߚᵤࠡ䴶ゴ㡖ⱘⳌ݇䯂乬੠ᄺдㅫ⊩Ǆ՟བˈ៥Ӏ䆕ᯢњ೼⡍ᅮ
ࠡᦤϟˈ޴Ͼࠡ䗄ⱘㅫ⊩㛑䕧ߎᵕ໻Ԑ✊‫؛‬䆒Ǆ࠽ԭⱘ޴㡖߭ҟ㒡њ޴⾡Ⳉ᥹᪡԰ὖ⥛ⱘᄺ
дㅫ⊩Ǆࣙᣀ䋱৊ᮃ᳔Ӭߚ㉏఼ǃGibbs ㅫ⊩੠ᴈ㋴䋱৊ᮃߚ㉏఼Ǆ᳔ৢˈ៥Ӏ䅼䆎њ䋱৊
ᮃ㕂ֵ㔥ˈᅗᰃϔ⾡෎Ѣὖ⥛᥼⧚ⱘ䕗ᮄⱘᄺдᮍ⊩˗ҹঞ EM ㅫ⊩ˈᰃᔧᄬ೼᳾㾖⌟ࠄব
䞣ᯊᑓ⊯Փ⫼ⱘᄺдㅫ⊩Ǆ
6.2 䍓ਬᯥ⌋ࡏ
೼ᴎ఼ᄺдЁˈ䗮ᐌ៥Ӏᛳ݈䍷ⱘᰃ೼㒭ᅮ䆁㒗᭄᥂ D ᯊˈ⹂ᅮ‫؛‬䆒ぎ䯈 H Ёⱘ᳔Շ

‫؛‬䆒Ǆ᠔䇧᳔Շ‫؛‬䆒ˈϔ⾡ࡲ⊩ᰃᡞᅗᅮНЎ೼㒭ᅮ᭄᥂ D ҹঞ H Ёϡৠ‫؛‬䆒ⱘ‫ܜ‬偠ὖ⥛
ⱘ᳝݇ⶹ䆚ᴵӊϟⱘ᳔ৃ㛑˄most probable˅‫؛‬䆒Ǆ䋱৊ᮃ⧚䆎ᦤկњ䅵ㅫ䖭⾡ৃ㛑ᗻⱘϔ
⾡Ⳉ᥹ⱘᮍ⊩Ǆ᳈㊒⹂ഄ䆆ˈ䋱৊ᮃ⊩߭ᦤկњϔ⾡䅵ㅫ‫؛‬䆒ὖ⥛ⱘᮍ⊩ˈᅗ෎Ѣ‫؛‬䆒ⱘ‫ܜ‬
偠ὖ⥛ǃ㒭ᅮ‫؛‬䆒ϟ㾖ᆳࠄϡৠ᭄᥂ⱘὖ⥛ǃҹঞ㾖ᆳⱘ᭄᥂ᴀ䑿Ǆ
㽕㊒⹂ഄᅮН䋱৊ᮃ⧚䆎ˈ‫ܜ‬ᓩܹϔѯ䆄োǄ៥Ӏ⫼ P(h)ᴹҷ㸼䖬≵᳝䆁㒗᭄᥂ࠡˈ‫؛‬
䆒 h ᢹ᳝ⱘ߱ྟὖ⥛ǄP(h)ᐌ㹿⿄Ў h ⱘ‫ܜ‬偠ὖ⥛˄prior probability ˅ˈᅗড᯴њ៥Ӏ᠔ᢹ
᳝ⱘ݇Ѣ h ᰃϔℷ⹂‫؛‬䆒ⱘᴎӮⱘ㚠᱃ⶹ䆚Ǆབᵰ≵᳝䖭ϔ‫ܜ‬偠ⶹ䆚ˈ䙷Мৃҹㅔऩഄᇚ↣
ϔ‫׭‬䗝‫؛‬䆒䌟ќⳌৠⱘ‫ܜ‬偠ὖ⥛ǄⳌԐഄˈৃ⫼ P(D)ҷ㸼ᇚ㽕㾖ᆳⱘ䆁㒗᭄᥂ D ⱘ‫ܜ‬偠ὖ
⥛˄ᤶ㿔Пˈ೼≵᳝⹂ᅮᶤϔ‫؛‬䆒៤ゟᯊˈD ⱘὖ⥛˅Ǆϟϔℹˈҹ P(D|h)ҷ㸼‫؛‬䆒 h ៤ゟ
ⱘᚙᔶϟ㾖ᆳࠄ᭄᥂ D ⱘὖ⥛Ǆ᳈ϔ㠀ഄˈ៥ӀՓ⫼ P(x|y)ҷ㸼㒭ᅮ y ᯊ x ⱘὖ⥛Ǆ೼ᴎ఼
ᄺдЁˈ៥Ӏᛳ݈䍷ⱘᰃ P(h|D)ˈे㒭ᅮ䆁㒗᭄᥂ D ᯊ h ៤ゟⱘὖ⥛ǄP(h|D)㹿⿄Ў h ⱘৢ
偠ὖ⥛˄posterior probability˅ˈ಴Ўᅗড᯴њ೼ⳟࠄ䆁㒗᭄᥂ D ৢ h ៤ゟⱘ㕂ֵᑺǄᑨ⊼ᛣˈ
ৢ偠ὖ⥛ P(h|D)ড᯴њ䆁㒗᭄᥂ D ⱘᕅડ˗Ⳍডˈ‫ܜ‬偠ὖ⥛ P(h)ᰃ⣀ゟѢ D ⱘǄ
䋱৊ᮃ⊩߭ᰃ䋱৊ᮃᄺдᮍ⊩ⱘ෎⸔ˈ಴ЎᅗᦤկњҢ‫ܜ‬偠ὖ⥛ P(h)ҹঞ P(D)੠ P(D|h)

䅵ㅫৢ偠ὖ⥛ P(h|D)ⱘᮍ⊩Ǆ
䋱৊ᮃ݀ᓣ
P ( D | h) P ( h)
P(h | D) ˄6.1˅
P( D)
Ⳉ㾖ৃⳟߎˈP(h|D)䱣ⴔ P(h)੠ P(D|h)ⱘ๲䭓㗠๲䭓Ǆৠᯊгৃⳟߎ P(h|D)䱣 P(D)ⱘ๲

ࡴ㗠‫ޣ‬ᇥˈ䖭ᰃᕜড়⧚ⱘˈ಴Ўབᵰ D ⣀ゟѢ h 㹿㾖ᆳࠄⱘৃ㛑ᗻ䍞໻ˈ䙷М D ᇍ h ⱘᬃ
ᣕᑺ䍞ᇣǄ
೼䆌໮ᄺдഎ᱃Ёˈᄺд఼㗗㰥‫׭‬䗝‫؛‬䆒䲚ড়Hᑊ೼݊Ёᇏᡒ㒭ᅮ᭄᥂Dᯊৃ㛑ᗻ᳔໻
ⱘ‫؛‬䆒hęH˄៪㗙ᄬ೼໮Ͼ䖭ḋⱘ‫؛‬䆒ᯊ䗝ᢽ݊ЁПϔ˅Ǆ䖭ḋⱘ‫݋‬᳔᳝໻ৃ㛑ᗻⱘ‫؛‬䆒㹿
⿄Ўᵕ໻ৢ偠˄maximum a posteriori, MAP˅‫؛‬䆒Ǆ⹂ᅮMAP‫؛‬䆒ⱘᮍ⊩ᰃ⫼䋱৊ᮃ݀ᓣ䅵
ㅫ↣Ͼ‫׭‬䗝‫؛‬䆒ⱘৢ偠ὖ⥛Ǆ᳈㊒⹂ഄ䇈ᔧϟᓣ៤ゟᯊˈ⿄hMAPЎüMAP‫؛‬䆒˖
hMAP { arg max P(h | D)

hH
P ( D | h) P ( h)
arg max
hH P( D)
arg max P( D | h) P(h) ˄6.2˅
hH
⊼ᛣ೼᳔ৢϔℹ៥Ӏএᥝњ P(D)ˈ಴Ўᅗᰃϡձ䌪Ѣ h ⱘᐌ䞣Ǆ
೼ ᶤ ѯ ᚙ ‫ މ‬ϟ ˈ ৃ ‫ ؛‬ᅮ H Ё ↣ Ͼ ‫ ؛‬䆒 ᳝ Ⳍ ৠ ⱘ ‫ ܜ‬偠 ὖ ⥛ ˄ े ᇍ H Ё ӏ ᛣ hi ੠ hj ˈ

P(hi)=P(hj)˅Ǆ䖭ᯊৃᡞㄝᓣ 6-2 䖯ϔℹㅔ࣪ˈা䳔㗗㰥P(D|h)ᴹᇏᡒᵕ໻ৃ㛑‫؛‬䆒ǄP(D|h)
ᐌ⿄Ў㒭ᅮ h ᯊ᭄᥂Dⱘ Ԑ✊ᑺ˄likelihood˅ˈ㗠ՓP(D|h)᳔໻ⱘ‫؛‬䆒㹿⿄Ўᵕ໻ Ԑ✊
˄maximum likelihoodˈML˅‫؛‬䆒hMLǄ
hML { arg max P( D | h) ˄6.3˅

hH
ЎњՓϞ䴶ⱘ䅼䆎Ϣᴎ఼ᄺд䯂乬Ⳍ㘨㋏ˈ៥Ӏᡞ᭄᥂ D ⿄԰ᶤⳂᷛߑ᭄ⱘ䆁㒗ḋ՟ˈ
㗠ᡞ H ⿄Ў‫׭‬䗝Ⳃᷛߑ᭄ぎ䯈Ǆᅲ䰙Ϟˈ䋱৊ᮃ݀ᓣ᳝ⴔ᳈Ў᱂䘡ⱘᛣНǄᅗৠḋৃҹᕜ
དഄ⫼ѢӏᛣѦ᭹ੑ乬ⱘ䲚ড় Hˈা㽕䖭ѯੑ乬ⱘὖ⥛П੠Ў 1˄՟བ˖Ā໽ぎᰃ݄㡆ⱘā
੠Ā໽ぎϡᰃ݄㡆ⱘā˅ǄᴀゴЁ᳝ᯊᇚ H ԰Ўࣙ৿Ⳃᷛߑ᭄ⱘ‫؛‬䆒ぎ䯈ˈ㗠 D ԰Ў䆁㒗՟
䲚ড়Ǆ݊Ҫϔѯᯊ‫׭‬㗗㰥ᇚ H ⳟ԰ϔѯѦ᭹ੑ乬ⱘ䲚ড়ˈ㗠 D Ўᶤ⾡᭄᥂Ǆ
6.2.1 ⽰ׁ
Ў䇈ᯢ䋱৊ᮃ㾘߭ˈৃ㗗㰥ϔए⭫䆞ᮁ䯂乬ˈ݊Ё᳝ϸϾৃ䗝ⱘ‫؛‬䆒˖ ˄1˅⮙Ҏ᳝ᶤ⾡
㉏ൟⱘⰠ⮛ˈ ˄2˅⮙Ҏ᮴Ⱐ⮛Ǆৃ⫼ⱘ᭄᥂ᴹ㞾Ѣϔ࣪偠⌟䆩ˈᅗ᳝ϸ⾡ৃ㛑ⱘ䕧ߎ˖˄ℷ˅
੠ ˄䋳˅Ǆ៥Ӏ᳝‫ܜ‬偠ⶹ䆚˖೼᠔᳝ҎষЁা᳝ 0.008 ⱘҎᙷ᳝䆹⮒⮙Ǆ঺໪ˈ䆹࣪偠⌟
䆩াᰃ䆹⮙ⱘϔϾϡᅠܼⱘ乘䅵Ǆ䆹⌟䆩䩜ᇍ⹂ᅲ᳝⮙ⱘᙷ㗙᳝ 98%ⱘৃ㛑䖨ಲℷ⹂ⱘ㒧
ᵰˈ㗠ᇍ᮴䆹⮙ⱘᙷ㗙᳝ 97%ⱘৃ㛑ℷ⹂䖨ಲ 㒧ᵰǄ䰸ℸҹ໪ˈ⌟䆩䖨ಲⱘ㒧ᵰᰃ䫭䇃
ⱘǄϞ䴶ⱘᚙ‫⬅ৃމ‬ҹϟⱘὖ⥛ᓣὖᣀ˖
P (cancer ) 0.008 , P (cancer ) 0.992

P ( | cancer ) 0.98 , P (4 | cancer ) 0.02
P ( | cancer ) 0.03 , P (4 | cancer ) 0.97
‫؛‬ᅮ⦄᳝ϔᮄ⮙Ҏˈ࣪偠⌟䆩䖨ಲњ㒧ᵰǄᰃ৺ᑨᇚ⮙ҎᮁᅮЎ᳝Ⱐ⮛ਸ਼˛ᵕ໻ৢ偠
‫؛‬䆒ৃ⫼ᓣ 6.2 ᴹ䅵ㅫ˖
P ( | cancer ) P (cancer ) (0.98) (0.008) 0.0078

P ( | cancer ) P (cancer ) (0.03) (0.992) 0.0298
಴ℸˈhMAP=cancerǄ⹂ߛⱘৢ偠ὖ⥛ৃᇚϞ䴶ⱘ㒧ᵰᔦϔ࣪ҹՓᅗӀⱘ੠Ўü ˄े
0.0078
P(cancer|)= =0.21˅Ǆ䆹ℹ偸ⱘḍ᥂೼Ѣˈ䋱৊ᮃ݀ᓣ䇈ᯢৢ偠ὖ⥛ህᰃ
0.0078 0.0298
Ϟ䴶ⱘ䞣䰸ҹ᭄᥂P()Ǆ㱑✊P()≵᳝԰Ў䯂乬䰜䗄ⱘϔ䚼ߚⳈ᥹㒭ߎˈԚ಴ЎᏆⶹ
ˈ಴ℸৃҹ䖯㸠ᔦ
P(cancer|)੠P(cancer|)ⱘ੠ᖙЎ 1˄े䆹⮙Ҏ㽕М᳝Ⱐ⮛ˈ㽕М≵᳝˅
ϔ࣪Ǆ⊼ᛣ㱑✊᳝Ⱐ⮛ⱘৢ偠ὖ⥛↨‫ܜ‬偠ὖ⥛㽕໻ˈԚ᳔ৃ㛑ⱘ‫؛‬䆒ҡЎℸҎ≵᳝Ⱐ⮛Ǆ
བϞ՟᠔⼎ˈ䋱৊ᮃ᥼⧚ⱘ㒧ᵰᕜ໻ഄձ䌪Ѣ‫ܜ‬偠ὖ⥛ˈ㽕Ⳉ᥹ᑨ⫼䆹ᮍ⊩ᖙ乏‫ܜ‬㦋প
䆹ؐǄ䖬㽕⊼ᛣ䆹՟Ёᑊ≵᳝ᅠܼഄ㹿᥹ফ៪ᢦ㒱‫؛‬䆒ˈ㗠াᰃ೼㾖ᆳࠄ䕗໮ⱘ᭄᥂ৢ‫؛‬䆒
ⱘৃ㛑ᗻ๲໻៪‫ޣ‬ᇣњǄ
䅵ㅫὖ⥛ⱘ෎ᴀ݀ᓣ೼㸼 6-1 Ё߫ВǄ
Ь⊩݀ᓣ(Product rule)˖ϸџӊ A ੠ B ⱘѸⱘὖ⥛ P(AB)

P( A B) P(A | B)P(B) P(B | A)P(A)
ࡴ⊩݀ᓣ(Sum Rule)˖ϸџӊ A ੠ B ⱘᑊⱘὖ⥛ P(AB)
P( A B) P(A) P(B) - P(A B)
䋱৊ᮃ⊩߭(Bayes theorem)˖㒭ᅮ D ᯊ h ⱘৢ偠ὖ⥛ P(h|D)
P ( D | h) P ( h)
P(h | D)
P( D)
¦
n
ܼὖ⥛݀ᓣ(Theorem of total probability)˖བᵰџӊA1, …, AnѦ᭹Ϩ P( Ai ) 1 ˈ߭˖
i 1
n
P( B) ¦ P( B | A ) P( A )
i 1
i i
㸼 6-1 ෎ᴀὖ⥛݀ᓣ㸼
6.3 䍓ਬᯥ⌋ࡏૂᾸᘫᆜҖ
䋱৊ᮃ⊩߭੠ὖᗉᄺд䯂乬ⱘ݇㋏ᰃҔМ˛಴Ў䋱৊ᮃ⊩߭Ў䅵ㅫ㒭ᅮ䆁㒗᭄᥂ϟӏ
ϔ‫؛‬䆒ⱘৢ偠ὖ⥛ᦤկњॳ߭ᗻᮍ⊩ˈ៥ӀৃⳈ᥹ᇚ݊԰ЎϔϾ෎ᴀⱘᄺдㅫ⊩˖䅵ㅫ↣Ͼ
‫؛‬䆒ⱘὖ⥛ˈ‫ݡ‬䕧ߎ݊Ёὖ⥛᳔໻ⱘǄᴀ㡖㗗㰥њ䖭ḋϔϾ Brute-Force 䋱৊ᮃὖᗉᄺдㅫ
⊩ˈ✊ৢᇚ݊Ϣ㄀ 2 ゴҟ㒡ⱘὖᗉᄺдㅫ⊩Ⳍ↨䕗Ǆ䗮䖛↨䕗ৃҹⳟࠄϔϾ᳝䍷ⱘ㒧䆎ˈे
೼⡍ᅮᴵӊϟˈࠡ䴶ᦤࠄⱘ޴⾡ㅫ⊩䛑䕧ߎϢ Brute-Force 䋱৊ᮃㅫ⊩Ⳍৠⱘ‫؛‬䆒ˈাϡ䖛
ࠡ䴶ⱘㅫ⊩ϡᯢ⹂䅵ㅫὖ⥛ˈ㗠Ϩ೼Ⳍᔧ⿟ᑺϞᬜ⥛᳈催Ǆ
6.3.1 Brute-Force 䍓ਬᯥᾸᘫᆜҖ
㗗㰥㄀ 2 ゴ佪‫ⱘࠄᦤܜ‬ὖᗉᄺд䯂乬Ǆ݊Ёˈ៥Ӏ‫؛‬ᅮᄺд఼㗗㰥ⱘᰃᅮН೼ᅲ՟ぎ䯈
XϞⱘ᳝䰤ⱘ‫؛‬䆒ぎ䯈HˈӏࡵᰃᄺдᶤϾⳂᷛὖᗉc:Xė{0,1}Ǆབ䗮ᐌ䙷ḋˈ‫؛‬ᅮ㒭ќᄺд
఼ᶤ䆁㒗ḋ՟ᑣ߫ǇǇx1ˈd1ˈǈĂǇxmˈdmǈǈˈ݊ЁxiЎXЁⱘᶤᅲ՟ˈdiЎxiⱘⳂᷛߑ᭄ؐ
˄ेdi=c(xi)˅
ǄЎㅔ࣪䅼䆎ˈ‫؛‬ᅮᅲ՟ᑣ߫Ǉx1Ăxmǈᰃ೎ᅮϡবⱘˈ಴ℸ䆁㒗᭄᥂Dৃ㹿ㅔ
ऩഄ‫ݭ‬԰Ⳃᷛߑ᭄ؐᑣ߫˖D=Ǉd1ĂdmǈǄৃҹⳟࠄ˄㾕д乬 6.4˅ˈ䖭ϔㅔ࣪ϡӮᬍবᴀ㡖
ⱘЏ㽕㒧䆎Ǆ
෎Ѣ䋱৊ᮃ⧚䆎ⱘⳈ᥹ὖᗉᄺдㅫ⊩ᅮНབϟ˖
Brute-Force MAP ᄺдㅫ⊩
1ˊᇍѢ H Ё↣Ͼ‫؛‬䆒 hˈ䅵ㅫৢ偠ὖ⥛˖

P ( D | h) P ( h)
P(h | D)
P( D)
2ˊ䕧ߎ᳔᳝催ৢ偠ὖ⥛ⱘ‫؛‬䆒hMAP
hMAP { arg max P(h | D)

hH
ℸㅫ⊩䳔㽕䕗໻ⱘ䅵ㅫ䞣ˈ಴Ўᅗᇍ H Ё↣Ͼ‫؛‬䆒䛑ᑨ⫼њ䋱৊ᮃ݀ᓣҹ䅵ㅫ P(h|D)Ǆ

㱑✊ᇍѢ໻ⱘ‫؛‬䆒ぎ䯈䖭ᕜϡߛᅲ䰙ˈԚ䆹ㅫ⊩ҡ✊ؐᕫ݇⊼ˈ಴ЎᅗᦤկњϔϾᷛ‫ˈޚ‬ҹ
߸ᮁ݊Ҫὖᗉᄺдㅫ⊩ⱘᗻ㛑Ǆ
ϟ䴶Ў Brute-Force MAP ᄺдㅫ⊩ᣛᅮϔᄺд䯂乬ˈ៥Ӏᖙ乏⹂ᅮ P(h)੠ P(D|h)ߚ߿ᑨ

পԩؐ˄ৃҹⳟߎˈP(D)ⱘؐӮձ䖭ϸ㗙㗠ᅮ˅Ǆ៥Ӏৃҹҹӏᛣᮍ⊩䗝ᢽ P(h)੠ P(D|h)ⱘ
ὖ⥛ߚᏗˈҹᦣ䗄䆹ᄺдӏࡵⱘ‫ܜ‬偠ⶹ䆚Ǆ䖭䞠Ҹ݊Ϣϟ䴶ⱘࠡᦤϔ㟈˖
˗
1ˊ䆁㒗᭄᥂Dᰃ᮴ాໄⱘ˄ेdi=c(xi)˅
2ˊⳂᷛὖᗉ c ࣙ৿೼‫؛‬䆒ぎ䯈 H Ё˗
3ˊ≵᳝ӏԩ⧚⬅䅸Ўᶤ‫؛‬䆒↨݊Ҫⱘ‫؛‬䆒ⱘৃ㛑ᗻ໻Ǆ
᳝њ䖭ѯ‫؛‬ᅮˈབԩ⹂ᅮ P(h)ⱘؐ˛⬅Ѣӏϔ‫؛‬䆒ϡ↨݊Ҫ‫؛‬䆒ৃ㛑ᗻ໻ˈᕜᰒ✊ৃᇍ
H Ё↣Ͼ‫؛‬䆒 h 䌟ҹⳌৠⱘ‫ܜ‬偠ὖ⥛Ǆ䖯ϔℹഄˈ⬅ѢⳂᷛὖᗉ೼ H Ёˈ᠔ҹৃ㽕∖ H Ё
᠔᳝‫؛‬䆒ⱘὖ⥛੠Ў 1Ǆᇚ䖭ѯ䰤ࠊড়䍋ᴹৃᕫ˖
1
ᇍ H Ёӏϔ h ˈ P ( h)
H
བԩ䗝ᢽP(D|h)ⱘؐ˛P(D|h)ᰃᏆⶹ‫؛‬䆒h៤ゟⱘᴵӊϟ˄ेᏆⶹhЎⳂᷛὖᗉcⱘℷ⹂ᦣ
䗄˅ˈ㾖ᆳࠄⳂᷛؐD=Ǉd1…dmǈⱘὖ⥛Ǆ⬅Ѣ‫؛‬ᅮ䆁㒗᭄᥂᮴ాໄˈ䙷М㒭ᅮhᯊˈབᵰ
di=h(xi)߭ߚ㉏diЎ 1ˈབᵰdiĮh(xi)߭diЎ 0Ǆ಴ℸ˖
1 བᵰᇍ D Ё᠔᳝ diˈdi=h(xi)

P(D|h)= ˄6.4˅
0 ݊Ҫᚙ‫މ‬
ᤶ㿔Пˈ㒭ᅮ‫؛‬䆒 hˈ᭄᥂ D ⱘὖ⥛೼݊Ϣ‫؛‬䆒 h ϔ㟈ᯊؐЎ 1ˈ৺߭ؐЎ 0Ǆ
᳝њ P(h)੠ P(D|h)ⱘؐˈ⦄೼៥ӀᇍѢϞ䴶ⱘ Brute-Force MAP ᄺдㅫ⊩᳝њϔϾᅠᭈ

ᅮНⱘ䯂乬Ǆ᥹ϟᴹ㗗㰥䆹ㅫ⊩ⱘ㄀ϔℹˈՓ⫼䋱৊ᮃ݀ᓣ䅵ㅫ↣Ͼ‫؛‬䆒 h ⱘৢ偠ὖ⥛
P(h|D)˖
P ( D | h) P ( h)
P(h | D)
P( D)
佪‫ܜ‬㗗㰥 h Ϣ䆁㒗᭄᥂ D ϡϔ㟈ⱘᚙᔶǄ⬅Ѣᓣ 6.4 ᅮНᔧ h Ϣ D ϡϔ㟈ᯊ P(D|h)Ў 0ˈ

᳝˖
0 P ( h)
P(h | D) 0 ˈᔧ h Ϣ D ϡϔ㟈Ǆ
P( D)
Ϣ D ϡϔ㟈ⱘ‫؛‬䆒 h ⱘৢ偠ὖ⥛Ў 0Ǆ
‫ݡ‬㗗㰥 h Ϣ D ϔ㟈ⱘᚙ‫މ‬Ǆ⬅Ѣᓣ 6.4 ᅮНᔧ h Ϣ D ϔ㟈ᯊ P(D |h)Ў 1ˈ᳝˖

1
1
H
P(h|D)
P( D)
1
1
H
VS H , D
H
1
ᔧ h Ϣ D ϔ㟈
VS H , D
݊ЁVSH,DᰃHЁϢDϔ㟈ⱘ‫؛‬䆒ᄤ䲚˄ेVSH,DᰃⳌᇍѢDⱘবൟぎ䯈ˈབ㄀ 2 ゴⱘᅮН˅Ǆ
VS H , D
ᕜᆍᯧৃ偠䆕P(D)= ˈ಴Ў೼᠔᳝‫؛‬䆒ϞP(h|D)ⱘ੠ᖙЎ 1ˈᑊϨHЁϢDϔ㟈ⱘ‫؛‬
H
䆒᭄䞣Ў|VSH,D|Ǆ঺໪ˈৃҢܼὖ⥛݀ᓣ˄㾕㸼 6-1˅ҹঞ᠔᳝‫؛‬䆒ᰃѦ᭹ⱘᴵӊ˄े
(i z j )( P(hi h j ) 0) ˅ˈ᥼ᇐߎP(D)ⱘؐ˖
P(D) ¦ P( D | h ) P (h )
hi H
i i
1 1
¦
hi VS H , D
1 ¦ 0
H hi VS H ,D H
1
¦
hi VS H , D
1
H
VS H , D
H
ὖ㗠㿔Пˈ䋱৊ᮃ݀ᓣ䇈ᯢ೼៥Ӏⱘ P(h)੠ P(D|h)ⱘᅮНϟˈৢ偠ὖ⥛ P(h|D)Ў˖

1
བᵰ h Ϣ D ϔ㟈
P(h|D)= VS H , D ˄6.5˅
0 ݊Ҫᚙ‫މ‬
݊Ё|VSH,D|ᰃHЁϢDϔ㟈ⱘ‫؛‬䆒᭄䞣Ǆ‫؛‬䆒ⱘὖ⥛ⓨ࣪ᚙ‫މ‬བ೒ 6-1 Ё᠔⼎Ǆ߱ྟᯊ˄೒
6-1a˅᠔᳝‫؛‬䆒‫݋‬᳝Ⳍৠⱘὖ⥛Ǆᔧ䆁㒗᭄᥂䗤ℹߎ⦄ৢˈ˄೒ 6-1b੠ 6-1c˅ˈϡϔ㟈‫؛‬䆒ⱘ
ὖ⥛বЎ 0ˈ㗠ᭈϾὖ⥛ⱘ੠ҡЎ 1ˈᅗӀഛࣔഄߚᏗࠄ࠽ԭⱘϔ㟈‫؛‬䆒ЁǄ
Ϟ䴶ⱘߚᵤ䇈ᯢˈ೼៥Ӏ䗝ᅮⱘP(h)੠P(D|h)পؐϟˈ↣Ͼϔ㟈ⱘ‫؛‬䆒ৢ偠ὖ⥛Ў
˄1/|VSH,D|˅
ˈ↣Ͼϡϔ㟈‫؛‬䆒ৢ偠ὖ⥛Ў 0Ǆ಴ℸˈ↣Ͼϔ㟈ⱘ‫؛‬䆒䛑ᰃMAP‫؛‬䆒Ǆ
原书页码： 162
hypotheses: ‫؛‬䆒
೒ 6-1 ৢ偠ὖ⥛䱣ⴔ䆁㒗᭄᥂๲䭓ⱘⓨ࣪
(a)ᇍ↣Ͼ‫؛‬䆒䌟ќഛࣔⱘ‫ܜ‬偠ὖ⥛Ǆᔧ䆁㒗᭄᥂佪‫ܜ‬๲䭓ࠄ D1(b)ˈ✊ৢ๲䭓ࠄ D1ġD2(c)ˈ

ϡϔ㟈‫؛‬䆒ⱘৢ偠ὖ⥛ব៤ 0ˈ㗠ֱ⬭೼বൟぎ䯈Ёⱘ‫؛‬䆒ⱘৢ偠ὖ⥛๲ࡴǄ
6.3.2 MAP ‫ٽ‬䇴ૂж㠪ᆜҖಞ
Ϟ䴶ⱘߚᵤ䇈ᯢ೼㒭ᅮᴵӊϟˈϢDϔ㟈ⱘ↣Ͼ‫؛‬䆒䛑ᰃMAP‫؛‬䆒Ǆḍ᥂䖭ϔ㒧䆎ৃⳈ
᥹ᕫࠄϔ㉏᱂䘡ⱘᄺд఼ˈ⿄Ўϔ㟈ᄺд఼Ǆᶤᄺдㅫ⊩㹿⿄Ўϔ㟈ᄺд఼ˈ䇈ᯢᅗ䕧ߎⱘ
‫؛‬䆒೼䆁㒗՟Ϟ᳝䳊䫭䇃⥛Ǆ⬅ҹϞⱘߚᵤৃᕫˈབᵰ‫؛‬ᅮHϞ᳝ഛࣔⱘ‫ܜ‬偠ὖ⥛˄े
P(hi)=P(hj)ˈᇍ᠔᳝ⱘi,j˅
ˈϨ䆁㒗᭄᥂ᰃ⹂ᅮᗻⱘ੠᮴ాໄⱘ˄ेᔧD੠hϔ㟈ᯊˈP(D|h)=1˅
ˈ
৺߭Ў 0˅ᯊˈӏᛣϔ㟈ᄺд఼ᇚ䕧ߎϔϾMAP‫؛‬䆒Ǆ
՟བ㄀ 2 ゴ䅼䆎ⱘ Find-S ὖᗉᄺдㅫ⊩ǄFind-S ᣝ✻⡍⅞ࠄϔ㠀ⱘ乎ᑣ᧰㋶‫؛‬䆒ぎ䯈 Hˈ

ᑊ䕧ߎϔϾᵕ໻⡍⅞ᗻⱘϔ㟈‫؛‬䆒ˈৃⶹ೼Ϟ䴶ᅮНⱘ P(h)੠ P(D|h)ὖ⥛ߚᏗϟˈᅗ䕧ߎ
MAP ‫؛‬䆒Ǆᔧ✊ˈFind-S ᑊϡⳈ᥹᪡԰ὖ⥛ˈᅗাㅔऩഄ䕧ߎবൟぎ䯈ⱘᵕ໻⡍⅞ᗻ៤ਬǄ
✊㗠ˈ䗮䖛‫އ‬ᅮ P(h)੠ P(D|h)ⱘߚᏗˈҹՓ݊䕧ߎЎ MAP ‫؛‬䆒ˈ៥Ӏ᳝њϔ⾡ࠏ⬏ Find-S
ㅫ⊩ⱘ᳝ᬜ䗨ᕘǄ
ᰃ৺䖬᳝݊Ҫৃ㛑ⱘP(h)੠P(D|h)ߚᏗˈՓFind-S䕧ߎMAP‫؛‬䆒˛ಲㄨᰃ㚃ᅮⱘǄ಴Ў
Find-SҢবൟぎ䯈Ё䕧ߎᵕ໻⡍⅞ᗻ(maximally specific)‫؛‬䆒ˈ᠔ҹᇍѢ‫ܜ‬偠ὖ⥛‫أ‬㹦Ѣ᳈⡍
⅞‫؛‬䆒ⱘӏԩὖ⥛ߚᏗˈᅗ䕧ߎⱘ‫؛‬䆒䛑ᇚᰃMAP‫؛‬䆒Ǆ᳈㊒⹂ഄ䆆ˈ‫؛‬བHᰃHϞӏᛣὖ
⥛ߚᏗP(h)ˈᅗ೼h1↨h2᳈⡍⅞ᯊ䌟ќP(h1)ıP(h2)Ǆৃ㾕ˈ೼‫؛‬ᅮ᳝‫ܜ‬偠ߚᏗH੠ϢϞ䴶Ⳍ
ৠⱘP(D|h)ߚᏗᯊˈFind-S䕧ߎϔMAP‫؛‬䆒Ǆ
ὖᣀҹϞ䅼䆎ˈ䋱৊ᮃḚᶊᦤߎњϔ⾡ࠏ⬏ᄺдㅫ⊩˄བ Find-S ㅫ⊩˅㸠Ўⱘᮍ⊩ˈ

ेՓ䆹ᄺдㅫ⊩ϡ䖯㸠ὖ⥛᪡԰Ǆ䗮䖛⹂ᅮㅫ⊩䕧ߎ᳔Ӭ˄བ MAP˅‫؛‬䆒ᯊՓ⫼ⱘὖ⥛ߚ
Ꮧ P(h)੠ P(D|h)ˈৃҹࠏ⬏ߎㅫ⊩‫݋‬᳔᳝Ӭ㸠Ўᯊⱘ䱤৿‫؛‬ᅮǄ
Փ⫼䋱৊ᮃⱘ䗨ᕘࠏ⬏ᄺдㅫ⊩ˈϢᧁ⼎ᄺд఼Ёⱘᔦ㒇‫أ‬㕂೼ᗱᛇϞᰃⳌ㉏ԐⱘǄ⊼
ᛣ೼㄀ 2 ゴᇚᄺдㅫ⊩ⱘᔦ㒇‫أ‬㕂ᅮНЎᮁ㿔䲚ড় Bˈ䗮䖛ᅗৃ‫ߚܙ‬ഄⓨ㒢᥼ᮁߎᄺд఼᠔
ᠻ㸠ⱘᔦ㒇᥼⧚㒧ᵰǄ՟བˈ‫׭‬䗝⍜䰸ㅫ⊩ⱘᔦ㒇‫أ‬㕂Ўˈ‫؛‬ᅮⳂᷛὖᗉ c ࣙ৿೼‫؛‬ᅮぎ䯈
H ЁǄ䖯ϔℹഄˈ៥Ӏ䖬䆕ᯢᄺдㅫ⊩ⱘ䕧ߎᰃ⬅݊䕧ܹҹঞ䖭ϔ䱤৿ⱘᔦ㒇‫أ‬㕂‫؛‬ᅮ᠔ⓨ
㒢ᕫߎⱘǄϞ䴶ⱘ䋱৊ᮃ㾷䞞ᇍѢᦣ䗄ᄺдㅫ⊩Ёⱘ䱤৿‫؛‬ᅮᦤկњ঺ϔ⾡ᮍ⊩Ǆ䖭䞠ˈϡ
ᰃ⫼ϔㄝᬜⱘⓨ㒢㋏㒳এᇍᔦ㒇᥼⧚ᓎ῵ˈ㗠ᰃ⫼෎Ѣ䋱৊ᮃ⧚䆎ⱘϔϾㄝᬜⱘὖ⥛᥼⧚
˄probabilistic reasoning˅㋏㒳Ǆ䖭䞠ᑨ⫼Ѣᄺд఼ⱘ䱤৿‫؛‬ᅮᔶᓣЎ˖ĀH Ϟⱘ‫ܜ‬偠ὖ⥛⬅
P(h)ߚᏗ㒭ߎˈ㗠᭄᥂ᢦ㒱៪᥹ফ‫؛‬䆒ⱘᔎᑺ⬅ P(D|h)㒭ߎǄ āᴀкⱘ P(h)੠ P(D|h)ᅮНࠏ
⬏њ‫׭‬䗝⍜䰸੠ Find-S ㋏㒳Ёⱘ䱤৿‫؛‬ᅮǄ೼Ꮖⶹ䖭ѯ‫؛‬ᅮⱘὖ⥛ߚᏗৢˈϔϾ෎Ѣ䋱৊
ᮃ⧚䆎ⱘὖ⥛᥼⧚㋏㒳ᇚѻ⫳ㄝᬜѢ䖭ѯㅫ⊩ⱘ䕧ܹ-䕧ߎ㸠ЎǄ
ᴀ㡖Ёⱘ䅼䆎ᰃ䋱৊ᮃ᥼⧚ⱘϔ⾡⡍⅞ᔶᓣˈ಴Ў៥Ӏা㗗㰥њ P(D|h)পؐা㛑Ў 0
៪ 1 ⱘᚙ‫ˈމ‬ᅗড᯴њ‫؛‬䆒乘⌟ⱘ⹂ᅮᗻҹঞ᮴ాໄ᭄᥂ⱘࠡᦤǄབৢϔ㡖᠔⼎ˈ䖬ৃҹ䗮
䖛‫ܕ‬䆌 P(D|h)পؐЎ 0 ੠ 1 П໪ⱘؐˈҹঞ೼ P(D|h)Ёࣙ৿䰘ࡴⱘᦣ䗄ҹ㸼⼎ాໄ᭄᥂ⱘߚ
Ꮧᚙ‫ˈމ‬ᴹ῵ᢳҢ᳝ాໄ䆁㒗᭄᥂Ёᄺдⱘ㸠ЎǄ
6.4 ᶷཝղ❬ૂᴶቅ䈥ᐤᒩᯯ‫ٽ‬䇴
བϞ㡖᠔⼎ˈ䋱৊ᮃߚᵤৃ⫼ᴹ㸼ᯢϔϾ⡍ᅮᄺдㅫ⊩Ӯ䕧ߎ MAP ‫؛‬䆒ˈेՓ䆹ㅫ⊩

≵᳝ᰒᓣഄՓ⫼䋱৊ᮃ㾘߭ˈ៪ҹᶤ⾡ᔶᓣ䅵ㅫὖ⥛Ǆ
ᴀ㡖㗗㰥ᄺд䖲㓁ؐⳂᷛߑ᭄ⱘ䯂乬ˈ䖭ᰃ೼䆌໮ᄺдㅫ⊩Ё䛑Ӯ䘛ࠄⱘˈབ⼲㒣㔥㒰
ᄺдǃ㒓ᗻಲᔦǃҹঞ໮乍ᓣ᳆㒓ᢳড়Ǆ䗮䖛ㅔऩⱘ䋱৊ᮃߚᵤˈৃҹ㸼ᯢ೼⡍ᅮࠡᦤϟˈ
ӏϔᄺдㅫ⊩བᵰՓ䕧ߎⱘ‫؛‬䆒乘⌟੠䆁㒗᭄᥂П䯈ⱘ䇃Ꮒᑇᮍ᳔ᇣ࣪ˈᅗᇚ䕧ߎϔᵕ໻Ԑ
✊‫؛‬䆒Ǆ䖭ϔ㒧䆎ⱘᛣН೼ѢˈᇍѢ䆌໮⼲㒣㔥㒰੠᳆㒓ᢳড়ⱘᮍ⊩ˈབᵰᅗӀ䆩೒೼䆁㒗
᭄᥂ϞՓ䇃Ꮒᑇᮍ੠᳔ᇣ࣪ˈℸ㒧䆎ᦤկњϔ⾡䋱৊ᮃⱘ䆎䆕ᮍ⊩˄೼⡍ᅮࠡᦤϟ˅Ǆ
䆒ᛇ䯂乬ᅮНབϟˈᄺд఼LᎹ԰೼ᅲ՟ぎ䯈X੠‫؛‬䆒ぎ䯈HϞˈHЁⱘ‫؛‬䆒ЎXϞᅮН
ⱘᶤ⾡ᅲ᭄ؐߑ᭄˄ेˈHЁ↣ϾhЎϔߑ᭄˖h: Xėˈ݊Ёҷ㸼ᅲ᭄䲚˅ǄL䴶Јⱘ䯂乬
ᰃᄺдϔϾҢHЁᢑপߎⱘ᳾ⶹⳂᷛߑ᭄f˖XėǄ㒭ᅮmϾ䆁㒗ḋ՟ⱘ䲚ড়ˈ↣Ͼḋ՟ⱘ
Ⳃᷛؐ㹿ᶤ䱣ᴎాໄᑆᡄˈℸ䱣ᴎాໄ᳡ҢℷᗕߚᏗǄ᳈㊒⹂ഄ䆆ˈ↣Ͼ䆁㒗ḋ՟ᰃᑣ‫ي‬Ǉxi,
diǈˈ݊Ёdi=f(xi)+eiǄ䖭䞠f(xi)ᰃⳂᷛߑ᭄ⱘ᮴ాໄؐˈeiᰃϔҷ㸼ాໄⱘ䱣ᴎব䞣Ǆ‫؛‬ᅮeiⱘ
ؐᰃ⣀ゟᢑপⱘˈᑊϨᅗӀⱘߚᏗ᳡Ң䳊ഛؐⱘℷᗕߚᏗǄᄺд఼ⱘӏࡵᰃ೼᠔᳝‫؛‬䆒᳝Ⳍ
ㄝⱘ‫ܜ‬偠ὖ⥛ࠡᦤϟˈ䕧ߎᵕ໻ৃ㛑‫؛‬䆒˄ेMAP‫؛‬䆒˅ˈ Ǆ
㱑✊៥Ӏⱘߚᵤᑨ⫼Ѣӏᛣᅲ᭄ؐߑ᭄ᄺдˈ✊㗠ৃҹ⫼ϔϾㅔऩⱘ՟ᄤᴹᦣ䗄䖭ϔ䯂
乬ˈेᄺд㒓ᗻߑ᭄Ǆ೒ 6-2 ᠔⼎Ўϔ㒓ᗻⳂᷛߑ᭄f˄ҹᅲ㒓㸼⼎˅ˈҹঞ䆹Ⳃᷛߑ᭄ⱘ᳝
ాໄ䆁㒗ḋ՟䲚Ǆ㰮㒓ᇍᑨ᳔᳝ᇣᑇᮍ䆁㒗䇃Ꮒⱘ‫؛‬䆒hMLˈгेᵕ໻Ԑ✊‫؛‬䆒Ǆ⊼ᛣˈ݊
Ёᵕ໻Ԑ✊‫؛‬䆒ϡϔᅮㄝѢℷ⹂‫؛‬䆒fˈ಴ЎᅗᰃҢ᳝䰤ⱘᏺాໄ᭄᥂Ё᥼䆎ᕫߎⱘǄ
原书页码： 164
೒ 6-2 ᄺдϔᅲؐߑ᭄Ǆ
Ⳃᷛߑ᭄fᇍᑨᅲ㒓Ǆ‫؛‬ᅮ䆁㒗ḋ՟Ǉxi, diǈЎⳳᅲⳂᷛؐࡴϞϔ䳊ഛؐⱘℷᗕߚᏗాໄeiǄ㰮
㒓ҷ㸼Փ䇃ᏂᑇᮍП੠᳔ᇣⱘ㒓ᗻߑ᭄Ǆ಴ℸˈᅗህᰃ䖭 5 Ͼ䆁㒗ḋ՟ϟⱘᵕ໻Ԑ✊‫؛‬䆒hMLǄ
䖭䞠ⱘ䇃Ꮒᑇᮍ੠᳔ᇣⱘ‫؛‬䆒ेЎᵕ໻Ԑ✊‫؛‬䆒ˈЎ䇈ᯢ䖭ϔ⚍ˈ佪‫ܜ‬ᖿ䗳ഄಲ乒ϔϟ
㒳䅵⧚䆎ЁⱘϸϾ෎ᴀὖᗉ˖ὖ⥛ᆚᑺ੠ℷᗕߚᏗǄ佪‫ˈܜ‬Ў䅼䆎‫ڣ‬e䖭ḋⱘ䖲㓁ব䞣Ϟⱘ
ὖ⥛ˈ៥Ӏᓩܹὖ⥛ᆚᑺ˄probability density˅Ǆㅔऩⱘ㾷䞞ᰃˈ៥Ӏ䳔㽕䱣ᴎব䞣᠔᳝ৃ㛑
ؐⱘὖ⥛੠Ў 1Ǆ⬅Ѣব䞣ᰃ䖲㓁ⱘˈ಴ℸϡ㛑Ў䱣ᴎব䞣ⱘ᮴䰤⾡ৃ㛑ⱘؐ䌟ќϔϾ᳝䰤
ὖ⥛Ǆ䖭䞠䳔㽕⫼ὖ⥛ᆚᑺᴹҷ᳓ˈҹՓe䖭ḋⱘ䖲㓁ব䞣೼᠔᳝ؐϞⱘὖ⥛ᆚᑺⱘ⿃ߚЎ
1Ǆϔ㠀ഄˈ⫼ᇣ‫ݭ‬ᄫ↡pᴹҷ㸼ὖ⥛ᆚᑺߑ᭄ˈҹऎߚ᳝䰤ὖ⥛P˄ᅗ᳝ᯊজ⿄Ўὖ⥛䋼䞣
˄probability mass˅˅Ǆὖ⥛ᆚᑺp(x0)ᰃᔧ¦䍟䖥Ѣ 0 ᯊˈxপؐ೼[x0, x0+¦)ऎ䯈‫ⱘݙ‬ὖ⥛Ϣ
1/¦Ь⿃ⱘᵕ䰤Ǆ
ὖ⥛ᆚᑺߑ᭄˖
1
p ( x 0 ) { lim P( x0 d x x 0 H )
H o0 H
݊⃵ˈ䖭䞠ᮁᅮ䱣ᴎాໄব䞣 e ⬅ℷᗕߚᏗ⫳៤ǄℷᗕߚᏗᰃϔᑇ⒥ⱘ䩳ᔶߚᏗˈᅗৃ
⬅݊ഛؐ੠ᷛ‫ޚ‬Ꮒ³ᅠܼࠏ⬏Ǆ㾕㸼 5-4 Ёⱘ㊒⹂ᅮНǄ
᳝њҹϞⱘϸϾὖᗉˈ‫ݡ‬ᴹ䅼䆎೼៥Ӏⱘ䯂乬䞠ЎҔМ᳔ᇣ䇃Ꮒᑇᮍ‫؛‬䆒ᅲ䰙Ϟህᰃᵕ
໻Ԑ✊‫؛‬䆒Ǆ䆕ᯢⱘ䖛⿟‫ܜ‬Փ⫼ࠡ䴶ⱘᓣ 6-3 ⱘᅮНᴹ᥼ᇐᵕ໻ৃ㛑‫؛‬䆒ˈԚՓ⫼ᇣ‫ ⱘݭ‬p
ҷ㸼ὖ⥛ᆚᑺ˖
hML arg max P( D | h)

hH
བࠡ᠔䗄ˈ‫؛‬ᅮ᳝ϔ೎ᅮⱘ䆁㒗ᅲ՟䲚ড়Ǉx1…xmǈˈ಴ℸা㗗㰥ⳌᑨⱘⳂᷛؐᑣ߫D=
Ǉd1…dmǈǄ䖭䞠di=f(xi)+eiǄ‫؛‬ᅮ䆁㒗ḋ՟ᰃⳌѦ⣀ゟⱘˈ㒭ᅮhᯊˈৃᇚP(D|h)‫ݭ‬៤৘p(di|h)
ⱘ⿃˖
m
hML arg max p (d i | h)
hH i 1
བᵰ䇃Ꮒei᳡Ң䳊ഛؐ੠᳾ⶹᮍᏂ³2ⱘℷᗕߚᏗˈ↣Ͼdiгᖙ乏᳡ҢℷᗕߚᏗˈ݊ᮍᏂ
Ў³2ˈ㗠ϨҹⳳᅲⱘⳂᷛؐf(xi)ЎЁᖗ˄㗠ϡᰃ 0˅Ǆ಴ℸˈp(di)ⱘৃ㹿‫ݭ‬ЎᮍᏂ³2ˈഛؐ
=f(xi)ⱘℷᗕߚᏗǄ⦄Փ⫼㸼 5-4 ЁⱘℷᗕߚᏗ݀ᓣᑊᇚⳌᑨⱘ੠³2ҷܹˈ‫ᦣߎݭ‬䗄p(di|h)
ⱘℷᗕߚᏗǄ⬅Ѣὖ⥛diⱘ㸼䖒ᓣᰃ೼hЎⳂᷛߑ᭄fⱘℷ⹂ᦣ䗄ᴵӊϟⱘˈ᠔ҹ䖬㽕᳓ᤶ
=f(xi)=h(xi)ǄᕫࠄǄ
m 1
1 ( di P )2
hML arg max e 2V 2
hH i 1 2SV 2
m 1
1 ( d i h ( xi )) 2
arg max e 2V 2
hH i 1 2SV 2
⦄೼Փ⫼ϔϾᵕ໻Ԑ✊䅵ㅫЁᐌ⫼ⱘ䕀ᤶ˖ϡᰃ⫼Ϟ䴶䖭Ͼ໡ᴖⱘ㸼䖒ᓣপ᳔໻ؐˈ㗠
ᰃՓ⫼݊ᇍ᭄প᳔໻ˈ䖭ḋ䕗ᆍᯧǄॳ಴ᰃ lnp ᰃ p ⱘऩ䇗ߑ᭄Ǆ಴ℸՓ lnp ᳔໻гህՓ p
᳔໻˖
m
1 1
hML arg max ¦ ln (d i h( xi )) 2
hH i 1 2SV 2 2V 2
ℸ㸼䖒ᓣЁ㄀ϔ乍Ўϔ⣀ゟѢ h ⱘᐌ᭄ˈৃ㹿ᗑ⬹ˈ಴ℸᕫࠄ˖
m
1
hML arg max ¦ (d i h( xi )) 2
hH i 1 2V 2
ՓϔϾ䋳ⱘ䞣᳔໻ㄝᬜѢՓⳌᑨⱘℷⱘ䞣᳔ᇣ˖
m
1
hML arg min ¦ (d i h( xi )) 2
hH i 1 2V 2
᳔ৢˈৃҹ‫ݡ‬ϔ⃵ᗑ⬹ᥝϢ h ᮴݇ⱘᐌ᭄˖
m
hML arg min ¦ (d i h( xi )) 2 (6.6)
hH i 1
䖭ḋˈᓣ 6.6 䇈ᯢњᵕ໻Ԑ✊‫؛‬䆒hMLЎˈՓ䆁㒗ؐdi੠‫؛‬䆒乘⌟ؐh(xi)П䯈ⱘ䇃Ꮒⱘᑇ

ᮍ੠᳔ᇣⱘ䙷ϔϾǄ䆹㒧䆎ࠡᦤᰃ㾖ᆳⱘ䆁㒗ؐdi⬅ⳳᅲⳂᷛؐࡴϞ䱣ᴎాໄѻ⫳ˈ݊Ё䱣
ᴎాໄᰃҢϔ䳊ഛؐⱘℷᗕߚᏗЁ⣀ゟᢑপⱘǄҢϞ䴶ⱘ᥼ᇐЁৃᯢ⹂ⳟߎˈ䇃Ꮒᑇᮍ乍
(di-h(xi))2ᰃҢℷᗕߚᏗᅮНЁⱘᣛ᭄乍ЁᕫᴹǄབᵰ‫؛‬ᅮాໄߚᏗ᳝঺໪ⱘᔶᓣˈৃ䖯㸠㉏
Ԑⱘ᥼ᇐᕫࠄϡৠⱘ㒧ᵰǄ
⊼ᛣϞ䴶ⱘ᥼ᇐࣙ৿њ䗝ᢽ‫؛‬䆒ՓԐ✊ⱘᇍ᭄ؐ˄lnp(D|h)˅Ў᳔໻ˈҹ⹂ᅮ᳔ৃ㛑ⱘ
‫؛‬䆒Ǆབࠡ᠔䗄ˈ䖭ᇐ㟈њϢՓ p(D|h)䖭ϾԐ✊ᗻ᳔໻࣪Ⳍৠⱘ㒧ᵰǄ䖭ϔ⫼ᇍ᭄Ԑ✊ᗻᴹ
䅵ㅫⱘᮍ⊩೼䆌໮䋱৊ᮃߚᵤЁ䛑⫼ࠄњˈ಴Ўᅗ↨Ⳉ᥹䅵ㅫԐ✊ᗻ䳔㽕ⱘ᭄ᄺ䖤ㅫ䞣ᇣᕜ
໮Ǆᔧ✊ˈབࠡ᠔䗄ˈᵕ໻Ԑ✊‫؛‬䆒г䆌ϡᰃ MAP ‫؛‬䆒ˈԚབᵰ᠔᳝‫؛‬䆒᳝Ⳍㄝⱘ‫ܜ‬偠ὖ
⥛ˈϸ㗙ⳌৠǄ
ЎҔМ⫼ℷᗕߚᏗᴹᦣ䗄ాໄᰃড়⧚ⱘ˛ϔϾᖙ乏ᡓ䅸ⱘॳ಴ᰃЎњ᭄ᄺ䅵ㅫⱘㅔ⋕
ᗻǄ঺ϔॳ಴ᰃˈ䖭ϔᑇ⒥ⱘ䩳ᔶߚᏗᇍ䆌໮⠽⧚㋏㒳ⱘాໄ䛑ᰃ㡃དⱘ䖥ԐǄᅲ䰙Ϟˈ㄀
5 ゴ䅼䆎ⱘЁᖗᵕ䰤ᅮᕟᰒ⼎ˈ䎇໳໮ⱘ⣀ゟৠߚᏗ䱣ᴎব䞣ⱘ੠᳡ҢϔℷᗕߚᏗˈ㗠ϡ䆎
⣀ゟব䞣ᴀ䑿ⱘߚᏗᰃҔМǄ䖭䇈ᯢ⬅䆌໮⣀ゟৠߚᏗⱘ಴㋴ⱘ੠᠔⫳៤ⱘాໄᇚ៤Ўℷᗕ
ߚᏗǄᔧ✊ˈ೼⦄ᅲЁϡৠⱘߚ䞣ᇍాໄⱘ䋵⤂г䆌ϡᰃৠߚᏗⱘˈ䖭ḋ䆹ᅮ⧚ᇚϡ㛑䆕ᯢ
៥Ӏⱘ䗝ᢽǄ
Փ䇃Ꮒᑇᮍ᳔ᇣ࣪ⱘᮍ⊩㒣ᐌ㹿⫼Ѣ⼲㒣㔥㒰ǃ᳆㒓ᢳড়ঞ݊Ҫᅲߑ᭄䘐䖥ⱘ䆌໮ㅫ⊩
ЁǄ㄀ 4 ゴ䅼䆎њẃᑺϟ䰡ᮍ⊩ˈᅗ೼⼲㒣㔥㒰Ё᧰㋶᳔ᇣ䇃Ꮒᑇᮍⱘ‫؛‬䆒Ǆ
೼㒧ᴳ䖭䞠݇Ѣᵕ໻Ԑ✊‫؛‬䆒੠᳔ᇣᑇᮍ䇃Ꮒ‫؛‬䆒ⱘ݇㋏ⱘ䅼䆎Пࠡˈᖙ乏䅸䆚ࠄ䆹䯂
乬ḚᶊЁⱘᶤѯ䰤ࠊǄϞ䴶ⱘߚᵤা㗗㰥њ䆁㒗ḋ՟ⱘⳂᷛؐЁⱘాໄˈ㗠≵᳝㗗㰥ᅲ՟ሲ
ᗻЁⱘాໄǄ՟བˈབᵰᄺд䯂乬ᰃ෎ѢᶤҎⱘᑈ啘੠催ᑺˈ乘⌟Ҫⱘ䞡䞣ˈ䙷МϞ䴶ⱘߚ
ᵤ㽕∖ˈ䞡䞣ⱘ⌟䞣Ёৃҹ᳝ాໄˈ㗠ᑈ啘੠催ᑺⱘ⌟䞣ᖙ乏ᰃ㊒⹂ⱘǄབᵰᇚ䖭ѯㅔ࣪‫؛‬
ᅮএᥝˈߚᵤ䖛⿟ᇚकߚ໡ᴖǄ
6.5 ⭞ӄ人⎁Ᾰ⦽Ⲻᶷཝղ❬‫ٽ‬䇴
೼ࠡϔゴⱘ䯂乬ḚᶊЁˈ៥Ӏ⹂ᅮњᵕ໻Ԑ✊‫؛‬䆒ᰃՓ݊೼䆁㒗ḋ՟Ϟⱘ䇃Ꮒᑇᮍ੠᳔
ᇣⱘ‫؛‬䆒Ǆᴀ㡖ᇚ᥼ᇐϔϾ㉏Ԑⱘ‫ˈ߭ޚ‬ᅗ䩜ᇍ⼲㒣㔥㒰ᄺд䖭ḋⱘ䯂乬˖ेᄺд乘⌟ὖ⥛Ǆ
㗗㰥䯂乬ⱘḚᶊЎᄺдϔϾϡ⹂ᅮᗻ˄ὖ⥛ⱘ˅ߑ᭄ f˖Xė{0, 1}ˈᅗ᳝ϸϾ⾏ᬷⱘؐ
䕧ߎǄ՟བˈᅲ՟ぎ䯈 X ҷ㸼᳝ᶤѯ⮛⢊ⱘ⮙ҎˈⳂᷛߑ᭄ f(x)೼⮙Ҏ㛑ᄬ⌏ϟᴹᯊЎ 1ˈ
৺߭Ў 0Ǆ៪㗙䇈ˈX ҷ㸼‫׳‬䋻⬇䇋㗙ˈ㸼⼎Ў݊䖛এⱘֵ⫼ग़৆ˈབᵰҪ㛑៤ࡳഄᔦ䖬ϟ
ϔ⃵‫׳‬䌋ˈf(x)Ў 1ˈ৺߭Ў 0Ǆ䖭ϸ⾡ᚙ‫މ‬ϟ䛑㽕 f ᳝ϡ⹂ᅮᗻǄ՟བˈϔ㕸᳝Ⳍৠ⮛⢊ⱘ
⮙ҎЎ 92%ৃҹᄬ⌏ˈ8%ϡ㛑Ǆ䖭⾡ϡৃ乘⌟ᗻᴹ⑤Ѣ᳾㛑㾖ᆳࠄⱘ⮛⢊⡍ᕕˈ៪㗙ᰃ⮒
⮙䕀࣪Ё⹂ᅲᄬ೼ⱘϡ⹂ᅮᗻᴎࠊǄ᮴䆎䯂乬ⱘᴹ⑤ᰃҔМˈ㒧ᵰ䛑ᰃ㽕∖Ⳃᷛߑ᭄ⱘ䕧ߎ
Ў䕧ܹⱘὖ⥛ߑ᭄Ǆ
᳝њ䖭ḋⱘ䯂乬ᦣ䗄ˈ៥ӀᏠᳯᄺдᕫࠄⱘ⼲㒣㔥㒰˄៪݊Ҫᅲߑ᭄䘐䖥఼˅ⱘ䕧ߎᰃ
f(x)=1 ⱘὖ⥛Ǆᤶ㿔Пˈ䳔㽕ᡒࠄⳂᷛߑ᭄ f´=Xė[0, 1]ˈՓ f´=P(f(x)=1˅
Ǆ೼Ϟ䴶ⱘ⮙Ҏᄬ
⌏乘⌟ⱘ՟ᄤЁˈབᵰ x Ўᄬ⌏⥛ᰃ 92%ⱘ⮙ҎПϔˈ䙷М f´(x)=0.92ˈὖ⥛ߑ᭄ f(x)ᇚ᳝
92%ⱘᴎӮㄝѢ 1ˈ࠽ԭⱘ 8%ⱘᴎӮㄝѢ 0Ǆ
བԩՓ⫼ϔϾ⼲㒣㔥㒰ᴹᄺд f´˛ϔϾᕜᯢᰒⱘǃ㳂࡯ⱘᮍ⊩ᰃ佪‫ܜ‬ᬊ䲚ᇍ x ⱘ↣Ͼৃ

㛑ؐ㾖ᆳࠄⱘ 1 ੠ 0 ⱘ乥⥛ˈ✊ৢ䆁㒗⼲㒣㔥㒰ˈᇍ↣Ͼ x 䕧ߎⳂᷛ乥⥛Ǆϟ䴶ᇚ㾕ࠄˈ៥
ӀৃҹⳈ᥹Ң f ⱘ䆁㒗ḋ՟Ё䆁㒗⼲㒣㔥㒰ˈ㗠Ϩҡ㛑᥼ᇐߎ f´ⱘᵕ໻ৃ㛑ᗻ‫؛‬䆒Ǆ
೼ℸᚙ‫މ‬ϟЎᇏᡒᵕ໻Ԑ✊‫؛‬䆒ˈᑨՓ⫼ᗢḋⱘӬ࣪‫˛߭ޚ‬Ўಲㄨ䆹䯂乬佪‫ܜ‬䳔㽕㦋ᕫ
P(D|h)ⱘ㸼⼎Ǆ䖭䞠‫؛‬ᅮ䆁㒗᭄᥂DⱘᔶᓣЎD={Ǉx1ˈd1ǈĂǇxmˈdmǈ}ˈ݊ЁdiЎ㾖ᆳࠄ
ⱘf(xi)ⱘ 0 ៪ 1 ؐǄ
ಲᖚࠡϔ㡖Ёⱘᵕ໻Ԑ✊ঞ᳔ᇣ䇃Ꮒᑇᮍߚᵤˈ݊Ёㅔऩഄ‫؛‬ᅮᅲ՟Ǉx1ˈĂxmǈᰃ೎
ᅮⱘǄ䖭ḋህৃҹা⫼Ⳃᷛؐdiᴹࠏ⬏᭄᥂Ǆ㱑✊䖭䞠гৃҹ԰䖭ḋⱘㅔऩ‫؛‬ᅮˈԚ៥Ӏ䖭
䞠ৃҹ䙓‫ܡ‬䖭ϔ‫؛‬ᅮҹ䇈ᯢ䖭ᇍ᳔ৢⱘ䕧ߎ≵᳝ᕅડǄᇚxi੠di䛑ⳟ԰䱣ᴎব䞣ˈᑊ‫؛‬ᅮ↣
Ͼ䆁㒗ḋ՟䛑ᰃ⣀ゟᢑপⱘˈৃᡞP(D|h)‫ݭ‬԰˖
m
P ( D | h) P( x , d
i 1
i i | h) ˄6.7˅
䖯ϔℹഄˈৃҹ‫؛‬ᅮ䘛ࠄ↣ϔ⡍ᅮᅲ՟xiⱘὖ⥛⣀ゟѢ‫؛‬䆒hǄ՟བˈ䆁㒗᭄᥂䲚Ёࣙ
৿ϔ⡍ᅮ⮙Ҏxiⱘὖ⥛⣀ゟѢ݇Ѣᄬ⌏⥛ⱘ‫؛‬䆒˄㱑✊⮙Ҏⱘᄬ⌏Ϣ৺di⹂ᅲᔎ⚜ձ䌪Ѣh˅
Ǆ
ᔧx⣀ゟѢhᯊˈৃᇚϞᓣ䞡‫˄ݭ‬ᑨ⫼㸼 6-1 ⱘЬ⊩㾘߭˅Ў˖
m m
P ( D | h) P ( x i , d i | h)
i 1
P(d
i 1
i | h, xi ) P( xi ) ˄6.8˅
⦄೼䅵ㅫ೼‫؛‬䆒h៤ゟⱘᴵӊϟˈᇍϔϾᅲ՟xi㾖ᆳࠄdi=1 ⱘὖ⥛P(di|h, xi˅Ǆ⊼ᛣhᰃᇍ

ᑨⳂᷛߑ᭄ⱘ‫؛‬䆒ˈᅗℷད㛑䅵ㅫ䖭ϔὖ⥛Ǆ಴ℸˈP(di=1|h, xi˅=h(xi)ˈᑊϨϔ㠀ᚙ‫މ‬ϟ˖
P(di|h, xi)= ˄6.9˅

h(xi) བᵰ di=1
(1-h(xi)) བᵰ di=0
Ўᇚ݊ҷܹࠄᓣ 6.8 Ё∖ P(D|h)ˈ佪‫ܜ‬ᇚ݊㸼䖒Ўϔৃ᭄ᄺ᪡԰ᔶᓣǄ
P (d i | h, xi ) h( xi ) d i (1 h( xi ))1 di ˄6.10˅
1 d i
ᆍᯧ偠䆕ˈㄝᓣ 6.9 ੠ 6.10 ᰃㄝӋⱘǄ⊼ᛣᔧdi=1 ᯊˈᓣ 6-10 Ё㄀Ѡ乍 (1 h( xi ))
ㄝѢ 1Ǆ಴ℸP(di=1|h, xi˅=h(xi)ˈᅗϢᓣ 6.9 ㄝӋǄৠḋৃߚᵤdi=0 ᯊⱘᚙᔶǄ
ᇚᓣ 6.10 ҷᤶᓣ 6.8 ЁⱘP(di|h, xi)ᕫࠄ˖
m
P ( D | h) h( x )
i 1
i
di
(1 h( xi ))1 d i P ( xi ) ˄6.11˅
⦄‫ߎݭ‬ᵕ໻Ԑ✊‫؛‬䆒ⱘ㸼䖒ᓣ˖
m
hML arg max h( xi ) d i (1 h( xi ))1 d i P ( xi )
hH i 1
᳔ৢϔ乍Ў⣀ゟѢ h ⱘᐌ䞣ˈৃএᥝ˖
m
hML arg max h( xi ) di (1 h( xi ))1 di (6.12)
hH i 1
ᓣ 6.12 Ёে䖍ⱘ㸼䖒ᓣৃⳟ԰ᰃ㸼 5-3 ЁѠ乍ߚᏗ˄Binomial distribution˅ⱘϔ㠀࣪ᔶ

ᓣǄ䆹ᓣᦣ䗄ⱘὖ⥛ⳌᔧѢᡩᦋmϾϡৠ⹀Ꮥˈ䕧ߎᕫࠄǇd1…dmǈⱘὖ⥛ˈ݊Ё‫؛‬ᅮ↣Ͼ
⹀Ꮥxiѻ⫳ℷ䴶ⱘὖ⥛Ўh(xi)Ǆ⊼ᛣ㸼 5-3 ᦣ䗄ⱘѠ乍ߚᏗᕜㅔऩˈԚᅗ䰘ࡴњϔϾ‫؛‬ᅮˈ
े᠔᳝⹀Ꮥᦋߎℷ䴶ⱘὖ⥛ᰃⳌৠⱘ˄ेh(xi)=h(xj), i, j˅Ǆϸ⾡ᚙ‫މ‬ϟ៥Ӏ䛑‫؛‬ᅮ⹀Ꮥᡩᦋ
ⱘ䕧ߎᰃⳌѦ⣀ゟⱘˈ䖭ϔ‫؛‬䆒г䗖⫼Ѣᔧࠡⱘ䯂乬Ǆ
Ϣࠡ䴶ⱘᚙ‫Ⳍމ‬ৠˈབᵰ⫼Ԑ✊ᗻⱘᇍ᭄䅵ㅫӮ↨䕗ᆍᯧˈᕫࠄ˖
m
hML arg max ¦ d i ln h( xi ) (1 d i ) ln(1 h( xi )) ˄6.13˅
hH i 1
ᓣ 6.13 ᦣ䗄њ೼៥Ӏⱘ䯂乬Ёᖙ乏㹿᳔໻࣪ⱘ䞣Ǆℸ㒧ᵰৃϢࠡ䴶ⱘՓ䇃Ꮒᑇᮍ᳔ᇣ
࣪ⱘߚᵤⳌ㉏↨Ǆ⊼ᛣᓣ 6.13 Ϣ➉ߑ᭄ⱘϔ㠀ᓣ-ipilogpi˄೼㄀ 3 ゴ䅼䆎䖛˅ⱘⳌԐᗻǄ
ℷ಴ЎℸⳌԐᗻˈҹϞ䞣ⱘ䋳᳝ؐᯊ㹿⿄ЎѸঝ➉˄cross entropy˅Ǆ
6.5.1 ൞⾔㔅㖇㔒ѣụᓜᩒ㍘ԛ䗴ࡦղ❬ᙝᴶཝौ
Ϟ䴶䅼䆎њՓᓣ 6.13 Ёⱘ䞣᳔໻࣪ৃᕫࠄᵕ໻ৃ㛑‫؛‬䆒Ǆ⦄⫼ G(h, D)ҷ㸼䆹䞣Ǆᴀ㡖

Ў⼲㒣㔥㒰ᄺд᥼ᇐϔϾᴗؐ䆁㒗㾘߭ˈᅗՓ⫼ẃᑺϞछҹՓ G(h, D)᳔໻࣪Ǆ
བ㄀ 4 ゴЁⱘ䅼䆎ˈG(h, D)ⱘẃᑺৃ⬅G(h, D) ݇Ѣϡৠⱘ㔥㒰ᴗؐⱘ‫أ‬ᇐⱘ৥䞣㒭ߎˈ

ᅗᅮНњ⬅ℸᄺдࠄⱘ㔥㒰㸼⼎ⱘ‫؛‬䆒h˄㾕㄀ 4 ゴЁẃᑺϟ䰡᧰㋶ⱘϔ㠀䅼䆎ˈҹঞ䖭䞠
᠔Փ⫼ⱘᴃ䇁ⱘ㒚㡖˅Ǆ೼ℸᚙ‫މ‬ϟˈᇍᑨѢᴗؐwjk˄Ң䕧ܹkࠄऩ‫ܗ‬j˅ⱘG(h, D)ⱘ‫أ‬ᇐЎ˖
wG (h, D) m
wG (h, D) wh( xi )
ww jk
¦
i 1 wh( xi ) ww jk
m
w (d i ln h( xi ) (1 d i ) ln(1 h( xi ))) wh( xi )
¦
i 1 wh( xi ) ww jk
m
d i h( x i ) wh( xi )
¦ h( x )(1 h( x ))
i 1 ww jk
˄6.14˅
i i
ЎՓߚᵤ䖛⿟ㅔᯢˈ‫؛‬ᅮ⼲㒣㔥㒰ҢϔϾऩሖⱘ sigmoid ऩ‫ܗ‬ᓎゟǄ䖭⾡ᚙ‫މ‬ϟ᳝˖
wh( xi )
V c( xi ) xijk h( xi )(1 h( xi )) xijk
wZ jk
݊Ёxijkᰃᇍ㄀iϾḋ՟ⱘࠄऩ‫ ܗ‬jⱘ㄀kϾ䕧ߎˈ㗠³´(x)Ўsigmoid ᣸य़˄squashing˅ߑ

᭄ⱘᇐ᭄˄㾕㄀ 4 ゴ˅Ǆ᳔ৢˈᇚℸ㸼䖒ᓣҷܹࠄㄝᓣ 6.14ˈৃᕫࠄ㒘៤ẃᑺⱘᇐ᭄ⱘㅔऩ
㸼⼎˖
wG (h, D) m
ww jk
¦ (d
i 1
i h( xi )) xijk
಴Ў䳔㽕Փ⫼ P(D|h)᳔໻࣪㗠ϡᰃ᳔ᇣ࣪ˈ᠔ҹ៥Ӏᠻ㸠ẃᑺϞछ᧰㋶㗠ϡᰃẃᑺϟ䰡
᧰㋶Ǆ೼᧰㋶ⱘ↣ϔ⃵䗁ҷЁˈᴗؐ৥䞣ᣝẃᑺⱘᮍ৥䇗ᭈˈՓ⫼ᴗؐ᳈ᮄ㾘߭˖
w jk m w jk 'w jk
݊Ё˖
m
'w jk K ¦ (d i h( xi )) xijk ˄6.15˅
i 1
݊Ё¨ᰃϔᇣⱘℷᐌ䞣ˈ㸼⼎ẃᑺϞछ᧰㋶ⱘℹ䖯໻ᇣǄ
ᇚ䖭ϔᴗؐ᳈ᮄ㾘߭Ϣড৥Ӵ᪁ㅫ⊩˄݊⫼䗨ᰃՓ乘⌟੠㾖ᆳⱘ㔥㒰䕧ߎⱘ䇃Ꮒᑇᮍ੠
᳔ᇣ࣪˅Ё⫼ࠄⱘᴗؐ᳈ᮄ㾘߭Ⳍ↨䕗ˈৃҹᕫࠄ᳝䍷ⱘ㒧䆎Ǆ⫼Ѣ䕧ߎऩ‫ܗ‬ᴗؐⱘড৥Ӵ
᪁᳈ᮄ㾘߭˄㾕㄀ 4 ゴ˅
ˈՓ⫼䖭䞠ⱘ䆄োৃ䞡ᮄ㸼⼎Ў˖
w jk m w jk 'w jk
݊Ё
m
'w jk K ¦ h( xi )(1 h( xi ))(d i h( xi )) xijk
i 1
⊼ᛣᅗϢᓣ 6.15 Ёⱘ㾘߭ⳌԐˈাᰃ䰸њϔ乍h(xi)(1-h(xi))ˈᅗᰃsigmoidߑ᭄ⱘᇐ᭄Ǆ
ὖᣀϔϟˈ䖭ϸϾᴗؐ᳈ᮄ㾘߭೼ϸ⾡ϡৠⱘ䯂乬㚠᱃ϟᬊᬯࠄᵕ໻Ԑ✊‫؛‬䆒ǄՓ䇃Ꮒ
ᑇᮍ᳔ᇣ࣪ⱘ㾘߭ᇏᡒࠄᵕ໻Ԑ✊‫؛‬䆒෎Ѣⱘࠡᦤᰃˈ䆁㒗᭄᥂ৃҹ⬅Ⳃᷛߑ᭄ؐࡴϞℷᗕ
ߚᏗాໄᴹ῵ᢳǄՓѸঝ➉᳔ᇣ࣪ⱘ㾘߭ᇏᡒᵕ໻Ԑ✊‫؛‬䆒෎Ѣⱘࠡᦤᰃˈ㾖ᆳࠄⱘᏗᇨؐ
Ў䕧ܹᅲ՟ⱘὖ⥛ߑ᭄Ǆ
6.6 ᴶቅᨅ䘦䮵ᓜ߼ࡏ
ಲᖚϔϟ㄀ 3 ゴ݇ѢĀ༹ഢൊࠗߔāⱘ䅼䆎ˈ䖭ᰃϔϾᕜᐌ⫼ⱘᔦ㒇‫أ‬㕂ˈᅗৃ㹿ὖᣀ
Ў˖ĀЎ㾖ᆳࠄⱘ᭄᥂䗝ᢽ᳔ⷁⱘ㾷䞞āǄᴀゴ៥Ӏ㽕䅼䆎೼ᇍ༹ഢൊࠗߔⱘ䭓ᳳѝ䆎Ёⱘ޴
Ͼ䆎⚍Ǆ䖭䞠ᇍℸ㒭ߎϔϾ䋱৊ᮃⱘߚᵤˈᑊ䅼䆎ϔ㋻ᆚⳌ݇ⱘ‫⿄ˈ߭ޚ‬Ў᳔ᇣᦣ䗄䭓ᑺ‫ޚ‬
߭˄Minimum Description Length, MDL˅Ǆ
ᦤߎ᳔ᇣᦣ䗄䭓ᑺⱘⳂⱘᰃЎњḍ᥂ֵᙃ䆎Ёⱘ෎ᴀὖᗉᴹ㾷䞞hMAPⱘᅮНǄ‫⃵ݡ‬㗗㰥
Ꮖᕜ❳ᙝⱘhMAPᅮН˖
hMAP arg max P( D | h) P(h)

hH
ৃ㹿ㄝӋഄ㸼⼎ЎՓҹ 2 Ўᑩⱘᇍ᳔᭄໻࣪˖
hMAP arg max log 2 P( D | h) log 2 P(h)

hH
៪Փℸ᳔ⱘ䋳᳔ؐᇣ࣪˖
hMAP arg min log 2 P( D | h) log 2 P(h) ˄6.16˅

hH
ҸҎৗ᚞ⱘᰃˈᓣ 6.16 ৃ㹿㾷䞞Ў೼⡍ᅮⱘ‫؛‬䆒㓪ⷕ㸼⼎ᮍḜϞĀӬ‫ܜ‬䗝ᢽⷁⱘ‫؛‬䆒āǄ

Ў㾷䞞䖭ϔ⚍ˈ‫ܜ‬ᓩֵܹᙃ䆎ЁⱘϔϾ෎ᴀ㒧䆎Ǆ䆒ᛇ㽕Ў䱣ᴎӴ䗕ⱘ⍜ᙃ䆒䅵ϔϾ㓪ⷕˈ
݊Ё䘛ࠄ⍜ᙃiⱘὖ⥛ᰃpiǄ䖭䞠᳔ᛳ݈䍷ⱘᰃ᳔ㅔⷁⱘ㓪ⷕˈेЎњӴ䕧䱣ᴎֵᙃⱘ㓪ⷕ᠔
㛑ᕫࠄⱘ᳔ᇣᳳᳯӴ䗕ԡ᭄Ǆᰒ✊ˈЎՓᳳᳯⱘ㓪ⷕ䭓ᑺ᳔ᇣˈᖙ乏Ўৃ㛑ᗻ䕗໻ⱘ⍜ᙃ䌟
ќ䕗ⷁⱘ㓪ⷕǄShannon & Weaver˄1949˅䆕ᯢ᳔Ӭ㓪ⷕ˄Փᕫᳳᳯ⍜ᙃ䭓ᑺ᳔ⷁⱘ㓪ⷕ˅
ᇍ⍜ᙃiⱘ㓪ⷕ䭓ᑺЎ-log2piԡǄ៥ӀᇚՓ⫼ҷⷕCᴹ㓪ⷕ⍜ᙃi᠔䳔ⱘԡ᭄⿄Ў⍜ᙃiⱘ݇ѢC
ⱘᦣ䗄䭓ᑺ˄description length of message i with respect to C˅Ǆᷛ䆄ЎLC(i)Ǆ
ϟ䴶ᇚՓ⫼ҹϞ㓪ⷕ⧚䆎ⱘ㒧䆎ᴹ㾷䞞ㄝᓣ 6.16˖
x -log2P(h)ᰃ೼‫؛‬䆒ぎ䯈Hⱘ᳔Ӭ㓪ⷕϟhⱘᦣ䗄䭓ᑺǄᤶ㿔Пˈ䖭ᰃ‫؛‬䆒hՓ⫼݊
᳔Ӭ㸼⼎ᯊⱘ໻ᇣǄҹ䖭䞠ⱘ䆄োˈLCH(h)= -log2P(h)ˈ݊ЁCHЎ‫؛‬䆒ぎ䯈Hⱘ
᳔Ӭ㓪ⷕǄ
x -log2P(D|h)ᰃ೼㒭ᅮ‫؛‬䆒hᯊ䆁㒗᭄᥂Dⱘᦣ䗄䭓ᑺ˄೼ℸ᳔Ӭ㓪ⷕϟ˅Ǆҹ䖭䞠
ⱘ䆄ো㸼⼎ˈLCD|h(D|h)= -log2P(D|h)ˈ݊Ё˟D|hᰃ‫؛‬ᅮথ䗕㗙੠᥹䗕㗙䛑ⶹ䘧‫؛‬
䆒hᯊᦣ䗄᭄᥂Dⱘ᳔Ӭ㓪ⷕǄ
x ಴ℸৃᡞᓣ 6.16 䞡‫ˈݭ‬ҹᰒ⼎ߎhMAPᰃՓ‫؛‬䆒ᦣ䗄䭓ᑺ੠㒭ᅮ‫؛‬䆒ϟ᭄᥂ᦣ䗄
䭓ᑺП੠᳔ᇣ࣪ⱘ‫؛‬䆒hǄ
hMAP arg min LCH (h) LCD|h ( D | h)
h
݊ЁCH੠CD|hߚ߿ЎHⱘ᳔Ӭ㓪ⷕ੠㒭ᅮhᯊDⱘ᳔Ӭ㓪ⷕˈ᳔ᇣᦣ䗄䭓ᑺ
˄Minimum Description Length, MDL˅‫߭ޚ‬ᓎ䆂ˈᑨ䗝ᢽՓ䖭ϸϾᦣ䗄䭓ᑺⱘ੠
᳔ᇣ࣪ⱘ‫؛‬䆒Ǆᔧ✊Ўᑨ⫼ℸ‫ˈ߭ޚ‬೼ᅲ䏉Ёᖙ乏䗝ᢽ䗖ড়Ѣᄺдӏࡵⱘ⡍ᅮ
㓪ⷕ៪㸼⼎Ǆ‫؛‬ᅮՓ⫼ҷⷕC1੠C2ᴹ㸼⼎‫؛‬䆒੠㒭ᅮ‫؛‬䆒ϟⱘ᭄᥂ˈৃᇚMDL
‫߭ޚ‬䰜䗄Ў˖
᳔ᇣᦣ䗄䭓ᑺ‫˖߭ޚ‬䗝ᢽhMDLՓ
hMDL arg min LC1 (h) LC2 ( D | h) ˄6.17˅

hH
Ϟ䴶ⱘߚᵤᰒ⼎ˈབᵰ䗝ᢽC1Ў‫؛‬䆒ⱘ᳔Ӭ㓪ⷕCHˈᑊϨ䗝ᢽC2Ў᳔Ӭ㓪ⷕCD|hˈ䙷М
hMDL=hMAPǄ
Ⳉ㾝Ϟˈৃᇚ MDL ‫߭ޚ‬ᛇ䈵Ў䗝ᢽ᳔ⷁⱘᮍ⊩ᴹ䞡ᮄ㓪ⷕ䆁㒗᭄᥂ˈ݊Ёϡҙ䅵ㅫ‫؛‬

䆒ⱘ໻ᇣˈᑊϨ䅵ㅫ㒭ᅮ‫؛‬䆒ᯊ㓪ⷕ᭄᥂ⱘ䰘ࡴᓔ䫔Ǆ
В՟䇈ᯢˈ‫؛‬ᅮᇚMDL‫߭ޚ‬ᑨ⫼ࠄ‫އ‬ㄪᷥᄺдⱘ䯂乬ᔧЁǄᗢḋ䗝ᢽ‫؛‬䆒੠᭄᥂ⱘ㸼
⼎C1੠C2˛ᇍѢC1ˈৃҹᕜ㞾✊ഄ䗝ᢽᶤ⾡ᯢ⹂ⱘ‫އ‬ㄪᷥ㓪ⷕᮍ⊩ˈ݊Ёᦣ䗄䭓ᑺ䱣ⴔᷥЁ
㡖⚍੠䖍ⱘ๲䭓㗠๲ࡴǄབԩ䗝ᢽ㒭ᅮϔ‫އ‬ㄪᷥᯊ‫؛‬䆒ⱘ᭄᥂㓪ⷕC2ਸ਼˛ЎՓ䅼䆎ㅔऩ࣪ˈ
‫؛‬ᅮᅲ՟ᑣ߫Ǉx1…xmǈᰃ᥹ᬊ㗙੠থ䗕㗙䛑ⶹ䘧ⱘˈ䙷МৃҹাӴ䕧ߚ㉏㒧ᵰǇf(x1)…f(xm)ǈǄ
˄⊼ᛣӴ䗕ᅲ՟ⱘᓔ䫔⣀ゟѢℷ⹂ⱘ‫؛‬䆒ˈ಴ℸᅗϡӮᕅડࠄhMDLⱘ䗝ᢽǄ˅⦄೼ˈབᵰ䆁
㒗ߚ㉏Ǉf(x1)…f(xm)ǈϢ‫؛‬䆒ⱘ乘䅵Ⳍㄝˈ䙷Мህ≵ᖙ㽕Ӵ䕧᳝݇䖭ѯḋ՟ⱘӏԩֵᙃ˄᥹
ᬊ㗙ৃ೼݊ᬊࠄ‫؛‬䆒ৢ䅵ㅫ䖭ѯؐ˅Ǆ಴ℸ೼ℸᚙ‫މ‬ϟˈ㒭ᅮ‫؛‬䆒ⱘߚ㉏ᚙ‫މ‬ᯊⱘᦣ䗄䭓ᑺ
Ў 0Ǆབᵰᶤѯḋ՟㹿h䇃ߚ㉏ˈ䙷Мᇍ↣ϔ䇃ߚ㉏䳔㽕Ӵ䗕ϔϾ⍜ᙃҹ⹂ᅮાϾḋ՟㹿䇃
ߚ㉏њ˄ৃ⫼㟇໮log2mԡӴ䗕˅ˈᑊӴ䗕݊ℷ⹂ߚ㉏ؐ˄ৃ⫼㟇໮log2kԡˈ݊ЁkЎৃ㛑ߚ
㉏ؐⱘ᭄Ⳃ˅ Ǆ೼㓪ⷕC1੠C2ϟhMDL䖭ϔ‫؛‬䆒ህᰃՓ䖭ѯᦣ䗄䭓ᑺ੠᳔ᇣⱘ‫؛‬䆒Ǆ
಴ℸˈMDL ‫ᦤ߭ޚ‬կњϔ⾡ᮍ⊩ҹ೼‫؛‬䆒ⱘ໡ᴖᗻ੠‫؛‬䆒ѻ⫳䫭䇃ⱘ᭄䞣П䯈䖯㸠ᡬ
Ёˈᅗ᳝ৃ㛑䗝ᢽϔϾѻ⫳ᇥ䞣䫭䇃䕗ⷁⱘ‫؛‬䆒˗㗠ϡᰃ㛑ᅠ㕢ഄߚ㉏䆁㒗᭄᥂ⱘ䭓ⱘ‫؛‬䆒Ǆ
ⳟࠄ䖭ϔ⚍ˈህ᳝њϔ⾡໘⧚᭄᥂䖛ᑺᢳড়ⱘᮍ⊩Ǆ
Quinlan & Rivest˄1989˅ᦣ䗄њᑨ⫼ MDL ‫߭ޚ‬ҹ䗝ᢽ‫އ‬ㄪ᳔ᷥՇ໻ᇣⱘ޴Ͼᅲ偠Ǆ᡹

ਞᣛߎ෎Ѣ MDL ⱘᮍ⊩ѻ⫳ⱘ‫އ‬ㄪᷥⱘ㊒ᑺⳌᔧѢ㄀ 3 ゴЁ䅼䆎ⱘᷛ‫ⱘޚ‬ׂᷥ࠾ᮍ⊩Ǆ
Mehta et al.˄1995˅ᦣ䗄њ঺ϔϾ෎Ѣ MDL ⱘᮍ⊩䖯㸠‫އ‬ㄪׂᷥ࠾ˈᑊᅲ偠䆕ᯢ䆹ᮍ⊩ᕫ
ࠄⱘ㒧ᵰϢᷛ‫ޚ‬ׂᷥ࠾ᮍ⊩ⳌᔧǄ
Ң᳔ᇣᦣ䗄䭓ᑺⱘॳ߭ⱘߚᵤЁৃᕫࠄҔМ㒧䆎˛ᰃ৺䇈ᯢ᠔᳝ᚙ‫މ‬ϟⷁ‫؛‬䆒䛑᳔
ད˛㒧䆎ᰃ৺ᅮⱘǄᏆ㒣䆕ᯢⱘাᰃˈᔧ䗝ᅮ‫؛‬䆒㸼⼎ҹՓ‫؛‬䆒hⱘ໻ᇣЎ-log2P(h)ˈᑊϨ
䗝ᢽ՟໪ᚙ‫ⱘމ‬㸼⼎ҹՓ㒭ᅮhϟDⱘ㓪ⷕ䭓ᑺㄝѢ-log2P(D|h)ᯊˈMDL‫߭ޚ‬ѻ⫳MAP‫؛‬䆒Ǆ
✊㗠Ў䇈ᯢҹϞϸ㗙ৃҹབℸ㸼⼎ˈᖙ乏ⶹ䘧᠔᳝ⱘ‫ܜ‬偠ὖ⥛P(h)ˈҹঞP(D|h)Ǆ≵᳝⧚⬅
ⳌֵMDL‫؛‬䆒ᇍѢӏᛣ㓪ⷕC1੠C2䛑ᰃ᳔དⱘǄߎѢᅲ䰙ⱘ㗗㰥ˈ᳈ᆍᯧⱘࡲ⊩ᰃ⬅䆒䅵
㗙ᣛᅮϔϾ㸼⼎ˈҹᤩ㦋᳝݇‫؛‬䆒ὖ⥛ⱘⶹ䆚ˈ㗠ϡᰃᅠᭈഄᣛᅮ↣Ͼ‫؛‬䆒ⱘὖ⥛Ǆᄺᴃ⬠
ᇍMDLᑨ⫼ࠄᅲ䰙䯂乬ⱘѝ䆎ˈЏ㽕Ў䗝ᢽC1੠C2㓪ⷕᦤկᶤ⾡ᔶᓣⱘ䆎䆕Ǆ
6.7 䍓ਬᯥᴶՎ࠼㊱ಞ
䖘Ҟ៥ӀᏆ䅼䆎њ䯂乬Ā㒭ᅮ䆁㒗᭄᥂ˈ᳔ৃ㛑ⱘ‫؛‬䆒ᰃҔМ˛āᅲ䰙Ϟˈ䆹䯂乬䗮ᐌ
Ϣ঺ϔ᳈᳝ᛣНⱘ䯂乬㋻ᆚⳌ݇˖Ā㒭ᅮ䆁㒗᭄᥂ˈᇍᮄᅲ՟ⱘ᳔ৃ㛑ߚ㉏ᰃҔМ˛ā㱑✊
ৃⳟߎ㄀ѠϾ䯂乬ৃㅔऩഄ⬅ᑨ⫼ MAP ‫؛‬䆒ࠄᮄᅲ՟ᴹᕫࠄˈᅲ䰙Ϟ䖬ৃ㛑᳈དⱘㅫ⊩Ǆ
Ўњ᳈Ⳉ㾖ѯˈ㗗㰥ϔࣙ৿ϝϾ‫؛‬䆒h1ˈh2ˈh3ⱘ‫؛‬䆒ぎ䯈Ǆ‫؛‬ᅮᏆⶹ䆁㒗᭄᥂ᯊϝϾ
‫؛‬䆒ⱘৢ偠ὖ⥛ߚ߿Ў 0.4ˈ0.3ˈ0.3Ǆ಴ℸˈh1ЎMAP‫؛‬䆒Ǆ㢹ϔᮄᅲ՟x㹿h1ߚ㉏Ўℷˈ
Ԛ㹿h2੠h3ߚ㉏ЎডǄ䅵ㅫ᠔᳝‫؛‬䆒ˈxЎℷ՟ⱘὖ⥛Ў 0.4˄ेϢh1Ⳍ㘨㋏ⱘὖ⥛˅ˈ㗠Ўড
՟ⱘὖ⥛ᰃ 0.6Ǆ䖭ᯊ᳔ৃ㛑ⱘߚ㉏˄ড՟ ˅ϢMAP‫؛‬䆒⫳៤ⱘߚ㉏ϡৠǄ
ϔ㠀ⱘ䇈ˈᮄᅲ՟ⱘ᳔ৃ㛑ߚ㉏ৃ䗮䖛ড়ᑊ᠔᳝‫؛‬䆒ⱘ乘⌟ᕫࠄˈ݊ᴗ䞡ЎᅗӀⱘৢ偠
ὖ⥛Ǆབᵰᮄⱘḋ՟ⱘৃ㛑ⱘߚ㉏ৃপᶤ䲚ড়VЁⱘӏϔؐvjˈ䙷Мὖ⥛P(vj|D)Ўᮄᅲ՟ℷ
⹂ߚ㉏Ўvjⱘὖ⥛ˈ݊ؐЎ˖
P (v j | D ) ¦ P (v
hi H
j | hi ) P (hi | D)
ᮄᅲ՟ⱘ᳔Ӭߚ㉏ЎP(vj|D)Ў᳔໻ᯊⱘvjؐǄ
䋱৊ᮃ᳔Ӭߚ㉏఼
arg max ¦ P(v j | hi ) P(hi | D) ˄6.18˅

v j V hi H
⫼Ϟ䴶ⱘ՟ᄤ䇈ᯢˈᮄᅲ՟ⱘৃ㛑ߚ㉏䲚ড়Ў V={,}ˈ㗠
P(h1|D)=0.4, P(|h1)=0, P(|h1)=1
P(h2|D)=0.3, P(|h2)=1, P(|h2)=0
P(h3|D)=0.3, P(|h3)=1, P(|h3)=0
಴ℸ
¦ P ( | h ) P ( h
hi H
i i | D) 0.4
¦ P (4 | h ) P ( h
hi H
i i | D) 0.6
ᑊϨ
arg max ¦ P(v j | hi ) P(hi | D) 4

v j { ,4} h H
i
ᣝ✻ᓣ 6.18 ߚ㉏ᮄᅲ՟ⱘ㋏㒳㹿Ў䋱৊ᮃ᳔Ӭߚ㉏఼˄Bayes optimal classifier˅ˈ៪䋱

৊ᮃ᳔Ӭᄺд఼ǄՓ⫼Ⳍৠⱘ‫؛‬䆒ぎ䯈੠Ⳍৠⱘ‫ܜ‬偠ὖ⥛ˈ≵᳝݊Ҫᮍ⊩㛑↨݊ᑇഛᗻ㛑᳈
དǄ䆹ᮍ⊩೼㒭ᅮৃ⫼᭄᥂ǃ‫؛‬䆒ぎ䯈ঞ䖭ѯ‫؛‬䆒ⱘ‫ܜ‬偠ὖ⥛ϟՓᮄᅲ՟ⱘℷ⹂ߚ㉏ⱘৃ㛑
ᗻ䖒ࠄ᳔໻Ǆ
՟བˈ೼Ꮧᇨὖᗉᄺд䯂乬ЁˈՓ⫼ࠡ䴶ゴ㡖ⱘবൟぎ䯈ᮍ⊩ˈᇍϔᮄᅲ՟ⱘ䋱৊ᮃ᳔
Ӭߚ㉏ᰃ೼বൟぎ䯈ⱘ᠔᳝៤ਬЁ䖯㸠ࡴᴗ䗝В㦋ᕫⱘˈ↣Ͼ‫׭‬䗝‫؛‬䆒ⱘᴗ䞡Ў݊ৢ偠ὖ
⥛Ǆ
䋱৊ᮃ᳔Ӭߚ㉏఼ⱘϔϾᵕ᳝䍷ⱘሲᗻᰃˈᅗ᠔‫ৃ㉏ߚⱘخ‬ҹᇍᑨѢ H Ёϡᄬ೼ⱘ‫؛‬
䆒Ǆ䆒ᛇՓ⫼ᓣ 6.18 ᴹߚ㉏ X Ё↣Ͼᅲ՟ǄᣝℸᅮНⱘⱘᅲ՟ᷛ⊼ϡᖙᇍᑨѢ H Ёⱘӏϔ
ऩϾ‫؛‬䆒 h ⱘᇍᅲ՟ⱘᷛ⊼Ǆ⧚㾷䆹ੑ乬ⱘϔ⾡ᮍ⊩ᰃᇚ䋱৊ᮃߚ㉏఼ⳟ៤ᰃϡৠѢ‫؛‬䆒ぎ
䯈 H ⱘ঺ϔぎ䯈 H´ˈ೼݊Ϟᑨ⫼䋱৊ᮃ݀ᓣǄ⹂ߛഄ䆆ˈH´᳝ᬜഄࣙ৿њϔ㒘‫؛‬䆒ˈᅗ㛑
೼ H Ё໮Ͼ‫؛‬䆒ⱘ㒓ᗻ㒘ড়᠔԰ⱘ乘㿔Ё䖯㸠↨䕗Ǆ
6.8 Gibbs ㇍⌋
㱑✊䋱৊ᮃ᳔Ӭߚ㉏఼㛑Ң㒭ᅮ䆁㒗᭄᥂Ё㦋ᕫ᳔དⱘᗻ㛑ˈᑨ⫼ℸㅫ⊩ⱘᓔ䫔ৃ㛑ᕜ
໻Ǆॳ಴೼Ѣᅗ㽕䅵ㅫ H Ё↣Ͼ‫؛‬䆒ⱘৢ偠ὖ⥛ˈ✊ৢড়ᑊ↣Ͼ‫؛‬䆒ⱘ乘⌟ˈҹߚ㉏ᮄᅲ
՟Ǆ
ϔϾ᳓ҷⱘǃ䴲᳔Ӭⱘᮍ⊩ᰃ Gibbs˄㾕 Opper & Haussler 1991˅ˈᅮНབϟ˖
1ˊᣝ✻ H Ϟⱘৢ偠ὖ⥛ߚᏗˈҢ H Ё䱣ᴎ䗝ᢽ‫؛‬䆒 hǄ
2ˊՓ⫼ h ᴹ乘㿔ϟϔᅲ՟ x ⱘߚ㉏Ǆ
ᔧ᳝ϔᕙߚ㉏ᮄᅲ՟ᯊˈGibbs ㅫ⊩ㅔऩഄᣝ✻ᔧࠡⱘৢ偠ὖ⥛ߚᏗˈՓ⫼ϔ䱣ᴎᢑপ
ⱘ‫؛‬䆒ǄҸҎৗ᚞ⱘᰃˈৃ䆕ᯢ೼ϔᅮᴵӊϟ Gibbs ㅫ⊩ⱘ䇃ߚ㉏⥛ⱘᳳᳯ᳔ؐ໮Ў䋱৊ᮃ
᳔Ӭߚ㉏఼ⱘϸ‫˄ס‬Haussher et al. 1994˅Ǆ᳈㊒⹂ഄ䆆ˈᳳᳯؐᰃ೼䱣ᴎᢑপⱘⳂᷛὖᗉϞ
԰ߎˈᢑপ䖛⿟ᣝ✻ᄺд఼‫؛‬ᅮⱘ‫ܜ‬偠ὖ⥛Ǆ೼ℸᴵӊϟˈGibbs ㅫ⊩ⱘ䫭䇃⥛ᳳᳯ᳔ؐᏂ
Ў䋱৊ᮃߚ㉏఼ⱘϸ‫ס‬Ǆ
䆹㒧䆎ᇍࠡ䗄ⱘὖᗉᄺд䯂乬᳝ϔ᳝䍷ⱘਃ⼎Ǆेབᵰᄺд఼‫؛‬ᅮ H Ϟ᳝ഛࣔⱘ‫ܜ‬偠
ὖ⥛ˈ㗠ϨབᵰⳂᷛὖᗉᅲ䰙Ϟгᣝ䆹ߚᏗᢑপˈ䙷Мᔧࠡবൟぎ䯈Ё䱣ᴎᢑপⱘ‫؛‬䆒ᇍϟ
ϔᅲ՟ߚ㉏ⱘᳳᳯ䇃Ꮒ᳔໮Ў䋱৊ᮃߚ㉏఼ⱘϸ‫ס‬Ǆ䖭䞠জ᳝њϔϾ՟ᄤ䇈ᯢ䋱৊ᮃߚᵤৃ
ҹᇍϔ䴲䋱৊ᮃㅫ⊩ⱘᗻ㛑䖯㸠䆘ԄǄ
6.9 ᵪ㍖䍓ਬᯥ࠼㊱ಞ
䋱৊ᮃᄺдᮍ⊩Ёᅲ⫼ᗻᕜ催ⱘϔ⾡Ўᴈ㋴䋱৊ᮃᄺд఼ˈᐌ㹿⿄Ўᴈ㋴䋱৊ᮃߚ㉏఼
˄naive Bayes classifier˅
Ǆ೼ᶤѯ乚ඳ‫݊ݙ‬ᗻ㛑ৃϢ⼲㒣㔥㒰੠‫އ‬ㄪᷥᄺдⳌᔧǄᴀ㡖ҟ㒡
ᴈ㋴䋱৊ᮃߚ㉏఼ˈϟϔ㡖ᇚ݊ᑨ⫼Ѣᅲ䰙ⱘ䯂乬ˈे㞾✊䇁㿔᭛ᴀ᭛ḷⱘߚ㉏䯂乬Ǆ
ᴈ㋴䋱৊ᮃߚ㉏఼ᑨ⫼ⱘᄺдӏࡵЁˈ↣Ͼᅲ՟xৃ⬅ሲᗻؐⱘড়পᦣ䗄ˈ㗠Ⳃᷛߑ᭄
f(x)Ңᶤ᳝䰤䲚ড়VЁপؐǄᄺд఼㹿ᦤկϔ㋏߫݇ѢⳂᷛߑ᭄ⱘ䆁㒗ḋ՟ˈҹঞᮄᅲ՟˄ᦣ
䗄Ўሲᗻؐⱘ‫ܗ‬㒘˅<a1,a2…an>ˈ✊ৢ㽕∖乘⌟ᮄᅲ՟ⱘⳂᷛؐ˄៪ߚ㉏˅Ǆ
䋱৊ᮃᮍ⊩ⱘᮄᅲ՟ߚ㉏Ⳃᷛᰃ೼㒭ᅮᦣ䗄ᅲ՟ⱘሲᗻؐ<a1,a2…an>ϟˈᕫࠄ᳔ৃ㛑ⱘ
ⳂᷛؐVMAPǄ
v MAP arg max P(v j | a1 , a 2 ...a n )

v j V
ৃՓ⫼䋱৊ᮃ݀ᓣᇚℸ㸼䖒ᓣ䞡‫ݭ‬Ў
P(a1 , a 2 ...a n | v j ) P(v j )
v MAP arg max
v j V P(a1 , a 2 ...a n )
arg max P(a1 , a 2 ...a n | v j ) P(v j ) ˄6.19˅

v j V
⦄೼㽕‫ⱘخ‬ᰃ෎Ѣ䆁㒗᭄᥂Ԅ䅵ᓣ 6.19 ЁϸϾ᭄᥂乍ⱘؐǄԄ䅵↣ϾP(vj)ᕜᆍᯧˈা

㽕䅵ㅫ↣ϾⳂᷛؐvjߎ⦄೼䆁㒗᭄᥂Ёⱘ乥⥛ህৃҹǄ✊㗠ˈ䰸䴲᳝ϔ䴲ᐌ໻ⱘ䆁㒗᭄᥂ⱘ
䲚ড়ˈ৺߭⫼䖭ḋᮍ⊩Ԅ䅵ϡৠⱘ P(a1,a2…an|vj)乍ϡ໾ৃ㸠Ǆ䯂乬೼Ѣ䖭ѯ乍ⱘ᭄䞣ㄝѢ
ৃ㛑ᅲ՟ⱘ᭄䞣Ьҹৃ㛑Ⳃᷛؐⱘ᭄䞣Ǆ಴ℸЎ㦋ᕫড়⧚ⱘԄ䅵ˈᅲ՟ぎ䯈Ё↣Ͼᅲ՟ᖙ乏
ߎ⦄໮⃵Ǆ
ᴈ㋴䋱৊ᮃߚ㉏఼෎ѢϔϾㅔऩⱘ‫؛‬ᅮ˖೼㒭ᅮⳂᷛؐᯊሲᗻؐП䯈ⳌѦᴵӊ⣀ゟǄᤶ
㿔Пˈ䆹‫؛‬ᅮ䇈ᯢ㒭ᅮᅲ՟ⱘⳂᷛؐᚙ‫މ‬ϟˈ㾖ᆳࠄ㘨ড়ⱘa1, a2…anⱘὖ⥛ℷདᰃᇍ↣Ͼऩ
⣀ሲᗻⱘὖ⥛Ь⿃˖
P (a1 , a 2 ...a n | v j ) P (a
i i | vj)
ᇚ݊ҷܹ 6.19 ᓣЁˈৃᕫࠄᴈ㋴䋱৊ᮃߚ㉏఼᠔Փ⫼ⱘᮍ⊩˖
ᴈ㋴䋱৊ᮃߚ㉏఼˖
v NB arg max P(v j )i P(ai | v j ) ˄6.20˅

v j V
݊ЁvNB㸼⼎ᴈ㋴䋱৊ᮃߚ㉏఼䕧ߎⱘⳂᷛؐǄ⊼ᛣ೼ᴈ㋴䋱৊ᮃߚ㉏఼Ёˈ乏Ң䆁㒗
᭄᥂ЁԄ䅵ⱘϡৠP(ai|vj)乍ⱘ᭄䞣াᰃϡৠⱘሲᗻ᭄ؐ䞣ЬҹϡৠⳂ᭄ᷛؐ䞣üü䖭↨㽕
Ԅ䅵P(a1,a2…an|vj)乍᠔䳔ⱘ䞣ᇣᕫ໮Ǆ
ὖᣀഄ䆆ˈᴈ㋴䋱৊ᮃᄺдᮍ⊩䳔㽕Ԅ䅵ϡৠⱘP(vj)੠P(ai|vj)乍ˈ෎ѢᅗӀ೼䆁㒗᭄᥂
Ϟⱘ乥⥛Ǆ䖭ѯԄ䅵ᇍᑨњᕙᄺдⱘ‫؛‬䆒Ǆ✊ৢ䆹‫؛‬䆒Փ⫼ᓣ 6.20 Ёⱘ㾘߭ᴹߚ㉏ᮄᅲ՟Ǆ
া㽕᠔䳔ⱘᴵӊ⣀ゟᗻ㛑໳㹿⒵䎇ˈᴈ㋴䋱৊ᮃߚ㉏vNBㄝѢMAPߚ㉏Ǆ
ᴈ㋴䋱৊ᮃᄺдᮍ⊩੠݊ҪᏆҟ㒡ⱘᄺдᮍ⊩П䯈᳝ϔ᳝䍷ⱘᏂ߿˖≵᳝ᯢ⹂ⱘ᧰㋶‫؛‬
䆒ぎ䯈ⱘ䖛⿟˄䖭䞠ˈৃ㛑‫؛‬䆒ⱘぎ䯈Ўৃ㹿䌟ќϡৠⱘP(vj)੠P(ai|vj)乍ⱘৃ㛑ؐǄⳌডˈ
‫؛‬䆒ⱘᔶ៤ϡ䳔㽕᧰㋶ˈাᰃㅔऩഄ䅵ㅫ䆁㒗ḋ՟Ёϡৠ᭄᥂㒘ড়ⱘߎ⦄乥⥛˅Ǆ
6.9.1 ⽰ׁ
⦄ᇚᴈ㋴䋱৊ᮃߚ㉏఼ᑨ⫼ࠄࠡ䴶‫އ‬ㄪᷥЁ䅼䆎䖛ⱘὖᗉᄺд䯂乬˖ᣝ✻ᶤҎᰃ৺㽕ᠧ
㔥⧗ᴹߦߚ໽⇨Ǆ㄀ 3 ゴⱘ㸼 3-2 ᦤկњⳂᷛὖᗉ PlayTennis ⱘ 14 Ͼ䆁㒗ḋ՟ˈ݊Ё↣ϔ
໽⬅ሲᗻ Outlook, Temprature, Humidity ੠ Wind ᴹᦣ䗄Ǆ䖭䞠៥ӀՓ⫼ℸ㸼Ёⱘ᭄᥂㒧ড়
ᴈ㋴䋱৊ᮃߚ㉏఼ᴹߚ㉏ϟ䴶ⱘᮄᅲ՟˖
<Outlook=sunny, Temperature=cool, Humidity=high, Wind=strong>
៥Ӏⱘӏࡵᰃᇍℸᮄᅲ՟乘⌟ⳂᷛὖᗉPlayTennis ⱘⳂᷛؐ˄yes ៪no˅

Ǆᇚᓣ 6.20 ᑨ
⫼ࠄᔧࠡⱘӏࡵˈⳂᷛؐvNB ⬅ϟᓣ㒭ߎ˖
v NB arg max P(v j )i P(ai | v j )

v j { yes , no}
arg max P(v j ) P(Outlook sunny | v j ) P(Temperature cool | v j )

v j { yes , no}
˄6.21˅
P ( Humidity high | v j ) P(Wind strong | v j )
⊼ᛣ೼᳔ৢϔϾ㸼䖒ᓣЁaiᏆ㒣⫼ᮄᅲ՟ⱘ⡍ᅮሲᗻؐᅲ՟࣪њǄЎ䅵ㅫvNBˈ⦄೼䳔㽕
10 Ͼὖ⥛ˈᅗӀ䛑ৃҹ䆁㒗᭄᥂ЁԄ䅵ߎǄ佪‫ܜ‬ϡৠⳂᷛؐⱘὖ⥛ৃҹ෎Ѣ䖭 14 Ͼ䆁㒗ḋ
՟ⱘ乥⥛ᕜᆍᯧഄԄ䅵ߎ˖
P(PlayTennis=yes)=9/14=0.64
P(PlayTennis=no)=5/14=0.36
ⳌԐഄˈৃҹԄ䅵ߎᴵӊὖ⥛ˈ՟བᇍѢ Wind=Strong ᳝˖
P(Wind=strong|PlayTennis=yes)=3/9=0.33
P(Wind=strong|PlayTennis=no)=3/5=0.60
Փ⫼䖭ѯὖ⥛Ԅ䅵ҹঞⳌԐⱘᇍ࠽ԭሲᗻⱘԄ䅵ˈৃᣝ✻ᓣ 6.21 䅵ㅫvNBབϟ˄Ўㅔᯢ

䍋㾕ᗑ⬹њሲᗻৡ˅Ǆ
P(yes)P(sunny|yes)P(cool|yes)P(high|yes)P(strong|yes)=0.0053
P(no)P(sunny|no)P(cool|no)P(high|no)P(strong|no)=0.0206
䖭ḋˈ෎ѢҢ䆁㒗᭄᥂Ёᄺдࠄⱘὖ⥛Ԅ䅵ˈᴈ㋴䋱৊ᮃߚ㉏఼ᇚℸᅲ՟䌟ҹⳂᷛؐ
PlayTennis= no Ǆ᳈䖯ϔℹˈ䗮䖛ᇚϞ䗄ⱘ䞣ᔦϔ࣪ˈৃ䅵ㅫ㒭ᅮ㾖ᆳؐϟⳂᷛؐЎ no ⱘ
ᴵӊὖ⥛ǄᇍѢℸ՟ˈὖ⥛Ў 0.0206/(0.0206+0.0053)=0.795Ǆ
6.9.1.1 զ䇗Ᾰ⦽
㟇ℸˈ៥Ӏ䗮䖛೼ܼ䚼џӊ෎⸔Ϟ㾖ᆳᶤџӊߎ⦄ⱘ↨՟ᴹԄ䅵ὖ⥛Ǆ՟བˈ೼Ϟ՟Ёˈ
Ԅ䅵P(Wind=Strong|PlayTennis=no)Փ⫼ⱘᰃ↨ؐnc/nˈ݊Ёn=5ˈЎ᠔᳝PlayTennis= no ⱘ䆁
㒗ḋ՟᭄Ⳃˈ㗠nc=3 ᰃ೼݊ЁWind=Strong ⱘ᭄ⳂǄ
ᰒ✊໮᭄ᚙ‫މ‬ϟˈ㾖ᆳࠄⱘ↨՟ᰃᇍὖ⥛ⱘϔϾ㡃དԄ䅵ˈԚᔧncᕜᇣᯊԄ䅵䛑䕗ᏂǄ
䲒ᑺ೼Ѣˈ䆒ᛇP(Wind=Strong|PlayTennis= no)ⱘؐЎ 0.08ˈ㗠ḋᴀЁা᳝ 5 Ͼḋ՟ⱘ
PlayTennis=noǄ䙷МᇍѢnc᳔ৃ㛑ⱘؐা᳝ 0Ǆ䖭ѻ⫳њϸϾ䲒乬ˈ佪‫ˈܜ‬nc/nѻ⫳њϔϾ
᳝‫ⱘأ‬䖛ԢԄ䅵˄underestimate˅ὖ⥛Ǆ݊⃵ˈᔧℸὖ⥛Ԅ䅵Ў 0 ᯊˈབᵰᇚᴹⱘᶹ䆶ࣙ৿
Wind=Strongˈℸὖ⥛乍Ӯ೼䋱৊ᮃߚ㉏఼ऴ᳝㒳⊏ഄԡǄॳ಴೼Ѣˈ⬅ᓣ 6.20 䅵ㅫⱘ䞣䳔
㽕ᇚ᠔᳝݊Ҫⱘὖ⥛乍Ьҹℸ 0 ؐǄ
Ў䙓‫ܡ‬䖭ѯ䲒乬ˈ䖭䞠䞛⫼ϔ⾡Ԅ䅵ὖ⥛ⱘ䋱৊ᮃᮍ⊩ˈՓ⫼བϟᅮНⱘ m-Ԅ䅵˖
ὖ⥛ⱘ m-Ԅ䅵
nc mp
˄6.22˅
nm
䖭䞠ˈnc੠nབࠡ䴶ᅮНˈpᰃᇚ㽕⹂ᅮⱘὖ⥛ⱘ‫ܜ‬偠Ԅ䅵ˈ㗠mᰃϔ⿄Ўㄝᬜḋᴀ໻ᇣ
ⱘᐌ䞣ˈᅗ⹂ᅮњᇍѢ㾖ᆳࠄⱘ᭄᥂བԩ㸵䞣pⱘ԰⫼Ǆ೼㔎ᇥ݊Ҫֵᙃᯊ䗝ᢽpⱘϔ⾡‫݌‬ൟ
ⱘᮍ⊩ᰃ‫؛‬ᅮഛࣔⱘ‫ܜ‬偠ὖ⥛ˈгህᰃˈབᵰᶤሲᗻ᳝kϾৃ㛑ؐˈ䙷М䆒㕂p=1/kǄ՟བˈ
ˈ⊼ᛣࠄሲᗻWind᳝ϸϾৃ㛑ؐˈ಴ℸഛࣔⱘ‫ܜ‬偠ὖ
ЎԄ䅵˄Wind=Strong|PlayTennis= no˅
⥛Ўp=0.5Ǆ⊼ᛣབᵰmЎ 0ˈm-Ԅ䅵ㄝᬜѢㅔऩⱘ↨՟nc/nǄབᵰn੠m䛑䴲 0ˈ䙷М㾖ᆳࠄ
ⱘ↨՟nc/n੠‫ܜ‬偠ὖ⥛pৃᣝ✻ᴗ䞡mড়ᑊǄm㹿⿄Ўㄝᬜḋᴀ໻ᇣⱘॳ಴ᰃˈᓣ 6.22 ৃ㹿㾷
䞞ЎᇚnϾᅲ䰙ⱘ㾖ᆳᠽ໻ˈࡴϞmϾᣝpߚᏗⱘ㰮ᢳḋᴀǄ
6.10 ⽰ׁφᆜҖ࠼㊱ᮽᵢ
Ўⓨ⼎䋱৊ᮃᄺдᮍ⊩೼ᅲ䏉Ϟⱘ䞡㽕ᗻˈ㗗㰥ϔϾᄺд䯂乬ˈ݊Ёⱘᅲ՟䛑Ў᭛ᴀ᭛
ḷǄ՟བˈ㽕ᄺдⳂᷛὖᗉ˖Ā៥ᛳ݈䍷ⱘ⬉ᄤᮄ䯏〓ā៪Ā䅼䆎ᴎ఼ᄺдⱘϛ㓈㔥义ā
Ǆ೼
䖭ϸ⾡ᚙ‫މ‬ϟˈབᵰ䅵ㅫᴎৃҹ㊒⹂ഄᄺдࠄⳂᷛὖᗉˈህৃҢ໻䞣೼㒓᭛ᴀ᭛ḷЁ㞾ࡼ䖛
Ⓒߎ᳔Ⳍ݇ⱘ᭛ḷᰒ⼎㒭䇏㗙Ǆ
䖭䞠ᦣ䗄њϔϾ෎Ѣᴈ㋴䋱৊ᮃߚ㉏఼ⱘ᭛ᴀߚ㉏ⱘ䗮⫼ㅫ⊩Ǆ᳝䍷ⱘᰃˈ䖭ḋⱘὖ⥛
ᮍ⊩ᰃⳂࠡ᠔ⶹ᭛ᴀ᭛ḷߚ㉏ㅫ⊩Ёⱘ᳔᳝ᬜⱘϔ㉏Ǆ䖭ḋⱘ㋏㒳՟ᄤ⬅ Lewis(1991)ˈ
Lang(1995)੠ Joachims(1996)ᦤߎǄ
ᇚ㽕ሩ⼎ⱘᴈ㋴䋱৊ᮃㅫ⊩䙉ᕾҹϟⱘ䯂乬㚠᱃Ǆ㗗㰥ᅲ՟ぎ䯈 X ࣙ৿њ᠔᳝ⱘ᭛ᴀ
᭛ḷ˄ेӏᛣ䭓ᑺⱘ᠔᳝ৃ㛑ⱘऩ䆡੠ᷛ⚍ヺোІ˅Ǆ㒭ᅮᶤ᳾ⶹⳂᷛߑ᭄ f(x) ⱘϔ㒘䆁㒗
ḋ՟ˈf(x)ⱘপؐᴹ㞾Ѣᶤ᳝䰤䲚ড় VǄℸӏࡵᰃҢ䆁㒗ḋ՟Ёᄺдˈҹ乘⌟ৢ㓁᭛ᴀ᭛ḷ
ⱘⳂᷛؐǄ԰Ў⼎՟ˈ䖭䞠㗗㰥ⱘⳂᷛߑ᭄ᰃˈᇚ᭛ḷߚ㉏ЎᇍᶤҎᰃ৺ᛳ݈䍷ˈՓ⫼Ⳃᷛ
ؐ like ੠ dislike ҷ㸼䖭ϸ㉏Ǆ
೼ᑨ⫼ᴈ㋴䋱৊ᮃߚ㉏఼ᯊࣙ৿ⱘϸϾЏ㽕䆒䅵䯂乬ᰃˈ佪‫ܜ‬㽕‫އ‬ᅮᗢḋᇚӏᛣ᭛ḷ㸼
⼎Ўሲᗻؐⱘᔶᓣˈ㄀Ѡ㽕‫އ‬ᅮབԩԄ䅵ᴈ㋴䋱৊ᮃߚ㉏఼᠔䳔ⱘὖ⥛Ǆ
䖭䞠㸼⼎ӏᛣ᭛ᴀ᭛ḷⱘ䗨ᕘߎ༛ഄㅔऩǄ㒭ᅮϔ᭛ᴀ᭛ḷˈ˄䖭䞠‫ܜ‬㗗㰥㣅᭛᭛ḷ˅ˈ
ৃᇍ↣Ͼऩ䆡ⱘԡ㕂ᅮНϔϾሲᗻˈ䆹ሲᗻⱘؐЎ೼ℸԡ㕂Ϟᡒࠄⱘ㣅᭛ऩ䆡Ǆ䆹᭛ᴀ᭛ḷ
བϟ՟᠔⼎˖
This is an example document for the naive Bayes classifier. This document contains only one
paragraph, or two sentences.
䖭ḋˈϞ՟Ёⱘ᭛ᴀ㹿㸼⼎Ў 19 Ͼሲᗻˈᇍᑨ 19 Ͼऩ䆡ԡ㕂Ǆ

㄀ϔϾሲᗻⱘؐЎĀThisāˈ
㄀ѠϾЎĀisāˈձ⃵㉏᥼Ǆ⊼ᛣ䕗䭓ⱘ᭛ḷг䳔㽕䕗໮ⱘሲᗻ᭄ⳂǄ៥Ӏᇚⳟࠄˈ䖭ϡӮ
ᏺᴹӏԩ咏⚺Ǆ
བᵰ᭛ᴀ᭛ḷ䖭ḋ㸼⼎ˈ⦄೼ህৃҹᑨ⫼ᴈ㋴䋱৊ᮃߚ㉏఼њǄЎњᯢ⹂䍋㾕ˈ‫؛‬ᅮ៥
Ӏ᳝ 700 Ͼ䆁㒗᭛ḷˈᑊϨᏆ⬅ҎᎹᇚ݊ߚ㉏Ў dislikeˈ㗠঺໪ 300 Ͼ᭛ḷ㹿ߚ㉏Ў likeǄ
⦄೼᳝њϔϾᮄ᭛ḷ㽕ߚ㉏ǄҡЎᯢ⹂䍋㾕ˈ䆹᭛ḷህᰃϞ䴶ⱘϸহ㣅᭛՟ᄤǄ೼ℸᚙ‫މ‬ϟˈ
ৃᑨ⫼ᓣ 6.20 䅵ㅫᴈ㋴䋱৊ᮃߚ㉏఼བ˖
19
v NB arg max P (v j ) P(ai | v j )
v j {like , dislike} i 1
arg max P(v j ) P(a1 " this"| v j ) P(a 2 " is"| v j )...P(a19 " sentence"| v j )
v j {like , dislike}
ὖᣀഄ䆆ˈᴈ㋴䋱৊ᮃߚ㉏vNBᰃՓ䆹᭛ḷЁⱘऩ䆡೼ℸ໘㹿㾖ᆳࠄⱘὖ⥛᳔໻ⱘϔϾ

19
ߚ㉏ˈᅗ䙉ᕾ䗮ᐌⱘᴈ㋴䋱৊ᮃ⣀ゟᗻ‫؛‬ᅮǄ⣀ゟᗻ‫؛‬ᅮ P ( a1 ,...a19 | v j ) 1
P (a i | v j )
䇈ᯢ೼ℸ䆒ᅮϟ೼ϔϾԡ㕂Ϟߎ⦄ᶤऩ䆡ⱘὖ⥛⣀ゟѢ঺໪ϔϾԡ㕂ⱘऩ䆡Ǆ䖭ϔ‫؛‬ᅮ೼᳝
ѯᯊ‫׭‬ᑊϡড᯴ⳳᅲᚙ‫މ‬Ǆ՟བˈ೼ᶤ໘㾖ᆳࠄऩ䆡learning ⱘὖ⥛Ӯ಴Ўᅗࠡϔԡ㕂ऩ䆡
ᰃmachine㗠๲໻Ǆ㱑✊ℸ⣀ゟᗻ‫؛‬ᅮᕜϡ㊒⹂ˈԚ䖭䞠߿᮴䗝ᢽˈᖙ乏԰ℸ‫؛‬ᅮüü≵᳝
䖭Ͼ‫؛‬ᅮˈ㽕䅵ㅫⱘὖ⥛乍ᇚᵕЎᑲ໻Ǆᑌ䖤ⱘᰃˈ೼ᅲ䏉Ёᴈ㋴䋱৊ᮃᄺд఼೼䆌໮᭛ᴀ
ߚ㉏䯂乬Ёᗻ㛑䴲ᐌདˈेՓℸ⣀ゟᗻ‫؛‬ᅮϡℷ⹂Ǆ Domingos ੠Pazzani(1996)ᇍ䖭ϔᑌ䖤
ⱘ⦄䈵԰њϔϾ᳝䍷ⱘߚᵤǄ
ЎՓ⫼Ϟᓣ䅵ㅫvNBˈ䳔㽕Ԅ䅵ὖ⥛乍P(vi)੠P(ai=wk|vi)Ǆ䖭䞠ᓩܹwkҷ㸼㣅᭛䆡‫݌‬Ёⱘ
㄀kϾऩ䆡Ǆࠡϔ乍ৃ෎Ѣ↣ϔ㉏೼䆁㒗᭄᥂Ёⱘ↨՟ᕜᆍᯧഄᕫࠄ˄ℸ՟ЁP(like)=0.3 Ϩ
P(dislike)= 0.7˅Ǆབҹᕔ䙷ḋˈԄ䅵㉏߿ⱘᴵӊὖ⥛˄བP(a1)=“This”| P(dislike)˅㽕ೄ䲒ⱘ
໮ˈ಴Ўᖙ乏ᇍ↣Ͼ᭛ᴀԡ㕂ǃ㣅᭛ऩ䆡੠Ⳃᷛؐⱘ㒘ড়䅵ㅫℸὖ⥛乍Ǆ䴲ᐌϡᑌˈ೼㣅᭛
䆡∛Ёࣙ৿㑺 5 ϛϾϡৠऩ䆡ˈ✊ৢᴀ՟Ё᳝ 2 Ͼৃ㛑ⱘⳂᷛؐ੠ 19 Ͼ᭛ᴀԡ㕂ˈ᠔ҹᖙ
乏Ң䆁㒗᭄᥂ЁԄ䅵 2×19×50000Ĭ200 ϛϾ䖭ḋⱘὖ⥛乍Ǆ
ᑌ䖤ⱘᰃˈৃҹ‫ݡ‬ᓩܹϔড়⧚ⱘ‫؛‬ᅮҹ‫ޣ‬ᇥ䳔㽕Ԅ䅵ⱘὖ⥛᭄䞣Ǆ⹂ߛഄ䆆ˈৃ‫؛‬ᅮ䘛
ࠄϔ⡍ᅮऩ䆡wkⱘὖ⥛⣀ゟѢऩ䆡᠔೼ԡ㕂Ǆᔶᓣ࣪ⱘ㸼䗄ᰃˈ೼㒭ᅮⳂᷛߚ㉏ⱘᚙ‫މ‬ϟˈ
‫؛‬ᅮ৘ሲᗻᰃ⣀ゟৠߚᏗⱘˈेᇍ᠔᳝ⱘi, j, k, mˈP(ai=wk|vj)=P(am=wk|vj)Ǆ಴ℸˈЎԄ䅵ᭈ
Ͼὖ⥛䲚ড়P(a1=wk|vj), P(a2=wk|vj)ˈৃ䗮䖛ϔϾԡ㕂᮴݇ⱘὖ⥛P(wk|vj)ˈ㗠ϡ㗗㰥ऩ䆡ⱘԡ
㕂Ǆ݊ᬜᵰᰃˈ⦄೼া䳔㽕 2×50000 Ͼϡৠⱘὖ⥛乍P(wk|vj)Ǆ㱑✊䖭ҡ✊ᰃϔϾ䕗໻ⱘ᭄
ؐˈԚैᰃৃㅵ⧚ⱘǄ⊼ᛣࠄབᵰ䆁㒗᭄᥂᳝䰤ˈ԰ℸ‫؛‬ᅮⱘϔϾЏ㽕Ӭ⚍೼ѢˈᅗՓৃ⫼
ѢԄ䅵↣Ͼ᠔䳔ὖ⥛ⱘḋ՟᭄๲ࡴњˈ಴ℸ๲ࡴњԄ䅵ⱘৃ䴴⿟ᑺǄ
Ўᅠ៤ᄺдㅫ⊩ⱘ䆒䅵ˈҡ䳔㽕䗝ᢽϔϾᮍ⊩Ԅ䅵ὖ⥛乍Ǆ䖭䞠䞛㒇њㄝᓣ 6.22 Ёⱘm

Ԅ䅵ˈे᳝㒳ϔⱘ‫ܜ‬偠ὖ⥛ᑊϨmㄝѢ䆡∛㸼ⱘ໻ᇣǄ಴ℸˈᇍP(wk|vj)ⱘԄ䅵Ў˖
nk 1
n Vocabulary
݊ЁnЎ᠔᳝ⳂᷛؐЎvjⱘ䆁㒗ḋ՟Ёऩ䆡ԡ㕂ⱘᘏ᭄ˈnkᰃ೼nϾऩ䆡ԡ㕂Ёᡒࠄwkⱘ
⃵᭄ˈ㗠|Vocabulary|Ў䆁㒗᭄᥂Ёⱘϡৠऩ䆡˄ҹঞ݊Ҫ䆄ো˅ⱘᘏ᭄Ǆ
ὖᣀഄ䇈ˈ᳔㒜ⱘㅫ⊩Փ⫼ⱘᴈ㋴䋱৊ᮃߚ㉏఼‫؛‬ᅮऩ䆡ߎ⦄ⱘὖ⥛Ϣᅗ೼᭛ᴀЁⱘԡ
㕂᮴݇Ǆ᳔㒜ⱘㅫ⊩ᰒ⼎೼㸼 6-2 ЁǄ⊼ᛣ䆹ㅫ⊩䴲ᐌㅔऩǄ೼ᄺд䖛⿟Ёˈ䖛⿟
Learn-naive-Bayes-textߚᵤ᠔᳝䆁㒗᭛ḷˈҢЁᢑপߎ᠔᳝ߎ⦄ⱘऩ䆡ⱘ䆄ো˗✊ৢ೼ϡৠ
Ⳃᷛ㉏Ё䅵ㅫ݊乥⥛ҹ㦋ᕫᖙ㽕ⱘὖ⥛Ԅ䅵Ǆҹৢˈ㢹㒭ᅮϔϾᕙߚ㉏ᮄᅲ՟ˈ䖛⿟
Classify-naive-Bayes-text Փ⫼ℸὖ⥛Ԅ䅵ᴹᣝ✻ᓣ 6.20 䅵ㅫVNBǄ⊼ᛣ೼ᮄ᭛ḷЁߎ⦄Ԛϡ
೼䆁㒗䲚ⱘ᭛ḷЁⱘӏԩऩ䆡ᇚ㹿ㅔऩഄᗑ⬹Ǆ䆹ㅫ⊩ⱘҷⷕҹঞ䆁㒗᭄᥂䲚ˈৃ೼ϛ㓈㔥
ⱘhttp://www.cs.cmu.edu/tom/book.htmlЁᡒࠄǄ
Learn_naive_Bayes_text(Examples, V)
ExamplesЎϔ㒘᭛ᴀ᭛ḷҹঞᅗӀⱘⳂᷛؐǄVЎ᠔᳝ৃ㛑Ⳃᷛؐⱘ䲚ড়Ǆℸߑ᭄԰⫼ᰃᄺдὖ⥛乍P(wk|vj)ˈ
ᅗᦣ䗄њҢ㉏߿vjЁⱘϔϾ᭛ḷЁ䱣ᴎᢑপⱘϔϾऩ䆡Ў㣅᭛ऩ䆡wkⱘὖ⥛Ǆ䆹ߑ᭄гᄺд㉏߿ⱘ‫ܜ‬偠ὖ⥛
P(vj)Ǆ
1.ᬊ䲚 Examples Ё᠔᳝ⱘऩ䆡ǃᷛ⚍ヺোҹঞ݊Ҫ䆄ো
Vocabularyĕ೼ Examples Ёӏᛣ᭛ᴀ᭛ḷЁߎ⦄ⱘ᠔᳝ऩ䆡ঞ䆄োⱘ䲚ড়
2.䅵ㅫ᠔䳔㽕ⱘὖ⥛乍P(vj)੠ P(wk|vj)
ᇍVЁ↣ϾⳂᷛؐvj
docsjĕExamplesЁⳂᷛؐЎvjⱘ᭛ḷᄤ䲚
P(vj) ĕ docs j
Examples
TextjĕᇚdocsjЁ᠔᳝៤ਬ䖲᥹䍋ᴹᓎゟⱘऩϾ᭛ḷ
nĕ೼TextjЁϡৠऩ䆡ԡ㕂ⱘᘏ᭄
ᇍVocabularyЁ↣Ͼऩ䆡wk
nkĕऩ䆡wkߎ⦄೼TextjЁⱘ⃵᭄
P(wk|vj) ĕ nk 1
n Vocabulary
Classify_naive_Bayes_text(Doc)
ᇍ᭛ḷDoc䖨ಲ݊Ԅ䅵ⱘⳂᷛؐǄaiҷ㸼೼DocЁⱘ㄀iϾԡ㕂Ϟߎ⦄ⱘऩ䆡Ǆ
positionsĕ೼ Doc Ёࣙ৿ⱘ㛑೼ Vocabulary Ёᡒࠄⱘ䆄োⱘ᠔᳝ऩ䆡ԡ㕂
䖨ಲ v
NB arg max P (v j )
v j V
P (a
i positions
i | vj)
㸼 6-2 ⫼Ѣᄺд੠ߚ㉏᭛ᴀⱘᴈ㋴䋱৊ᮃㅫ⊩
䰸䗮ᐌⱘᴈ㋴䋱৊ᮃ‫؛‬ᅮ໪ˈㅫ⊩䖬‫؛‬ᅮऩ䆡ߎ⦄ⱘὖ⥛⣀ゟѢ݊೼᭛ᴀЁⱘԡ㕂Ǆ
6.10.1 ᇔ僂㔉᷒
㸼 6-2 ⱘᄺдㅫ⊩ᬜ⥛བԩ˛೼ Joachims˄1996˅ⱘϔϾᅲ偠Ёˈℸㅫ⊩˄᳝ᖂᇣⱘব

࣪˅㹿ᑨ⫼Ѣߚ㉏ᮄ䯏㒘ⱘ᭛ゴǄ݊Ё↣ϔ᭛ゴⱘߚ㉏ᰃ䆹᭛ゴ᠔ሲⱘᮄ䯏㒘ৡ⿄Ǆℸᅲ偠
㗗㰥њ 20 Ͼ⬉ᄤᮄ䯏㒘ˈ✊ৢҢ↣Ͼᮄ䯏㒘Ё᧰䲚 1000 ㆛᭛ゴˈᔶ៤ϔϾࣙ৿ 2 ϛϾ᭛ḷ
ⱘ᭄᥂䲚Ǆ✊ৢᑨ⫼ᴈ㋴䋱৊ᮃㅫ⊩ˈ݊Ё 2/3 ԰Ў䆁㒗ḋ՟ˈ㗠ᗻ㛑ⱘ㸵䞣೼࠽ԭ 1/3 Ё
䖯㸠Ǆ಴Ў᳝ 20 Ͼৃ㛑ⱘᮄ䯏㒘ˈ䙷М䱣ᴎ⣰⌟ⱘߚ㉏㊒⹂ᑺЎ 5%Ǆ⬅⿟ᑣ㦋ᕫⱘ㊒⹂
ᑺЎ 89%Ǆℸᅲ偠ЁՓ⫼ⱘㅫ⊩Ϣ㸼 6-2 Ёⱘㅫ⊩া᳝ϔ⚍ϡৠ˖া᳝᭛ḷЁߎ⦄ऩ䆡ⱘϔ
Ͼᄤ䲚㹿䗝Ўㅫ⊩Ёⱘ䆡∛㸼ˈ⹂ߛഄ䇈ˈ100 Ͼ᳔ᐌ㾕ⱘऩ䆡㹿⿏এ˄བĀtheā੠Āofā
䖭ḋⱘऩ䆡˅ ˈ㗠Ϩӏԩߎ⦄ᇥѢ 3 ⃵ⱘऩ䆡г㹿⿏এǄᕫࠄⱘ䆡∛㸼ࣙ৿໻㑺 38,500 Ͼऩ
䆡Ǆ
݊Ҫᑨ⫼㉏Ԑⱘ㒳䅵ᄺдㅫ⊩䖯㸠᭛ᴀߚ㉏ⱘᅲ偠г㦋ᕫњৠḋདⱘ㒧ᵰǄ՟བˈLang
˄1995˅ᦣ䗄њᴈ㋴䋱৊ᮃㅫ⊩ⱘ঺ϔব⾡ˈᡞᅗᑨ⫼ࠄᄺдⳂᷛὖᗉĀ៥ᛳ݈䍷ⱘᮄ䯏㒘
᭛ゴāǄҪᦣ䗄њ NewsWeeder ㋏㒳üüᰃϔϾ䅽⫼᠋䯙䇏ᮄ䯏㒘᭛ゴᑊЎ݊䆘ߚⱘ㋏㒳Ǆ
✊ৢ NewsWeeder Փ⫼䖭ѯ䆘ߚⱘ᭛ゴ԰Ў䆁㒗ḋ՟ˈᴹ乘⌟ৢ㓁ⱘ᭛ゴાѯᰃ⫼᠋ᛳ݈䍷
ⱘˈ‫ݡ‬ᇚ݊䗕㒭⫼᠋䯙䇏ǄLang (1995)᡹ਞњҪⱘᅲ偠ˈ݊Ё⫼ NewsWeeder Ёᄺࠄⱘ⫼᠋
݈䍷䜡㕂᭛ӊˈ↣໽৥⫼᠋᥼㤤ߚ᳔ؐ催ⱘᮄ䯏᭛ゴǄ䗮䖛৥⫼᠋ሩ⼎ࠡ 10%ⱘ㞾ࡼ䆘ߚ
᭛ゴˈᅗᓎゟⱘ᭛ゴᑣ߫Ё⫼᠋ᛳ݈䍷ⱘ↨⥛↨䗮ᐌᚙ‫މ‬ϟ催ߎ 3-4 ‫ס‬Ǆ՟བˈ㢹ϔϾ⫼᠋
ᇍ䗮ᐌⱘ᭛ゴ᳝ 16%ᛳ݈䍷ˈ݊ᇍѢ NewsWeeder ᥼㤤ⱘ᭛ゴ᳝ 59%ᛳ݈䍷Ǆ
݊Ҫ޴⾡䴲䋱৊ᮃⱘ㒳䅵᭛ᴀᄺдㅫ⊩гᕜᐌ㾕ˈ݊Ё䆌໮෎ѢֵᙃẔ㋶乚ඳ
˄Information Retrieval˅Ёⱘ᳔‫ܜ‬থᯢⱘⳌԐᗻᑺ䞣˄㾕 Rocchio 1971˗Salton 1991˅Ǆ঺໪
ⱘ᭛ᴀᄺдㅫ⊩㾕 Hearst & Hirsh˄1996˅Ǆ
comp.graphics misc.forsale soc.religion.christian sci.space
comp.os.ms-windows.misc rec.autos talk.politics.guns sci.crypt
comp.sys.ibm.pc.hardware rec.motocycles talk.politics.mideast sci.electronics
comp.sys.mac.hardware rec.sport.baseball talk.politics.misc sci.med
comp.windows.x rec.sport.hockey talk.religion.misc
alt.atheism
㸼 6-3 ೼᭛ᴀߚ㉏ᅲ偠ЁՓ⫼ⱘ 20 Ͼᮄ䯏㒘
೼ᇍ↣Ͼᮄ䯏㒘⫼ 667 ㆛᭛ゴ䆁㒗ৢˈᴈ㋴䋱৊ᮃߚ㉏఼೼乘⌟ৢ㓁᭛ゴሲѢાϔϾᮄ䯏㒘

ᯊ㦋ᕫњ 89%ⱘ㊒ᑺǄ䱣ᴎ⣰⌟া㛑ᕫࠄ 5%ⱘ㊒⹂ᑺǄ
6.11 䍓ਬᯥ㖤‫ؗ‬㖇
བࠡϸ㡖᠔䅼䆎ⱘˈᴈ㋴䋱৊ᮃߚ㉏఼‫؛‬ᅮњሲᗻa1Ăanⱘؐ೼㒭ᅮⳂᷛؐvϟᰃᴵӊ
⣀ゟⱘǄ䖭ϔ‫؛‬ᅮᰒ㨫ഄ‫ޣ‬ᇣњⳂᷛߑ᭄ᄺдⱘ䅵ㅫ໡ᴖᑺǄᔧℸᴵӊ៤ゟᯊˈᴈ㋴䋱৊ᮃ
ߚ㉏఼ৃᕫࠄ᳔Ӭ䋱৊ᮃߚ㉏Ǆ✊㗠೼䆌໮ᚙᔶϟˈ䖭ϔᴵӊ⣀ゟ‫؛‬ᅮᯢᰒ䖛ѢϹḐњǄ
䋱৊ᮃ㕂ֵ㔥ᦣ䗄ⱘᰃϔ㒘ব䞣᠔䙉Ңⱘὖ⥛ߚᏗˈᅗ䗮䖛ϔ㒘ᴵӊὖ⥛ᴹᣛᅮϔ㒘ᴵ
ӊ⣀ゟᗻ‫؛‬ᅮǄᴈ㋴䋱৊ᮃߚ㉏఼‫؛‬ᅮ᠔᳝ব䞣೼㒭ᅮⳂᷛব䞣ؐᯊЎᴵӊ⣀ゟⱘˈϢℸϡ
ৠˈ䋱৊ᮃ㕂ֵ㔥Ёৃ㸼䗄ᑨ⫼ࠄব䞣ⱘϔϾᄤ䲚Ϟⱘᴵӊ⣀ゟᗻ‫؛‬ᅮǄ಴ℸˈ䋱৊ᮃ㕂ֵ
㔥ᦤկњϔ⾡Ё䯈ⱘᮍ⊩ˈᅗ↨ᴈ㋴䋱৊ᮃߚ㉏఼Ёᴵӊ⣀ゟᗻⱘܼሔ‫؛‬ᅮⱘ䰤ࠊ᳈ᇥˈজ
↨೼᠔᳝ব䞣Ё䅵ㅫᴵӊձ䌪᳈ৃ㸠Ǆ䋱৊ᮃ㕂ֵ㔥ᰃⳂࠡⷨおЁϔϾ䴲ᐌ⌏䎗ⱘ⛺⚍ˈ㗠
Ϩ᳝໮⾡ᮍ⊩㹿ᦤߎҹᄺдᅗ੠⫼ᅗ䖯㸠᥼⧚Ǆᴀ㡖ҟ㒡䋱৊ᮃ㕂ֵ㔥ⱘ݇䬂ὖᗉ੠㸼⼎Ǆ
᳈ 䆺㒚 ⱘ 䅼䆎㾕 Pearl(1988) ˈ Rusell & Norvig(1995) ˈ Herkerman et al. ˄ 1995 ˅ ҹ ঞ
Jensen(1996)Ǆ
ϔ㠀ᴹ䇈ˈ䋱৊ᮃ㕂ֵ㔥ᦣ䗄њ೼ϔ㒘ব䞣Ϟⱘὖ⥛ߚᏗǄ㗗㰥ϔӏᛣⱘ䱣ᴎব䞣䲚ড়
Y1ĂYnˈ݊Ё↣ϾYiৃপⱘؐ䲚ড়ЎV(Yi)ǄᅮНব䞣䲚ড়Yⱘ㘨ড়ぎ䯈(joint space)ЎঝЬV(Y1)
hV(Y2)ĂV(Yn)Ǆᤶ㿔Пˈ೼㘨ড়ぎ䯈Ёⱘ↣ϔ乍ᇍᑨব䞣‫ܗ‬㒘ⱘϔϾৃ㛑ⱘ䌟ؐ<Y1…Yn>Ǆ
೼ℸ㘨ড়ぎ䯈Ϟⱘὖ⥛ߚᏗ⿄Ў㘨ড়ὖ⥛ߚᏗ(joint probability distribution)Ǆ㘨ড়ὖ⥛ߚᏗᣛ
ᅮњ‫ܗ‬㒘<Y1…Yn>ⱘ↣Ͼৃ㛑ⱘব䞣㑺ᴳⱘὖ⥛Ǆ䋱৊ᮃ㕂ֵ㔥߭ᇍϔ㒘ব䞣ᦣ䗄њ㘨ড়ὖ
⥛ߚᏗǄ
6.11.1 ᶗԬ⤢㄁ᙝ
佪‫ˈܜ‬Ў䅼䆎䋱৊ᮃ㕂ֵ㔥ˈ䳔㽕㊒⹂ᅮНᴵӊ⣀ゟᗻǄҸ XˈY ੠ Z Ў 3 Ͼ⾏ᬷؐ䱣

ᴎব䞣Ǆᔧ X ᳡Ңⱘὖ⥛ߚᏗ⣀ゟѢ㒭ᅮ Z ؐᯊ Y ⱘؐˈ⿄ X ೼㒭ᅮ Z ᯊᴵӊ⣀ゟѢ Yˈ
े˖
(xi , y j , z k ) P ( X xi | Y yj,Z zk ) P( X xi | Z zk )
݊Ё xi V ( X ) ˈ y j V (Y ) ˈ z k V ( Z ) Ǆ䗮ᐌᇚϞᓣㅔ‫ݭ‬ЎP(X|Y,Z)=P(X|Z)Ǆ䖭ϔ
݇Ѣᴵӊ⣀ゟᗻⱘᅮНৃ㹿ᠽሩࠄব䞣䲚ড়Ǆᔧϟ߫ᴵӊ៤ゟᯊˈ⿄ব䞣䲚ড়X1…Xl㒭ᅮব
䞣䲚ড়Z1…Znᯊᴵӊ⣀ゟѢব䞣䲚ড়Y1…Ym˖
P( X 1 ... X l | Y1 ...Ym , Z 1 ...Z n ) P( X 1 ... X l | Z 1 ...Z n )
⊼ᛣℸᅮНϢᴈ㋴䋱৊ᮃߚ㉏఼ЁՓ⫼ⱘᴵӊ⣀ゟᗻП䯈ⱘ݇㋏Ǆᴈ㋴䋱৊ᮃߚ㉏఼‫؛‬
ᅮ㒭ᅮⳂᷛؐVᯊˈᅲ՟ሲᗻA1ᴵӊ⣀ゟѢᅲ՟ሲᗻA2Ǆ䖭Փᕫᴈ㋴䋱৊ᮃߚ㉏఼ৃҹᣝ✻
ϟᓣ䅵ㅫ 6.20 ᓣЁⱘP(A1,A2|V)˖
P(A1,A2|V)=P(A1|A2,V)P(A2|V) (6.23)
=P(A1|V)P(A2|V) (6.24)
ᓣ 6.23 াᰃ㸼 6-1 Ёὖ⥛ⱘЬ⊩㾘߭ⱘϔ㠀ᔶᓣǄᓣ 6.24 ៤ゟᰃ಴ЎA1೼㒭ᅮVᯊᴵӊ

⣀ゟѢA2ˈ✊ৢ⬅ᴵӊ⣀ゟᗻⱘᅮНৃҹᕫࠄP(A1|A2,V)=P(A1|V)Ǆ
原书页码： 186
೒ 6-3 ϔϾ䋱৊ᮃ㕂ֵ㔥Ǆ
Ꮊ䖍ⱘ㔥㒰㸼⼎њϔ㒘ᴵӊ⣀ゟᗻ‫؛‬ᅮǄ⹂ߛഄ䇈ˈ↣Ͼ㡖⚍೼㒭ᅮ݊⠊㒧⚍ᯊˈᴵӊ⣀ゟ
Ѣ݊䴲ৢҷ㒧⚍Ǆ↣Ͼ㒧⚍݇㘨ϔϾᴵӊὖ⥛㸼ˈᅗᣛᅮњ䆹ব䞣೼㒭ᅮ݊⠊㒧⚍ᯊⱘᴵӊߚᏗǄ
ে䖍߫ߎњ Campfire 㒧⚍ⱘᴵӊὖ⥛㸼ˈ݊Ё Campfire, Storm ੠ BusTourGroup ߚ߿㓽‫ݭ‬Ў CˈSˈ
BǄ
6.11.2 㺞⽰
䋱৊ᮃ㕂ֵ㔥˄ㅔ‫ݭ‬Ў䋱৊ᮃ㔥˅㸼⼎ϔ㒘ব䞣ⱘ㘨ড়ὖ⥛ߚᏗǄ՟བˈ೒ 6-3 Ёⱘ䋱

৊ᮃ㔥㸼⼎њ೼Ꮧᇨব䞣StormˈLightningˈThunderˈForestFireˈCampfire੠BusTourGroup
Ϟⱘ㘨ড়ὖ⥛ߚᏗǄϔ㠀ഄ䇈ˈ䋱৊ᮃ㔥㸼⼎㘨ড়ὖ⥛ߚᏗⱘᮍ⊩ᰃᣛᅮϔ㒘ᴵӊ⣀ゟᗻ‫؛‬
ᅮ˄ᅗ㸼⼎Ўϔ᳝৥᮴⦃೒˅ˈҹঞϔ㒘ሔ䚼ᴵӊὖ⥛䲚ড়Ǆ㘨ড়ぎ䯈Ё↣Ͼব䞣೼䋱৊ᮃ
㔥Ё㸼⼎Ўϔ㒧⚍Ǆᇍ↣ϔব䞣䳔㽕ϸ⾡㉏ൟⱘֵᙃǄ佪‫ˈܜ‬㔥㒰ᓻ㸼⼎ᮁ㿔Āℸব䞣೼㒭
ᅮ݊ゟेࠡ偅ᯊᴵӊ⣀ゟѢ݊䴲ৢ㒻āǄ ᔧҢYࠄXᄬ೼ϔᴵ᳝৥ⱘ䏃ᕘˈ៥Ӏ⿄XᰃYⱘৢ
㒻Ǆ㄀Ѡˈᇍ↣Ͼব䞣᳝ϔϾᴵӊὖ⥛㸼ˈᅗᦣ䗄њ䆹ব䞣೼㒭ᅮ݊ゟेࠡ偅ᯊⱘὖ⥛ߚᏗǄ
ᇍ㔥㒰ব䞣ⱘ‫ܗ‬㒘<Y1…Yn>䌟ҹ᠔Ꮰᳯⱘؐ˄y1…yn˅ⱘ㘨ড়ὖ⥛ৃ⬅ϟ䴶ⱘ݀ᓣ䅵ㅫǄ
n
P ( y1 ,... y n ) P( y
i 1
i | Parents (Yi ))
݊ЁParents(Yi)㸼⼎㔥㒰ЁYiⱘゟेࠡ偅ⱘ䲚ড়Ǆ⊼ᛣ P(yi|Parents(yi))ⱘؐㄝѢϢ㒧⚍
Yi݇㘨ⱘᴵӊὖ⥛㸼ЁⱘؐǄ
Ў䇈ᯢ䖭ϔ⚍ˈ೒ 6-3 㸼⼎њ೼Ꮧᇨব䞣 Storm, Lighting, Thunder, ForestFire, Campfire

ҹঞ BusTourGroup Ϟⱘ㘨ড়ὖ⥛ߚᏗǄ㗗㰥㒧⚍ CampfireǄ㔥㒰㒧⚍੠ᄸ㸼⼎њᮁ㿔˖
Campfire ೼㒭ᅮ݊⠊㒧⚍ Storm ੠ BusTourGroup ᯊᴵӊ⣀ゟѢ݊䴲ৢ㒻 Lighting ੠
Thunder Ǆ䖭ᛣੇⴔϔԚ៥Ӏⶹ䘧њব䞣 Storm ੠ BusTourGroup ⱘؐˈব䞣 Lighting ੠
Thunder ϡӮᦤկ᳝݇ Campfire ⱘ᳈໮ⱘֵᙃǄ೒ে䖍ᰒ⼎њϢব䞣 Campfire 㘨㋏ⱘᴵӊ
ὖ⥛㸼Ǆ↨བ㸼ⱘ᳔ᎺϞϔϾ᭄᥂㸼⼎њҹϟⱘᮁ㿔˖
P(Campfire=True|Storm=True, BusTourGroup=True)=0.4
⊼ᛣℸ㸼াᦤկњ㒭ᅮ݊⠊ব䞣 Storm ੠ BusTourGroup ϟ Campfire ⱘᴵӊὖ⥛Ǆ᠔᳝

ব䞣ⱘሔ䚼ᴵӊὖ⥛㸼ҹঞ⬅㔥㒰᠔ᦣ䗄ⱘϔ㒘ᴵӊ⣀ゟ‫؛‬ᅮˈᦣ䗄њ䆹㔥㒰ⱘᭈϾ㘨ড়ὖ
⥛ߚᏗǄ
䋱৊ᮃ㕂ֵ㔥ⱘϔϾ਌ᓩҎⱘ⡍ᗻ೼Ѣˈᅗᦤկњϔ⾡ᮍ֓ⱘ䗨ᕘҹ㸼⼎಴ᵰⶹ䆚ˈ↨
˄䮾⬉˅ᇐ㟈 Thunder˄ᠧ䳋˅Ǆҹᴵӊ⣀ゟᗻⱘᴃ䇁ˈ
བ Lighting ৃᇚ݊㸼䗄Ў೼㒭ᅮ Lighting
ⱘؐᚙ‫މ‬ϟˈThunder ᴵӊ⣀ゟѢ㔥㒰Ё݊Ҫব䞣Ǆ⊼ᛣℸᴵӊ⣀ゟᗻ‫؛‬ᅮᰃ⬅೒ 6-3 ⱘ䋱
৊ᮃ㔥ⱘᓻᣛᅮⱘǄ
6.11.3 ᧞⨼
ৃҹ⫼䋱৊ᮃ㔥೼㒭ᅮ݊Ҫব䞣ⱘ㾖ᆳؐᯊ᥼⧚ߎᶤѯⳂᷛব䞣˄བ ForestFire˅ⱘؐǄ
ᔧ✊ˈ⬅Ѣ᠔໘⧚ⱘᰃ䱣ᴎব䞣ˈ᠔ҹϔ㠀ϡӮ䌟ќⳂᷛব䞣ϔϾ⹂ߛⱘؐǄⳳℷ䳔㽕᥼⧚
ⱘᰃⳂᷛব䞣ⱘὖ⥛ߚᏗˈᅗᣛᅮњ೼㒭Ϣ݊Ҫব䞣ⱘ㾖ᆳؐᴵӊϟˈⳂᷛব䞣প↣ϔৃ㛑
ؐⱘὖ⥛Ǆ೼㔥㒰Ё᠔᳝݊Ҫব䞣䛑⹂ߛⶹ䘧њҹৢˈ䖭ϔ᥼⧚ℹ偸ᰃᕜㅔऩⱘǄ೼᳈䗮ᐌ
ⱘᚙ‫މ‬ϟ,៥ӀᏠᳯ೼ⶹ䘧ϔ䚼ߚব䞣ⱘؐ˄↨བ Thunder ੠ BusTourGroup Ўҙ᳝ৃ⫼ⱘ㾖
ᆳؐ˅ᯊ㦋ᕫᶤব䞣ⱘὖ⥛ߚᏗ˄བ ForestFire˅Ǆϔ㠀ഄˈ䋱৊ᮃ㔥㒰ৃ⫼Ѣ೼ⶹ䘧ᶤѯব
䞣ⱘؐ៪ߚᏗᯊ䅵ㅫ㔥㒰Ё঺ϔ䚼ߚব䞣ⱘὖ⥛ߚᏗǄ
ϔ㠀ᚙ‫މ‬ϟᇍӏᛣ䋱৊ᮃ㔥㒰ⱘὖ⥛ⱘ⹂ߛ᥼⧚Ꮖ㒣ⶹ䘧ᰃϔϾ NP 䲒乬˄Cooper
1990˅ǄᏆᦤߎњ໮⾡ᮍ⊩೼䋱৊ᮃ㔥㒰Ё䖯㸠ϡ⹂ᅮᗻ᥼⧚ˈࣙᣀ⹂ߛⱘ᥼⧚ҹঞ⡎⡆㊒
ᑺᤶপᬜ⥛ⱘ䖥Ԑ᥼⧚ᮍ⊩Ǆ՟བˈMonte Carlo ᮍ⊩ᦤկњϔ⾡䖥Ԑⱘᮍ⊩ˈ䗮䖛ᇍ᳾㾖
ᆳࠄব䞣䖯㸠䱣ᴎ䞛ḋ˄ Pradham Dagum 1996˅
Ǆ⧚䆎ϞˈेՓᰃ䋱৊ᮃ㔥㒰Ёⱘ䖥Ԑ᥼⧚
гৃᰃ NP 䲒乬˄Dagnm ੠ Luby1993˅Ǆᑌ䖤ⱘᰃˈᅲ䏉Ё䆌໮ᚙ‫މ‬ϟ䖥Ԑⱘᮍ⊩㹿䆕ᯢ
ᰃ᳝ᬜⱘˈᇍѢ䋱৊ᮃ㔥㒰᥼⧚ᮍ⊩ⱘ䅼䆎⬅ Russell & Norvig˄1995˅੠ Jensen˄1996˅
԰ߎǄ
6.11.4 ᆜҖ䍓ਬᯥ㖤‫ؗ‬㖇
ᰃ৺ৃҹ䆒䅵ߎ᳝ᬜⱘㅫ⊩ҹҢ䆁㒗᭄᥂Ёᄺࠄ䋱৊ᮃ㕂ֵ㔥˛䖭ᰃⳂࠡⷨおЁⱘϔ
Ͼ⛺⚍ⱘ䯂乬ǄᇍѢ䖭ϔ䯂乬᳝໮⾡ৃҹ㗗㰥ⱘḚᶊǄ佪‫ܜ‬㔥㒰㒧ᵘৃҹ乘‫ܜ‬㒭ߎˈ៪ৃ⬅
䆁㒗᭄᥂Ё᥼ᕫǄ㄀Ѡˈ᠔᳝ⱘ㔥㒰ব䞣ৃҹⳈ᥹Ң↣Ͼ䆁㒗ḋ՟Ё㾖ᆳࠄˈ៪ᶤѯব䞣ϡ
㛑㾖ᆳࠄǄ
೼㔥㒰㒧ᵘⱘ乘‫ܜ‬ᏆⶹˈᑊϨব䞣ৃҹҢ䆁㒗ḋ՟Ёᅠܼ㦋ᕫᯊˈ䗮䖛ᄺдᕫࠄᴵӊὖ
⥛㸼ህ↨䕗ㅔऩњǄা䳔㽕䈵೼ᴈ㋴䋱৊ᮃߚ㉏఼Ё䙷ḋԄ䅵㸼Ёⱘᴵӊὖ⥛乍Ǆ
㢹㔥㒰㒧ᵘᏆⶹˈԚা᳝ϔ䚼ߚব䞣ؐ㛑೼᭄᥂Ё㾖ᆳࠄˈᄺд䯂乬ህೄ䲒ᕫ໮њǄ䖭
ϔ䯂乬೼ᶤ⾡⿟ᑺϞ㉏ԐѢ೼ҎᎹ⼲㒣㔥㒰Ёᄺд䱤㮣ऩ‫ⱘܗ‬ᴗؐˈ݊Ё䕧ܹ੠䕧ߎ㒧⚍ؐ
⬅䆁㒗ḋ՟㒭ߎˈԚ䱤㮣ऩ‫ؐⱘܗ‬᳾ᣛᅮǄᅲ䰙ϞˈRusstll et al.˄1995˅ᦤߎњϔϾㅔऩⱘ
ẃᑺϞछ䖛⿟ҹᄺдᴵӊὖ⥛㸼Ёⱘ乍Ǆ䖭ϔẃᑺϞछ䖛⿟᧰㋶ϔϾ‫؛‬䆒ぎ䯈ˈᅗᇍᑨѢᴵ
ӊὖ⥛㸼Ё᠔᳝ৃ㛑ⱘ乍Ǆ೼ẃᑺϞछЁ᳔໻࣪ⱘⳂᷛߑ᭄ᰃ㒭ᅮ‫؛‬䆒 h ϟ㾖ᆳࠄ䆁㒗᭄᥂
D ⱘὖ⥛ P(D|h)Ǆᣝ✻ᅮНˈᅗᇍᑨѢᇍ㸼乍᧰㋶ᵕ໻Ԑ✊‫؛‬䆒Ǆ
6.11.5 䍓ਬᯥ㖇Ⲻụᓜрॽ䇣㓹
⬅Russell et al. (1995)㒭ߎⱘẃᑺϞछ㾘߭Փ⫼ⳌᑨѢᅮНᴵӊὖ⥛㸼ⱘখ᭄ⱘlnP(D|h)

ⱘẃᑺᴹՓP(D|h)᳔໻࣪ǄҸwijkҷ㸼ϔϾᴵӊὖ⥛㸼ⱘϔϾ㸼乍Ǆ⹂ߛഄ䆆ˈҸwijkЎ೼㒭
ᅮ⠊㒧⚍Uiপؐuikᯊˈ㔥㒰ব䞣YiؐЎyijⱘὖ⥛Ǆ՟བˈ㢹wijkЎ೒ 6-3 Ёᴵӊὖ⥛㸼Ё᳔ে
Ϟᮍⱘ㸼乍ˈ䙷МYiЎব䞣CampfireˈUiᰃ݊⠊㒧⚍ⱘ‫ܗ‬㒘<Storm, BusTourGroup>ˈyij=Trueˈ
w ln P ( D | h)
ᑊϨuik=<False, False>ǄᇍѢ↣Ͼwijk ˈlnP(D|h)ⱘẃᑺ⬅ᇐ᭄ 㒭ߎǄབϟ䴶ৃ
wwijk
㾕ˈ↣Ͼᇐ᭄ৃབϟ䅵ㅫ˖
w ln P( D | h) P(Yi y ij ,U i u ik | d )
wwijk
¦
d D wijk
˄6.25˅
՟བˈЎ䅵ㅫᇍᑨѢ೒ 6-3 Ё㸼ᎺϞᮍⱘ㸼乍ⱘ lnP(D|h)ⱘᇐ᭄ˈ䳔㽕ᇍ D Ё↣Ͼ䆁㒗

ḋ՟ d 䅵ㅫ P(Campfire=True, Storm=False, BusTourGroup=False | d)Ǆᔧ䖭ѯব䞣ᇍ䆁㒗ḋ՟
d ᮴⊩㾖ᆳࠄᯊˈ䖭ѯ᠔䳔ⱘὖ⥛ৃҹ d Ё㾖ᆳࠄⱘব䞣Ё⫼ᷛ‫ⱘޚ‬䋱৊ᮃ㔥㒰᥼⧚ᕫࠄǄ
ᅲ䰙Ϟ䖭ѯ᠔䳔ⱘ䞣ৃ೼໮᭄䋱৊ᮃ㔥㒰᥼⧚ⱘ䖛⿟Ё䅵ㅫᕫࠄˈ಴ℸ᮴䆎ԩᯊ䋱৊ᮃ㔥㒰
㹿⫼Ѣ᥼⧚ˈᑊϨৢᴹ㦋ᕫњᮄⱘ䆕᥂ˈᄺд䖛⿟޴Тϡ䳔㽕䰘ࡴⱘ㢅䫔ˈǄ
ϟ䴶ḍ᥂Russell et al.˄1995˅᥼ᇐᓣ 6.25Ǆᴀ㡖ⱘৢ䴶೼㄀ϔ⃵䯙䇏ᯊৃҹ㹿䏇䖛ˈ

㗠ϡӮϻ༅䖲㓁ᗻˈЎՓ䆄োㅔऩ࣪ˈϟ䴶ⱘ᥼ᇐᇚ⫼Ph(D)ᴹㅔ‫ݭ‬P(D|h)Ǆ಴ℸˈ៥Ӏⱘ䯂
wPh ( D)
乬ᰃ㦋ᕫᇐ᭄䲚ড় ˄ᇍ᠔᳝ⱘi, j, k˅ⱘẃᑺˈ‫؛‬ᅮ೼᭄᥂䲚DЁⱘ৘ḋ՟d䛑ᰃ⣀
wwijk
ゟᢑপⱘǄৃᇚℸᇐ᭄‫ݭ‬Ў
w ln Ph ( D) w
ln Ph (d )
wwijk wwijk dD
w ln Ph (d )
¦
d D wwijk
1 wPh (d )
¦ P (d )
d D wwijk
h
᳔ৢϔℹՓ⫼њㄝᓣ w ln f ( x) 1 wf ( x) Ǆ⦄೼ৃҹᓩܹব䞣Yi ੠Ui=Parents(Yi)ⱘؐˈ

wx f ( x) wx
ᮍ⊩ᰃ䗮䖛೼݊ৃ㛑ⱘؐyij´੠uik´Ϟ䅵ㅫࡴ੠Ǆ
w ln Ph ( D) 1 w
wwijk
¦ P (d ) ww ¦ P (d | y
d D
h ij ' , u ik ' ) Ph ( y ij ' , u ik ' )
h ijk j ', k '
1 w
¦ P (d ) ww ¦ P (d | y
d D
h ij ' , u ik ' ) Ph ( y ij ' | u ik ' ) Ph (u ik ' )
h ijk j ', k '
᳔ৢϔℹᴹ㞾Ѣὖ⥛ⱘЬ⊩݀ᓣˈ㾕㸼 6-1Ǆ⦄೼㗗㰥Ϟ䴶᳔ৢϔᓣ᳔ে䖍ⱘࡴ੠乍Ǆ
㒭ᅮњwijk įPh(yij|uik)ˈ೼ℸࡴ੠Ёᚳϔ w ϡㄝѢ 0 ⱘ乍ᰃ݊Ёj´=j੠i´=iⱘ乍ˈ಴ℸ˖
wwijk
w ln Ph ( D) 1 w
wwijk
¦ P (d ) ww
d D
Ph (d | yij , uik ) Ph ( yij | uik ) Ph (uik )
h ijk
1 w
¦ P (d ) ww
dD
Ph (d | yij , uik ) wijk Ph (uik )
h ijk
1
¦ P (d ) P (d | y , u
d D
h ij ik ) Ph (uik )
h
ᑨ⫼䋱৊ᮃ݀ᓣᴹ䞡‫ݭ‬Ph(d|yij,uik)ৃᕫ
w ln Ph ( D) 1 Ph ( yij , uik | d ) Ph (d ) Ph (uik )

wwijk
¦ P (d )
dD Ph ( yij , uik )
h
Ph ( yij , uik | d ) Ph (uik )

¦
d D Ph ( yij , uik )
Ph ( yij , uik | d )
¦
dD Ph ( yij | uik )
Ph ( yij , uik | d )
¦
d D wijk
˄6.26˅
䖭ḋ៥ӀᏆᇐߎњᓣ 6.25 ЁⱘẃᑺǄ೼ᦣ䗄ẃᑺϞछ䆁㒗ࠡ䖬㽕㗗㰥ϔϾ䯂乬Ǆ⹂ߛ

ഄ䇈ˈ៥Ӏ㽕∖ᔧᴗؐwijk᳈ᮄᯊˈᅗӀᖙ乏ֱᣕ೼ऎ䯈[0ˈ1]П䯈ˈ䖭ḋᠡᰃ᳝ᬜⱘὖ⥛Ǆ
៥Ӏ䖬㽕∖jwijkᇍ᠔᳝ⱘi,kֱᣕЎ 1Ǆ䖭ѯ䰤ࠊৃ⬅ϔϾϸℹ偸ⱘᴗؐ᳈ᮄᴹ⒵䎇Ǆ佪‫ܜ‬
⫼ẃᑺϞछᴹ᳈ᮄ↣Ͼwijk˖
Ph ( y ij , u ik | d )
wijk m wijk K ¦
d D wijk
݊Ё¨ᰃϔᇣⱘᐌ䞣ˈ⿄Ўᄺд⥛Ǆ݊⃵ˈ‫ݡ‬ᇚᴗؐwijkᔦϔ࣪ˈҹֱ䆕Ϟ䴶ⱘ䰤ࠊᕫ
ࠄ⒵䎇ǄབRussell᠔ᦣ䗄ⱘ䙷ḋˈ䖭ϔ䖛⿟ᇚᬊᬯࠄ䋱৊ᮃ㔥㒰Ёⱘᴵӊὖ⥛ⱘϔϾሔ䚼ⱘ
ᵕ໻Ԑ✊‫؛‬䆒Ǆ
བ೼݊Ҫ෎Ѣẃᑺⱘᮍ⊩Ё䙷ḋˈ䆹ㅫ⊩াֱ䆕ᇏᡒࠄሔ䚼᳔Ӭ㾷Ǆ᳓ҷẃᑺϞछⱘϔ
Ͼㅫ⊩ᰃ EM ㅫ⊩ˈᅗ೼ 6.12 㡖Ё䅼䆎ˈᅗгাᡒሔ䚼ᵕ໻ৃ㛑ᗻⱘ㾷Ǆ
6.11.6 ᆜҖ䍓ਬᯥ㖇Ⲻ㔉ᶺ
ᔧ㔥㒰㒧ᵘ乘‫ܜ‬᳾ⶹᯊˈᄺд䋱৊ᮃ㔥㒰гᕜೄ䲒Ǆ Cooper & Herskovits˄1992˅ ᦤ

ߎњϔϾ䋱৊ᮃ䆘ߚᑺ䞣˄Bayesian scoring metric˅ҹҢϡৠ㔥㒰Ё䖯㸠䗝ᢽǄҪӀ䖬ᦤߎ
ϔϾ⿄Ў K2 ⱘਃথᓣ᧰㋶ㅫ⊩⫼Ѣ೼᭄᥂ᅠܼৃ㾖ᆳࠄᯊᄺд㔥㒰㒧ᵘǄབ໮᭄ᄺд㔥㒰
㒧ᵘⱘㅫ⊩ˈK2 ᠻ㸠ⱘᰃϔϾ䋾ှ᧰㋶ˈҹ೼㔥㒰ⱘ໡ᴖᗻ੠ᅗ೼䆁㒗᭄᥂Ϟⱘ㊒ᑺП䯈
԰ߎᡬЁǄ೼ϔϾᅲ偠ЁˈK2 㹿㒭Ϣ 3000 Ͼ䆁㒗ḋ՟ˈ䖭ѯḋ՟ᰃҢࣙ৿њ 37 Ͼ㡖⚍੠
46 ᴵᓻⱘ᠟Ꮉ߯ᓎⱘ䋱৊ᮃ㔥㒰Ё䱣ᴎᢑপⱘǄ䖭ϔ㔥㒰ᦣ䗄њ೼ϔए䰶ⱘ᠟ᴃᅸЁ┰೼
ⱘ㒚㦠䯂乬Ǆ䰸њ᭄᥂ҹ໪ˈ⿟ᑣ䖬㹿㒭Ϣ 37 Ͼব䞣ⱘ߱ྟᥦᑣˈᅗϢᅲ䰙㔥㒰Ёব䞣П
䯈ⱘ‫أ‬ᑣ݇㋏ϔ㟈Ǆ䆹⿟ᑣ៤ࡳഄ߯ᓎߎњϢℷ⹂㔥㒰㒧ᵘ޴Тϔḋⱘ䋱৊ᮃ㔥ˈ䰸њϔϾ
ϡℷ⹂ഄ㹿ߴ䰸ⱘ੠ϔϡℷ⹂ഄ㹿ࡴܹⱘᓻǄ
෎Ѣ㑺ᴳⱘᄺд䋱৊ᮃ㔥㒰㒧ᵘⱘ䗨ᕘгᏆ㹿ᓔথߎᴹ˄՟བˈSpirtes et al. 1993˅

Ǆ䖭
ѯ䗨ᕘҢ᭄᥂Ё᥼䆎ߎϡⳌ݇੠Ⳍ݇ⱘ݇㋏ˈ✊ৢ⫼䖭ѯ݇㋏ᴹᵘ䗴䋱৊ᮃ㔥Ǆ݇Ѣᔧࠡᄺ
д䋱৊ᮃ㔥ⱘ䗨ᕘⱘ䇗ⷨ⬅ Heckerman˄1995˅੠ Buntine˄1994˅㒭ߎǄ
6.12 EM ㇍⌋
೼䆌໮ᅲ䰙ⱘᄺд䯂乬ḚᶊЁˈⳌ݇ᅲ՟⡍ᕕЁা᳝ϔ䚼ߚৃ㾖ᆳࠄǄ՟བˈ೼䆁㒗៪
Փ⫼೒ 6-3 Ёⱘ䋱৊ᮃ㕂ֵ㔥ᯊˈৃ㛑㔥㒰ব䞣 Storm, Lighting, Thunder, ForestFire, Campfire
੠ BusTourGroup Ёা᳝݊ϔϾᄤ䲚㛑೼᭄᥂Ё㾖ᆳࠄǄᏆ᳝䆌໮ᮍ⊩㹿ᦤߎ⫼ᴹ໘⧚ᄬ೼
᳾㾖ᆳࠄব䞣ᯊⱘ䯂乬Ǆབ೼㄀ 3 ゴⳟࠄⱘˈ㢹ᶤѯব䞣᳝ᯊ㛑㾖ᆳࠄˈ᳝ᯊϡ㛑ˈ䙷Мৃ
ҹ⫼㾖ᆳࠄⱘᅲ՟এ乘⌟᳾㾖ᆳࠄⱘǄ೼ᴀ㡖Ёᦣ䗄 EM ㅫ⊩˄ Dempster et al. 1977˅ˈ䖭
ᰃᄬ೼䱤৿ব䞣ᯊᑓ⊯Փ⫼ⱘϔ⾡ᄺдᮍ⊩ǄEM ㅫ⊩ৃ⫼Ѣব䞣ⱘؐҢᴹ≵᳝㹿Ⳉ᥹㾖ᆳ
ࠄⱘᚙᔶˈা㽕䖭ѯব䞣᠔䙉ᕾⱘὖ⥛ߚᏗⱘϔ㠀ᔶᓣᏆⶹǄEM ㅫ⊩Ꮖ㹿⫼Ѣ䆁㒗䋱৊ᮃ
㕂ֵ㔥˄㾕 Heckerman 1995˅ҹঞ 8.4 㡖䅼䆎ⱘᕘ৥෎ߑ᭄˄radial basis function˅㔥㒰Ǆ
EM ㅫ⊩䖬ᰃ䆌໮䴲ⲥⴷ㘮㉏ㅫ⊩ⱘ෎⸔˄བ Cheeseman et al. 1988˅ˈ㗠Ϩᅗᰃ⫼Ѣᄺд䚼
ߚৃ㾖ᆳ偀ᇨৃ໿῵ൟ˄Partially Observable Markov Model˅ⱘᑓ⊯Փ⫼ⱘ Baum-Welch ࠡ
৥ৢ৥ㅫ⊩ⱘ෎⸔˄Rabiner 1989˅Ǆ
6.12.1 զ䇗 k Ѡ 儎ᯥ࠼ᐹⲺൽ‫ٲ‬
ҟ㒡EMㅫ⊩᳔ᮍ֓ⱘᮍ⊩ᰃ䗮䖛ϔϾ՟ᄤǄ㗗㰥᭄᥂Dᰃϔᅲ՟䲚ড়ˈᅗ⬅kϾϡৠℷ
ᗕߚᏗⱘ⏋ড়᠔ᕫߚᏗ᠔⫳៤Ǆ䆹䯂乬Ḛᶊ೼೒ 6-4 Ё⼎ߎˈ݊Ёk=2 㗠Ϩᅲ՟Ў⊓ⴔx䕈ᰒ
⼎ⱘ⚍Ǆ↣Ͼᅲ՟Փ⫼ϔϾϸℹ偸䖛⿟ᔶ៤Ǆ佪‫ܜ‬њ䱣ᴎ䗝ᢽkϾℷᗕߚᏗ݊ЁПϔǄ݊⃵
䱣ᴎব䞣xiᣝ✻ℸ䗝ᢽⱘߚᏗ⫳៤Ǆ䖭ϔ䖛⿟ϡᮁ䞡໡ˈ⫳៤ϔ㒘᭄᥂⚍བ೒᠔⼎ǄЎՓ䅼
䆎ㅔऩ࣪ˈ៥Ӏ㗗㰥ϔϾㅔऩᚙᔶˈेऩϾℷᗕߚᏗⱘ䗝ᢽ෎Ѣ㒳ϔⱘὖ⥛䖯㸠䗝ᢽˈᑊϨ
kϾℷᗕߚᏗ᳝ⳌৠⱘᮍᏂ³2ˈϨ³2ᏆⶹǄᄺдӏࡵᰃ䕧ߎϔϾ‫؛‬䆒h=<1…k>ˈᅗᦣ䗄
њkϾߚᏗЁ↣ϔϾߚᏗⱘഛؐǄ៥ӀᏠᳯᇍ䖭ѯഛؐᡒࠄϔϾᵕ໻Ԑ✊‫؛‬䆒ˈेϔϾՓ
P(D|h)᳔໻࣪ⱘ‫؛‬䆒hǄ
原书页码： 192
೒ 6-4 ⬅ϸϾ‫݋‬᳝ⳌㄝᮍᏂ³2ⱘℷᗕߚᏗⱘ⏋ড়⫳៤ⱘᅲ՟䲚Ǆ
ᅲ՟Ў⊓ⴔ x 䕈ᰒ⼎ⱘ⚍䲚ǄབᵰℷᗕߚᏗⱘഛؐ᳾ⶹˈEM ㅫ⊩ৃ⫼Ѣ᧰㋶݊ᵕ໻Ԑ✊Ԅ䅵Ǆ
⊼ᛣࠄˈᔧ㒭ᅮҢϔϾℷᗕߚᏗЁᢑপⱘ᭄᥂ᅲ՟x1, x2, …, xmᯊˈᕜᆍᯧ䅵ㅫ䆹ߚᏗⱘ

ഛؐⱘᵕ໻Ԑ✊‫؛‬䆒Ǆ䖭ϔᇏᡒऩϾߚᏗഛؐⱘ䯂乬াᰃ೼ 6.4 㡖ⱘᓣ 6.6 Ё䅼䆎ⱘ䯂乬ⱘ
ϔϾ⡍՟ˈ೼݊Ё៥Ӏ䆕ᯢњᵕ໻Ԑ✊‫؛‬䆒ᰃՓmϾ䆁㒗ᅲ՟Ϟⱘ䇃Ꮒᑇᮍ੠᳔ᇣ࣪ⱘ‫؛‬
䆒ǄՓ⫼ᔧࠡⱘ䆄ো䞡ᮄ㸼䗄ϔϟᓣ 6.6ˈৃҹᕫࠄ˖
m
P ML arg min ¦ ( xi P ) 2 ˄6.27˅
P i 1
೼ℸᚙ‫މ‬ϟˈ䇃Ꮒᑇᮍ੠ᰃ⬅ḋᴀഛ᳔ؐᇣ࣪ⱘ˖
1 m
P ML ¦ xi
mi1
˄6.28˅
✊㗠ˈ೼䖭䞠៥Ӏⱘ䯂乬⍝ঞࠄkϾϡৠℷᗕߚᏗⱘ⏋ড়ˈ㗠Ϩ៥Ӏϡ㛑ⶹ䘧ાϾᅲ՟
ᰃાϾߚᏗѻ⫳ⱘǄ಴ℸ䖭ᰃϔϾ⍝ঞ䱤㮣ব䞣ⱘ‫݌‬ൟ՟ᄤǄ೼೒ 6-4 ⱘ՟ᄤЁˈৃᡞ↣Ͼ
ᅲ՟ⱘᅠᭈᦣ䗄ⳟ԰ᰃϝ‫ܗ‬㒘<xi, zi1, zi2>ˈ݊Ёxiᰃ㄀iϾᅲ՟ⱘ㾖⌟ؐˈzi1੠zi2㸼⼎ϸϾℷ
ᗕߚᏗЁાϾ㹿⫼Ѣѻ⫳ؐxiǄ⹂ߛഄ䆆ˈzij೼xi⬅㄀jϾℷᗕߚᏗѻ⫳ᯊؐЎ 1ˈ৺߭Ў 0Ǆ
䖭䞠xiᰃᅲ՟ⱘᦣ䗄ЁᏆ㾖ᆳࠄⱘব䞣ˈzi1੠zi2ᰃ䱤㮣ব䞣Ǆབᵰzi1੠zi2ⱘؐৃⶹˈህৃҹ
⫼ᓣ 6.27 ᴹ㾷‫އ‬ഛؐ1੠2Ǆ಴ЎᅗӀ᳾ⶹˈ಴ℸ៥Ӏা㛑⫼EMㅫ⊩Ǆ
EMㅫ⊩ᑨ⫼Ѣ៥Ӏⱘkഛؐ䯂乬ˈⳂⱘᰃ᧰㋶ϔϾᵕ໻Ԑ✊‫؛‬䆒ˈᮍ⊩ᰃḍ᥂ᔧࠡ‫؛‬䆒
<1…k>ϡᮁഄ‫ݡ‬Ԅ䅵䱤㮣ব䞣zijⱘᳳᳯؐǄ✊ৢ⫼䖭ѯ䱤㮣ব䞣ⱘᳳᳯؐ䞡ᮄ䅵ㅫᵕ໻
Ԑ✊‫؛‬䆒Ǆ䖭䞠佪‫ᦣܜ‬䗄䖭ϔᅲ՟࣪ⱘEMㅫ⊩ˈҹৢᇚ㒭ߎEMㅫ⊩ⱘϔ㠀ᔶᓣǄ
ЎњԄ䅵೒ 6-4 ЁⱘϸϾഛؐˈEMㅫ⊩佪‫ܜ‬ᇚ‫؛‬䆒߱ྟ࣪Ўh=<1, 2>ˈ݊Ё1੠

2Ўӏᛣⱘ߱ྟؐǄ✊ৢ䞡໡ҹϟⱘϸϾℹ偸ҹ䞡Ԅ䅵hˈⳈࠄ䆹䖛⿟ᬊᬯࠄϔϾ〇ᅮⱘh
ؐǄ
ℹ偸 1˖䅵ㅫ↣Ͼ䱤㮣ব䞣zijⱘᳳᳯؐE[zij]ˈ‫؛‬ᅮᔧࠡ‫؛‬䆒h=<1, 2>៤ゟǄ
ℹ偸 2˖䅵ㅫϔϾᮄⱘᵕ໻Ԑ✊‫؛‬䆒h´=<1´, 2´>ˈ‫؛‬ᅮ⬅↣Ͼ䱤㮣ব䞣zij᠔পⱘؐ
Ў㄀ 1 ℹЁᕫࠄⱘᳳᳯؐE[zij]ˈ✊ৢᇚ‫؛‬䆒h=<1, 2>᳓ᤶЎᮄⱘ‫؛‬䆒h´=<1´, 2´>ˈ
✊ৢᕾ⦃Ǆ
⦄೼㗗ᆳ㄀ϔℹᰃབԩᅲ⦄ⱘǄℹ偸 1 㽕䅵ㅫ↣Ͼzij ⱘᳳᳯؐǄℸE[zij]ℷᰃᅲ՟xi⬅㄀

jϾℷᗕߚᏗ⫳៤ⱘὖ⥛˖
p( x xi | P Pj)
E[ zij ]
¦
2
n 1
p( x xi | P Pn )
1
( xi P j ) 2
2V 2
e
1
( xi P n ) 2
¦
2
2V 2
n 1
e
಴ℸ㄀ϔℹৃ⬅ᇚᔧࠡؐ<1, 2>੠ᏆⶹⱘxiҷܹࠄϞᓣЁᅲ⦄Ǆ
೼㄀ѠℹˈՓ⫼㄀ 1 ℹЁᕫࠄⱘE[zij] ᴹᇐߎϔᮄⱘᵕ໻Ԑ✊‫؛‬䆒h´=<1´, 2´>Ǆབ

ৢ䴶ᇚ䅼䆎ࠄⱘˈ䖭ᯊⱘᵕ໻Ԑ✊‫؛‬䆒Ў˖
¦
m
E[ zij ]xi
Pj m i 1
¦
m
i 1
E[ zij ]
⊼ᛣℸ㸼䖒ᓣ㉏ԐѢᓣ 6.28 Ёⱘḋᴀഛؐˈᅗ⫼ѢҢऩϾℷᗕߚᏗЁԄ䅵Ǆᮄⱘ㸼

䖒ᓣাᰃᇍjⱘࡴᴗḋᴀഛؐˈ↣Ͼᅲ՟ⱘᴗ䞡Ў݊⬅㄀jϾℷᗕߚᏗѻ⫳ⱘᳳᳯؐǄ
Ϟ䴶Ԅ䅵kϾℷᗕߚᏗഛؐⱘㅫ⊩ᦣ䗄њEMᮍ⊩ⱘ㽕⚍˖ेᔧࠡⱘ‫؛‬䆒⫼ѢԄ䅵᳾ⶹব
䞣ˈ㗠䖭ѯব䞣ⱘᳳᳯؐ‫ݡ‬㹿⫼Ѣᬍ䖯‫؛‬䆒Ǆৃҹ䆕ᯢˈ೼ℸㅫ⊩㄀ϔ⃵ᕾ⦃ЁˈEMㅫ⊩
㛑ՓԐ✊ᗻP(D|h)๲ࡴˈ䰸䴲ᅗᏆ䖒ࠄሔ䚼ⱘ᳔໻Ǆ಴ℸ䆹ㅫ⊩ᬊᬯࠄᇍѢ<1, 2>ⱘϔ
Ͼሔ䚼ᵕ໻ৃ㛑ᗻ‫؛‬䆒Ǆ
6.12.2 EM ㇍⌋Ⲻж㡢㺞䘦
Ϟ䴶ⱘEMㅫ⊩䩜ᇍⱘᰃԄ䅵⏋ড়ℷᗕߚᏗഛؐⱘ䯂乬Ǆ᳈ϔ㠀ഄˈEMㅫ⊩ৃ⫼Ѣ䆌໮
䯂乬Ḛᶊˈ݊Ё䳔㽕Ԅ䅵ϔ㒘ᦣ䗄෎‫ޚ‬ὖ⥛ߚᏗⱘখ᭄©ˈা㒭ᅮњ⬅ℸߚᏗѻ⫳ⱘܼ䚼᭄
᥂Ё㛑㾖ᆳࠄⱘϔ䚼ߚǄ೼Ϟ䴶ⱘѠഛؐ䯂乬Ёˈᛳ݈䍷ⱘখ᭄Ў©=<1, 2>ˈ㗠ܼ䚼
᭄᥂Ўϝ‫ܗ‬㒘<xi, zi1, zi2>ˈ㗠া᳝xiৃ㾖ᆳࠄˈϔ㠀ഄҸX=<x1, …, xm>ҷ㸼೼ৠḋⱘᅲ՟Ё
᳾㾖ᆳࠄⱘ᭄᥂ˈᑊҸY=XĤZҷ㸼ܼԧ᭄᥂Ǆ⊼ᛣࠄ᳾㾖ᆳࠄⱘZৃ㹿ⳟ԰ϔ䱣ᴎব䞣ˈᅗ
ⱘὖ⥛ߚᏗձ䌪Ѣ᳾ⶹখ᭄©੠Ꮖⶹ᭄᥂XǄ㉏ԐഄˈYᰃϔ䱣ᴎব䞣ˈ಴Ўᅗᰃ⬅䱣ᴎব
䞣ZᴹᅮНⱘǄ೼ᴀ㡖ⱘৢ㓁䚼ߚˈᇚᦣ䗄EMㅫ⊩ⱘϔ㠀ᔶᓣǄՓ⫼hᴹҷ㸼খ᭄©ⱘ‫؛‬䆒
ؐˈ㗠h´ҷ㸼೼EMㅫ⊩ⱘ↣⃵䗁ҷЁׂᬍⱘ‫؛‬䆒Ǆ
EM ㅫ⊩䗮䖛᧰ᇏՓ E[lnP(Y|h´)]᳔໻ⱘ h´ᴹᇏᡒᵕ໻Ԑ✊‫؛‬䆒 h´Ǆℸᳳᳯؐᰃ೼ Y ᠔
䙉ᕾⱘὖ⥛ߚᏗϞ䅵ㅫˈℸߚᏗ⬅᳾ⶹখ᭄©⹂ᅮǄ㗗㰥ℸ㸼䖒ᓣおコᛣੇњҔМǄ佪‫ܜ‬
P(Y|h´)ᰃ㒭ᅮ‫؛‬䆒 h´ϟܼ䚼᭄᥂ Y ⱘԐ✊ᗻǄ݊ড়⧚ᗻ೼Ѣ៥Ӏ㽕ᇏᡒϔϾ h´Փ䆹䞣ⱘᶤ
ߑ᳔᭄ؐ໻࣪Ǆ݊⃵Փ䆹䞣ⱘᇍ᭄ lnP(Y|h´)᳔໻࣪гՓ P(Y|h´)᳔໻࣪ˈབᏆ㒣ҟ㒡䖛ⱘ䙷ḋǄ
㄀ϝˈᓩܹᳳᳯؐ E[lnP(Y|h´)]ᰃ಴Ўܼ䚼᭄᥂ Y ᴀ䑿гᰃϔ䱣ᴎব䞣ǄᏆⶹܼ䚼᭄᥂ Y ᰃ
㾖ᆳࠄⱘ X ੠᳾㾖ᆳࠄⱘ Z ⱘড়ᑊˈ៥Ӏᖙ乏೼᳾㾖ᆳࠄⱘ Z ⱘৃ㛑ؐϞপᑇഛˈᑊҹⳌ
ᑨⱘὖ⥛ЎᴗؐǄᤶ㿔Пˈ㽕೼䱣ᴎব䞣 Y 䙉ᕾⱘὖ⥛ߚᏗϞপᳳᳯؐ E[lnP(Y|h´)]Ǆ䆹ߚ
Ꮧ⬅ᅠܼᏆⶹⱘ X ؐࡴϞ Z ᳡ҢⱘߚᏗᴹ⹂ᅮǄ
Y 䙉Ңⱘὖ⥛ߚᏗᰃҔМ˛ϔ㠀ᴹ䇈ϡ㛑ⶹ䘧ℸߚᏗˈ಴Ўᅗᰃ⬅ᕙԄ䅵ⱘ©খ᭄⹂ᅮ
ⱘǄ✊㗠ˈEM ㅫ⊩Փ⫼݊ᔧࠡⱘ‫؛‬䆒 h ҷ᳓ᅲ䰙খ᭄©ˈҹԄ䅵 Y ⱘߚᏗǄ⦄ᅮНϔߑ᭄
Q(h´|h)ˈᅗᇚ E[lnP(Y|h´)]԰Ў h´ⱘϔϾߑ᭄㒭ߎˈ೼©=h ੠ܼ䚼᭄᥂ Y ⱘ㾖ᆳࠄⱘ䚼ߚ X
ⱘ‫؛‬ᅮПϟǄ
Q ( h'| h) E[ln p (Y | h' ) | h, X ]
ᇚ Q ߑ᭄‫ݭ‬៤ Q(h´|h)ᰃЎњ㸼⼎݊ᅮНᰃ೼ᔧࠡ‫؛‬䆒 h ㄝѢ©ⱘ‫؛‬ᅮϟǄ೼ EM ㅫ⊩

ⱘϔ㠀ᔶᓣ䞠ˈᅗ䞡໡ҹϟϸϾℹ偸Ⳉ㟇ᬊᬯǄ
ℹ偸 1˖Ԅ䅵˄E˅ℹ偸˖Փ⫼ᔧࠡ‫؛‬䆒 h ੠㾖ᆳࠄⱘ᭄᥂ X ᴹԄ䅵 Y Ϟⱘὖ⥛ߚᏗҹ

䅵ㅫ Q(h´|h)Ǆ
Q (h' | h) m E[ln P (Y | h' ) | h, X ]
ℹ偸 2˖᳔໻࣪˄M˅ℹ偸˖ᇚ‫؛‬䆒 h ᳓ᤶЎՓ Q ߑ᳔᭄໻࣪ⱘ‫؛‬䆒 h´˖
h m arg max Q(h'| h)

h'
ᔧߑ᭄ Q 䖲㓁ᯊˈEM ㅫ⊩ᬊᬯࠄԐ✊ߑ᭄ P(Y|h´)ⱘϔϾϡࡼ⚍Ǆ㢹ℸԐ✊ߑ᭄᳝ऩ

Ͼⱘ᳔໻ؐᯊˈEM ㅫ⊩ৃҹᬊᬯࠄ䖭Ͼᇍ h´ⱘܼሔⱘᵕ໻Ԑ✊Ԅ䅵Ǆ৺߭ˈᅗাֱ䆕ᬊᬯ
ࠄϔϾሔ䚼᳔໻ؐǄ಴ℸˈEM Ϣ݊Ҫ᳔Ӭ࣪ᮍ⊩᳝ৠḋⱘሔ䰤ᗻˈབ㄀ 4 ゴ䅼䆎ⱘẃᑺϟ
䰡ˈ㒓ᗻ᧰㋶੠বᔶẃᑺㄝǄ
6.12.3 k-ൽ‫ٲ‬㇍⌋Ⲻ᧞ሲ
Ў䇈ᯢϔ㠀ⱘEMㅫ⊩ˈ៥Ӏ⫼ᅗᴹ᥼ᇐ 6.12.1 㡖ЁԄ䅵kϾℷᗕߚᏗ⏋ড়ഛؐⱘㅫ⊩Ǆ

བϞ᠔䅼䆎ˈk-ഛؐㅫ⊩ᰃЎњԄ䅵kϾℷᗕߚᏗⱘഛؐ©=<1, …, k>ǄᏆ᳝ⱘ᭄᥂Ў㾖
ᆳࠄⱘX={<xi>}ˈ䖭䞠ⱘ䱤㮣ব䞣Z={<zi1, …, zik>}㸼⼎kϾℷᗕߚᏗЁાϔϾ⫼Ѣ⫳៤xiǄ
㽕ᑨ⫼EMㅫ⊩ˈᖙ乏᥼ᇐߎৃ⫼Ѣk-ഛؐ䯂乬ⱘ㸼䖒ᓣQ(h´|h)Ǆ佪‫᥼ܜ‬ᇐߎlnp(Y|h´)
ⱘ㸼䖒ᓣǄ⊼ᛣᇍ↣Ͼᅲ՟yi=<xi, zi1, …, zik>ⱘὖ⥛p(yi|h´)ৃ㹿‫ݭ‬԰˖
1
¦ j 1 z ij ( xi P ' j ) 2
k
1
2V 2
p( yi | h' ) p( xi , zi1 ,..., zik | h' ) e
2SV 2
㽕偠䆕ℸᓣˈᖙ乏⊼ᛣা᳝ϔϾzijؐЎ 1ˈ݊ҪⱘЎ 0Ǆ಴ℸˈ䆹ᓣ㒭ߎњ⬅᠔䗝ⱘℷ

ᗕߚᏗ⫳៤ⱘxiⱘὖ⥛ߚᏗǄᏆⶹњऩϾᅲ՟ⱘߚᏗp(yi|h´)ˈᇍ᠔᳝mϾᅲ՟ⱘὖ⥛ⱘᇍ᭄
lnP(Y|h´)Ў˖
m
ln P(Y | h' ) ln p ( yi | h' )
i 1
¦ ln p( y | h' )
i 1
i
§m
1 1 k ·
¦ ¨
¨
i 1©
ln ¦
2SV 2 2V j 1
2
z ij ( xi P ' j ) 2
¸
¸
¹
᳔ৢˈᖙ乏೼Y᠔䙉Ңⱘὖ⥛ߚᏗˈгህᰃYⱘ᳾㾖ᆳࠄ䚼ߚzij䙉Ңⱘὖ⥛ߚᏗϞˈ䅵ㅫ
ℸlnP(Y|h´)ⱘഛؐǄ⊼ᛣϞ䴶lnP(Y|h´)ⱘ㸼䖒ᓣЎ䖭ѯzijⱘ㒓ᗻߑ᭄Ǆϔ㠀ⱘˈᇍZⱘӏᛣ㒓
ᗻߑ᭄f(z)ᴹ䇈ˈϟ䴶ⱘㄝᓣ៤ゟ˖
E[f(z)]=f(E[z])
ḍ᥂ℸㄝᓣˈৃᕫ˖
ªm § 1 1 k ·º
E[ln P (Y | h' )] E «¦ ¨¨ ln 2 ¦ zij ( xi P ' j ) 2 ¸¸»
«¬ i 1 © 2SV 2 2V j 1 ¹¼»
§m
1 1 k ·
¦ ¨
¨
i 1©
ln ¦
2SV 2 2V j 1
2
E[ zij ]( xi P ' j ) 2 ¸¸
¹
ὖᣀഄ䇈ˈk-ഛؐ䯂乬Ёߑ᭄ Q(h´|h)Ў
m § 1 1 k ·
Q ( h'| h) ¦ ¨¨ ln 2SV 2

2V 2 ¦ E[ z ij ]( xi P ' j ) 2 ¸¸
i 1 © j 1 ¹
݊Ёh´=<1´, …, k´>ˈ㗠E[zij]෎Ѣᔧࠡ‫؛‬䆒h੠㾖ᆳࠄⱘ᭄᥂X䅵ㅫᕫߎǄབࠡ᠔䅼
䆎˖
1
( xi P j ) 2
2V 2
e
E[ zij ] 1
(6.29)
( xi P n ) 2
¦
k
2V 2
n 1
e
಴ℸˈEMㅫ⊩ⱘ㄀ 1 ℹ˄Ԅ䅵ℹ˅෎ѢԄ䅵ⱘE[zij]乍ᅮНњQߑ᭄Ǆ㄀ 2 ℹ˄᳔໻࣪

ℹ˅᥹ⴔᇏᡒՓℸQߑ᳔᭄໻ⱘؐ1´, …, k´Ǆ೼ᔧࠡ՟ᄤЁ˖
m § ·
1 1 k
arg max Q(h'| h) arg max ¦ ¨¨ ln ¦ E[ z ]( x P ' ) 2
¸
¸
2SV 2 2V j 1
2 ij i j
h' h' i 1© ¹
m k
arg min ¦¦ E[ zij ]( xi P ' j ) 2 ˄6.30˅
h' i 1 j 1
಴ℸˈ䖭䞠ⱘᵕ໻Ԑ✊‫؛‬䆒Փᑇᮍ䇃Ꮒⱘࡴᴗ੠᳔ᇣ࣪њˈ݊Ё↣Ͼᅲ՟xiᇍ䇃Ꮒⱘ䋵
⤂j´ᴗ䞡ЎE[zij]Ǆ⬅ㄝᓣ 6.30 㒭ߎⱘ䞣ᰃ䗮䖛ᇚ↣Ͼj´䆒Ўࡴᴗḋᴀഛؐᴹ᳔ᇣ࣪Ǆ
¦
m
E[ zij ]xi
Pj m i 1
˄6.31˅
¦
m
i 1
E[ zij ]
ᓣ 6.29 ੠ᓣ 6.31 ᅮНњ 6.12.1 㡖ЁᅮНⱘ k-ഛؐㅫ⊩ЁⱘϸϾℹ偸Ǆ
ὖ⥛ᄺдᮍ⊩߽⫼˄ᑊϨ㽕∖˅݇Ѣϡৠ‫؛‬䆒ⱘ‫ܜ‬偠ὖ⥛ҹঞ೼㒭ᅮ‫؛‬䆒ᯊ㾖ᆳࠄ
ϡৠ᭄᥂ⱘὖ⥛ⱘⶹ䆚Ǆ䋱৊ᮃᮍ⊩߭ᦤկњὖ⥛ᄺдᮍ⊩ⱘ෎⸔Ǆ䋱৊ᮃᮍ⊩䖬
ৃ෎Ѣ䖭ѯ‫ܜ‬偠੠᭄᥂㾖ᆳ‫؛‬ᅮˈ䌟ќ↣Ͼ‫׭‬䗝‫؛‬䆒ϔϾৢ偠ὖ⥛Ǆ
䋱৊ᮃᮍ⊩ৃ⫼Ѣ⹂ᅮ೼㒭ᅮ᭄᥂ᯊ᳔ৃ㛑ⱘ‫؛‬䆒üüᵕ໻ৢ偠ὖ⥛˄MAP˅‫؛‬
䆒Ǆᅗ↨݊Ҫⱘ‫؛‬䆒᳈ৃ㛑៤Ў᳔Ӭ‫؛‬䆒Ǆ
䋱৊ᮃ᳔Ӭߚ㉏఼ᇚ᠔᳝‫؛‬䆒ⱘ乘⌟㒧ড়䍋ᴹˈᑊҹ݊ৢ偠ὖ⥛Ўᴗ䞡ˈҹ䅵ㅫᇍ
ᮄᅲ՟ⱘ᳔ৃ㛑ߚ㉏Ǆ
ᴈ㋴䋱৊ᮃߚ㉏఼ᰃ೼䆌໮ᅲ䰙ᑨ⫼䯂乬Ёᕜ᳝⫼ⱘϔ⾡䋱৊ᮃᄺдᮍ⊩Ǆᅗ㹿⿄
Ўᴈ㋴ⱘ˄naive˅ᰃ಴Ў݊԰ⱘㅔ࣪‫؛‬ᅮ˖ሲᗻؐ೼㒭ᅮᅲ՟ⱘߚ㉏ᯊᴵӊ⣀ゟǄ
ᔧ䆹‫؛‬ᅮ៤ゟᯊˈᴈ㋴䋱৊ᮃߚ㉏఼ৃ䕧ߎ MAP ߚ㉏ǄेՓℸ‫؛‬ᅮϡ៤ゟˈ೼ᄺ
дߚ㉏᭛ᴀⱘᚙ‫މ‬ϟˈᴈ㋴䋱৊ᮃߚ㉏䗮ᐌгᰃᕜ᳝ᬜⱘǄ䋱৊ᮃ㕂ֵ㔥ᇍѢሲᗻ
ⱘᄤ䲚Ϟⱘϔ㒘ᴵӊ⣀ゟᗻ‫؛‬ᅮᦤկњ᳈ᔎⱘ㸼䖒㛑࡯Ǆ
䋱৊ᮃ᥼⧚Ḛᶊৃᇍ݊ҪϡⳈ᥹ᑨ⫼䋱৊ᮃ݀ᓣⱘᄺдᮍ⊩ⱘߚᵤᦤկ⧚䆎෎⸔Ǆ
՟བˈ೼⡍ᅮᴵӊϟᄺдϔϾᇍᑨѢᵕ໻Ԑ✊‫؛‬䆒ⱘᅲؐⳂᷛߑ᭄ᯊˈᅗৃՓ䇃Ꮒ
ᑇᮍ᳔ᇣ࣪Ǆ
᳔ᇣᦣ䗄䭓ᑺ‫߭ޚ‬ᓎ䆂䗝প䖭ḋⱘ‫؛‬䆒ˈᅗՓ‫؛‬䆒ⱘᦣ䗄䭓ᑺ੠㒭ᅮ‫؛‬䆒ϟ᭄᥂ⱘ
ᦣ䗄䭓ᑺⱘ੠᳔ᇣ࣪Ǆ䋱৊ᮃ݀ᓣ੠ֵᙃ䆎Ёⱘ෎ᴀ㒧䆎ৃᦤկℸ‫ⱘ߭ޚ‬ḍ᥂Ǆ
೼䆌໮ᅲ䰙ⱘᄺд䯂乬ЁˈᶤѯⳌ݇ⱘᅲ՟ব䞣ᰃϡৃ㾖ᆳࠄⱘǄEM ㅫ⊩ᦤկњ
ϔϾᕜ䗮⫼ⱘᮍ⊩ˈᔧᄬ೼䱤㮣ব䞣ᯊ䖯㸠ᄺдǄ䆹ㅫ⊩ᓔྟѢϔϾӏᛣⱘ߱ྟ‫؛‬
䆒Ǆ✊ৢ䗁ҷഄ䅵ㅫ䱤㮣ব䞣ⱘᳳᳯؐ˄‫؛‬ᅮᔧࠡ‫؛‬䆒ᰃℷ⹂ⱘ˅ˈ‫ݡ‬䞡ᮄ䅵ㅫᵕ
໻Ԑ✊‫؛‬䆒˄‫؛‬ᅮ䱤㮣ব䞣ㄝѢ㄀ 1 ℹЁᕫࠄⱘᳳᳯؐ˅Ǆ䖭ϔ䖛⿟ᬊᬯࠄϔϾሔ
䚼ⱘᵕ໻Ԑ✊‫؛‬䆒ˈҹঞ䱤㮣ব䞣ⱘԄ䅵ؐǄ
೼ὖ⥛੠㒳䅵ᮍ䴶᳝䆌໮ᕜདⱘҟ㒡ᗻ᭛ゴˈབ Casella & Berger(1990)Ǆ޴ᴀᖿ䗳খ㗗

㉏к㈡˄བ Maisel 1971; Speigel 1991)гᇍᴎ఼ᄺдⳌ݇ⱘὖ⥛੠㒳䅵⧚䆎ᦤկњӬ⾔ⱘ䯤
䗄Ǆ
ᇍ䋱৊ᮃߚ㉏఼੠᳔ᇣᑇᮍ䇃Ꮒߚ㉏఼ⱘ෎ᴀҟ㒡⬅ Duda & Hart(1973)㒭ߎˈDomigos

& Pazzani˄1996˅ߚᵤњ೼ᗢḋⱘᴵӊϟᴈ㋴䋱৊ᮃᮍ⊩ৃ䕧ߎ᳔Ӭⱘߚ㉏ˈेՓ݊⣀ゟ
ᗻ‫؛‬ᅮϡ៤ゟᯊ˄݇䬂೼Ѣ೼ᗢḋⱘᴵӊϟेՓⳌ݇㘨ⱘৢ偠ὖ⥛Ԅ䅵ϡℷ⹂гৃ䕧ߎ᳔Ӭ
ߚ㉏˅Ǆ
Cestnik(1990)䅼䆎њՓ⫼ m-Ԅ䅵ᴹԄ䅵ὖ⥛Ǆ
ᇚϡৠ䋱৊ᮃᮍ⊩Ϣ‫އ‬ㄪᷥㄝ݊Ҫㅫ⊩䖯㸠↨䕗ⱘᅲ偠㒧ᵰৃ೼ Michie et al.˄1994˅

ЁᡒࠄǄChauvin & Rumelhart(1995)ᦤկњ෎Ѣড৥Ӵ᪁ㅫ⊩ⱘ⼲㒣㔥㒰ⱘ䋱৊ᮃߚᵤǄ
ᇍ᳔ᇣᦣ䗄䭓ᑺ‫ⱘ߭ޚ‬䅼䆎ৃখ㗗 Rissanen(1983, 1989)ǄQuinlan & Rivest˄1989˅ᦣ

䗄њ݊Փ⫼ҹ䙓‫އܡ‬ㄪᷥⱘ䖛ᑺᢳড়Ǆ
д乬
6.1 ‫⃵ݡ‬㗗㰥 6.2.1 㡖Ёᑨ⫼䋱৊ᮃ㾘߭ⱘ՟ᄤǄ‫؛‬ᅮए⫳‫އ‬ᅮᇍ䆹⮙Ҏ‫خ‬㄀Ѡ⃵࣪偠⌟

䆩ˈ㗠Ϩ࣪偠㒧ᵰгЎℷǄḍ᥂䖭ϸ⃵⌟䆩ˈcancer ੠cancer ⱘৢ偠ὖ⥛ᰃ໮ᇥ˛‫؛‬ᅮϸ
Ͼ⌟䆩ᰃⳌѦ⣀ゟⱘǄ
6.2 ೼ 6.2.1 㡖ⱘ՟ᄤЁˈЎ䅵ㅫⰠ⮛ⱘৢ偠ὖ⥛ˈ䗮䖛ᇚ P(+|cancer)gP(cancer)੠

P(+|cancer)gP(cancer)ᔦϔ࣪ՓᅗӀⱘ੠Ў 1ǄՓ⫼䋱৊ᮃ݀ᓣ੠ܼὖ⥛݀ᓣ˄㾕㸼 6-1˅
䆕ᯢ䆹ᮍ⊩ᰃℷ⹂ⱘ˄े䖭ḋⱘᔦϔ࣪ৃҹᕫࠄ P(cancer|+)ⱘℷ⹂ؐ˅Ǆ
6.3 㗗㰥ϟ䴶ⱘὖ⥛ᄺдㅫ⊩ FindGˈᅗ䕧ߎϔϾᵕ໻ϔ㠀࣪ⱘϔ㟈‫؛‬䆒˄՟བˈবൟ

ぎ䯈ⱘᶤϾᵕ໻ϔ㠀៤ਬ˅Ǆ
(a)㒭ߎ P(h)੠ P(D|h)ⱘߚᏗˈҹՓ FindG ֱ䆕䕧ߎ MAP ‫؛‬䆒Ǆ
(b)㒭ߎ P(h)੠ P(D|h)ⱘߚᏗˈҹՓ FindG ϡ㛑ֱ䆕䕧ߎ MAP ‫؛‬䆒Ǆ
(c)㒭ߎ P(h)੠ P(D|h)ⱘߚᏗˈҹՓ FindG ֱ䆕䕧ߎ ML ‫؛‬䆒Ԛϡᰃ MAP ‫؛‬䆒Ǆ
6.4 ೼ 6.3 㡖ЁⱘὖᗉᄺдߚᵤЁˈ‫؛‬ᅮњᅲ՟ᑣ߫<x1…xm>ᰃ೎ᅮⱘǄ಴ℸˈ೼᥼ᇐ

P(D|h)㸼䖒ᓣᯊা䳔㗗㰥㗗ᆳࠄⳂᷛؐᑣ߫<d1…dm>ⱘὖ⥛Ǆ㗗㰥᳈ϔ㠀ⱘᚙ‫ेˈމ‬ᅲ՟乎
ᑣϡ೎ᅮˈԚᰃᅗӀᰃҢᅲ՟ぎ䯈XϞᅮНⱘᶤὖ⥛ߚᏗϞ⣀ゟᢑপⱘǄ᭄᥂D⦄೼ᖙ乏㹿
ᦣ䗄Ўϔ㒘ᑣ‫<{ي‬xi, di>}ˈ㗠P(D|h)ᖙ乏㛑ড᯴䘛ࠄ⡍ᅮᅲ՟xiⱘὖ⥛ˈҹঞⳂᷛؐdiⱘὖ⥛Ǆ
䆕ᯢ೼ℸϔ㠀ḚᶊЁᓣ 6.5 ҡ✊៤ゟǄᦤ⼎˖খ㗗 6.5 㡖ЁⱘߚᵤǄ
6.5 㗗㰥ᇚ᳔ᇣᦣ䗄䭓ᑺ‫߭ޚ‬ᑨ⫼ࠄϔϾ‫؛‬䆒ぎ䯈Hˈᅗࣙ৿㟇໮nϾᏗᇨሲᗻⱘড়প
˄བ˖SunnyġWarm˅Ǆ‫؛‬ᅮ↣Ͼ‫؛‬䆒ⱘ㓪ⷕЎㅔऩഄᇚ‫؛‬䆒Ёߎ⦄ⱘሲᗻ߫Вߎᴹˈ݊Ё
Ўњ㓪ⷕӏᛣϔϾnᏗᇨሲᗻ᠔䳔ԡ᭄Ўlog2nǄ䆒ᛇ㒭ᅮ‫؛‬䆒ϟḋ՟㓪ⷕᮍᓣЎ˖㢹ḋ՟Ϣ
‫؛‬䆒ϔ㟈㓪ⷕ䳔 0 ԡˈ৺߭⫼log2mԡ˄㸼⼎mϾḋ՟Ёાѯ㹿䇃ߚ㉏њüüℷ⹂ⱘߚ㉏ৃ
⬅䆹‫؛‬䆒乘⌟ⱘؐⱘ৺ᅮᕫࠄ˅Ǆ
(a)‫ߎݭ‬㽕㹿᳔ᇣ࣪ⱘ䞣ⱘ㸼䖒ᓣˈᣝ✻᳔ᇣᦣ䗄䭓ᑺ‫߭ޚ‬Ǆ
(b)ᰃ৺ৃ㛑ᓎゟϔ㒘䆁㒗᭄᥂ˈՓᄬ೼ϔϾϔ㟈‫؛‬䆒ˈԚ MDL 䗝ᢽњϔϾ䕗ϡϔ㟈ⱘ

‫؛‬䆒Ǆབᵰᰃ䖭ḋˈ㒭ߎ䖭ḋⱘ䆁㒗䲚˗৺߭㾷䞞ЎҔМǄ
(c)㒭ߎ P(h)੠ P(D|h)ⱘὖ⥛ߚᏗҹՓϞ䴶ⱘ MDL ㅫ⊩䕧ߎ MAP ‫؛‬䆒Ǆ
6.6 㗗㰥㡖 6.9.1 Ё PlayTennis 䯂乬ⱘᴈ㋴䋱৊ᮃߚ㉏఼ˈ⫼䋱৊ᮃ㕂ֵ㔥⬏ߎ݊ЁՓ⫼

ⱘᴵӊ⣀ゟᗻ‫؛‬ᅮǄ㒭ߎϢ㒧⚍ Wind Ⳍ݇㘨ⱘᴵӊὖ⥛㸼Ǆ
ㅢ7ㄖ䇗㇍ᆜҖ⨼䇰
ᵜㄐ⨶䇪ൠ࡫⭫Ҷ㤕ᒢ㊫රⲴᵪಘᆖҐ䰞仈ѝⲴഠ䳮ˈ઼㤕ᒢ㊫රⲴᵪಘᆖҐ㇇⌅Ⲵ㜭
࣋Ǆ䈕⨶䇪㠤࣋ҾഎㆄྲлⲴ䰞仈˖Ā൘ӰѸṧⲴᶑԦлᡀ࣏ⲴᆖҐᱟਟ㜭Ⲵ˛āԕ৺Ā൘
ӰѸᶑԦла⢩ᇊⲴᆖҐ㇇⌅ਟ‫؍‬䇱ᡀ࣏䘀㹼˛āѪҶ࠶᷀ᆖҐ㇇⌅ˈ䘉䟼㘳㲁Ҷє⿽ṶᷦǄ
൘ਟ㜭䘁լ↓⺞˄PAC˅ṶᷦлˈᡁԜ⺞ᇊҶ㤕ᒢ‫ٷ‬䇮㊫࡛ˈࡔᯝᆳԜ㜭੖Ӿཊ亩ᔿᮠ䟿Ⲵ
䇝㓳ṧֻѝᆖҐᗇࡠ˗ᡁԜ䘈ᇊѹҶањሩ‫ٷ‬䇮オ䰤Ⲵ㠚❦ᓖ䟿ˈ⭡ᆳਟԕ⭼ᇊᖂ㓣ᆖҐᡰ
䴰Ⲵ䇝㓳ṧֻᮠⴞǄ൘ࠪ䭉⭼䲀˄Mistake bound˅ṶᷦлˈᡁԜ㘳ḕҶањᆖҐಘ൘⺞ᇊ↓
⺞‫ٷ‬䇮ࡽਟ㜭ӗ⭏Ⲵ䇝㓳䭉䈟ᮠ䟿Ǆ
7.1 ԁ㔃
൘⹄ウᵪಘᆖҐ䗷〻ѝˈᖸ㠚❦ൠᜣ⸕䚃ᆖҐಘ˄ᵪಘⲴᡆ䶎ᵪಘⲴ˅ᓄ䚥ᗚӰѸṧⲴ
㿴ࡉǄᱟ੖ਟ㜭⤜・ҾᆖҐ㇇⌅⺞ᇊᆖҐ䰞仈ѝപᴹⲴ䳮ᓖ˛㜭੖⸕䚃Ѫ‫؍‬䇱ᡀ࣏ⲴᆖҐᴹ
ཊቁ䇝㓳ᱟᗵ㾱Ⲵᡆ‫ݵ‬䏣Ⲵ˛ྲ᷌ᆖҐಘ㻛‫ݱ‬䇨ੁᯭᮉ㘵ᨀࠪḕ䈒ˈ㘼нᱟ㿲ሏ䇝㓳䳶Ⲵ䲿
ᵪṧᵜˈՊሩᡰ䴰ṧֻᮠⴞᴹᘾṧⲴᖡ૽˛㜭੖࡫⭫ࠪᆖҐಘ൘ᆖࡠⴞḷ࠭ᮠࡽՊᴹཊቁ⅑
ࠪ䭉˛㜭੖࡫⭫ࠪа㊫ᆖҐ䰞仈ѝപᴹⲴ䇑㇇༽ᵲᓖ˛
㲭❦ሩᡰᴹ䘉Ӌ䰞仈Ⲵа㡜എㆄ䘈ᵚ⸕ˈնᱟнᆼᮤⲴᆖҐ䇑㇇⨶䇪ᐢ㓿ᔰ࿻ࠪ⧠Ǆᵜ
ㄐ䱀䘠Ҷ䈕⨶䇪ѝⲴаӋ‫ޣ‬䭞㔃䇪ˈᒦᨀ‫׋‬Ҷ൘⢩ᇊ䰞仈лаӋ䰞仈ⲴㆄṸǄ䘉䟼ᡁԜ⵰䟽
䇘䇪ਚ㔉ᇊⴞḷ࠭ᮠⲴ䇝㓳ṧ઼ֻ‫ى‬䘹‫ٷ‬䇮オ䰤ⲴᶑԦлˈሩ䈕ᵚ⸕Ⲵⴞḷ࠭ᮠⲴᖂ㓣ᆖҐ
䰞仈Ǆ൘䘉ṧⲴṶᷦлˈѫ㾱㾱䀓ߣⲴ䰞仈ྲ˖䴰㾱ཊቁ䇝㓳ṧֻ᡽䏣ԕᡀ࣏ൠᆖҐࡠⴞḷ
࠭ᮠˈԕ৺ᆖҐಘ൘䗮ࡠⴞḷࡽՊᴹཊቁ⅑ࠪ䭉Ǆྲਾ䶒ሶⴻࡠˈᴹਟ㜭ሩ䘉Ӌ䰞仈ᨀࠪᇊ
䟿Ⲵкл⭼ˈ䘉สҾᆖҐ䰞仈Ⲵྲл኎ᙗǄ
x ᆖҐಘᡰ㘳㲁Ⲵ‫ٷ‬䇮オ䰤Ⲵབྷሿ઼༽ᵲᓖ
x ⴞḷᾲᘥ享䘁լࡠᘾṧⲴ㋮ᓖ
x ᆖҐಘ䗃ࠪᡀ࣏Ⲵ‫ٷ‬䇮Ⲵਟ㜭ᙗ
x 䇝㓳ṧֻᨀ‫׋‬㔉ᆖҐಘⲴᯩᔿ
ᵜㄐⲴབྷ䜘࠶ሶнՊ⵰䟽Ҿঅ⤜ⲴᆖҐ㇇⌅ˈ㘼ᱟ൘䖳ᇭᒯⲴᆖҐ㇇⌅㊫࡛ѝ࡫⭫ᡰ㘳
㲁Ⲵ‫ٷ‬䇮オ䰤ˈԕ৺䇝㓳ṧֻⲴᨀ‫׋‬ᯩᔿㅹǄᡁԜⲴⴞḷᱟѪҶഎㆄԕлⲴ䰞仈˖
x ṧᵜ༽ᵲᓖ˄Sample complexity˅ǄᆖҐಘ㾱᭦ᮋࡠᡀ࣏‫ٷ‬䇮˄ԕ䖳儈Ⲵᾲ⦷˅ˈ
䴰㾱ཊቁ䇝㓳ṧֻ˛
x 䇑㇇༽ᵲᓖ(Computational complexity)ǄᆖҐಘ㾱᭦ᮋࡠᡀ࣏‫ٷ‬䇮˄ԕ䖳儈Ⲵᾲ
⦷˅䴰㾱ཊབྷⲴ䇑㇇䟿˛
x ࠪ䭉⭼䲀˄Mistake bound˅Ǆ൘ᡀ࣏᭦ᮋࡠањ‫ٷ‬䇮ࡽˈᆖҐಘሩ䇝㓳ṧֻⲴ
䈟࠶㊫ᴹཊቁ⅑˛
⌘᜿ѪҶ䀓ߣ䘉Ӌ䰞仈䴰㾱䇨ཊ⢩↺ⲴᶑԦ䇮ᇊǄֻྲˈᴹ䇨ཊᯩ⌅ᶕᤷᇊሩҾᆖҐಘ
ӰѸᱟĀᡀ࣏ⲴāǄа⿽ਟ㜭Ⲵࡔᯝᯩ⌅ᱟ˖ᆖҐಘᱟ੖䗃ࠪㅹҾⴞḷᾲᘥⲴ‫ٷ‬䇮Ǆਖа⿽
ᯩ⌅ᱟਚ㾱≲䗃ࠪⲴ‫ٷ‬䇮оⴞḷᾲᘥ൘ཊᮠᰦ䰤޵᜿㿱а㠤ˈᡆᱟᆖҐಘ䙊ᑨՊ䗃ࠪ䘉ṧⲴ
‫ٷ‬䇮Ǆ⴨լൠˈ䘈ᗵ享ᤷᇊᆖҐಘᱟྲօ㧧ᗇ䇝㓳ṧֻⲴǄਟԕᤷᇊ䇝㓳ṧֻ⭡ањᯭᮉ㘵
㔉ࠪˈᡆ⭡ᆖҐಘ㠚ᐡᇎ傼ᶕ㧧ᗇˈᡆ᤹➗Ḁ䗷〻䲿ᵪൠ⭏ᡀ㘼нਇᆖҐಘⲴ᧗ࡦǄਟԕ亴
ᯉˈሩк䘠䰞仈Ⲵഎㆄ‫׍‬䎆ҾᡁԜᡰ㘳㲁Ⲵ⢩ᇊṶᷦᡆᆖҐ⁑රǄ
ᵜㄐⲴਾ㔝ྲлᆹᧂǄ7.2 㢲ӻ㓽ਟ㜭䘁լ↓⺞˄PAC˅ᆖҐṶᷦǄ7.3 㢲൘↔ PAC Ṷ

ᷦл࠶᷀Ҷࠐ⿽ᆖҐ㇇⌅Ⲵṧᵜ༽ᵲᓖ઼䇑㇇༽ᵲᓖǄ7.4 㢲ӻ㓽Ҷ‫ٷ‬䇮オ䰤༽ᵲᓖⲴањ
䟽㾱ᓖ䟿ḷ߶ˈ〠Ѫ VC-㔤ˈᒦфሶ PAC ࠶᷀ᢙኅࡠ‫ٷ‬䇮オ䰤ᰐ䲀ⲴᛵߥǄ7.5 㢲ӻ㓽Ҷࠪ
䭉⭼䲀⁑රˈᒦᨀ‫׋‬Ҷࡽ䶒ㄐ㢲ѝࠐњᆖҐ㇇⌅ࠪ䭉ᮠ䟿Ⲵ⭼䲀Ǆᴰਾˈӻ㓽Ҷ࣐ᵳཊᮠ㇇
⌅ˈᆳᱟањ㔃ਸཊњᆖҐ㇇⌅ᶕӗ⭏ਸᒦⲴ亴⍻Ⲵᇎ⭘㇇⌅ˈ䘈ӻ㓽Ҷ䈕㇇⌅Ⲵ⨶䇪ࠪ䭉
⭼䲀Ǆ
7.2 ਥ㜳ᆜҖ䘇ղ↙⺤‫ٽ‬䇴
ᵜ㢲ᡁԜ㘳㲁ᆖҐ䰞仈Ⲵа⿽⢩↺Ṷᷦˈ〠Ѫਟ㜭䘁լ↓⺞ (probably approximately

correct, PAC)ᆖҐ⁑රǄ俆‫ݸ‬ᡁԜᤷᇊ PAC ᆖҐ⁑ර䘲⭘Ⲵ䰞仈ˈ❦ਾ࠶᷀൘↔ PAC ⁑ර
лᆖҐн਼㊫࡛Ⲵⴞḷ࠭ᮠ䴰㾱ཊቁ䇝㓳ṧ઼ֻཊབྷⲴ䇑㇇䟿ǄѪㆰ᰾䎧㿱ˈ䘉䟼Ⲵ䇘䇪ሶ
䲀ࡦ൘ᆖҐᐳቄ٬ᾲᘥˈф䇝㓳ᮠᦞᱟᰐಚ༠ⲴǄ❦㘼ˈ䇨ཊ㔃䇪ਟᢙኅࡠᴤа㡜Ⲵᛵᖒˈ
ྲᆖҐᇎ٬ⴞḷ࠭ᮠ˄∄ྲ Natarajan 1991 ˅ˈᡆӾḀ⿽㊫රⲴᴹಚ༠ᮠᦞѝ䘋㹼ᆖҐ˄ֻ
ྲˈ㿱 Laird 1988; Kearns & Vazirani 1994˅Ǆ
7.2.1 䰤从Ṽᷬ
ྲࡽ䶒Ⲵㄐ㢲ˈԔ X ԓ㺘ᡰᴹᇎֻⲴ䳶ਸˈⴞḷ࠭ᮠ൘ަкᇊѹǄֻྲˈX ਟ㺘⽪ᡰᴹ

ӪⲴ䳶ਸˈ⇿њӪ᧿䘠Ѫ኎ᙗ age˄young ᡆ old˅઼ height (short ᡆ long)ǄԔ C ԓ㺘ᆖҐ
ಘ㾱ᆖҐⲴⴞḷᾲᘥ䳶ਸǄC ѝ⇿њⴞḷᾲᘥ c ሩᓄҾ X ⲴḀњᆀ䳶ˈᡆањㅹ᭸Ⲵᐳቄ࠭
ᮠ c˖Xė{0,1}ǄֻྲˈC ѝањⴞḷ࠭ᮠ c Ѫᾲᘥ˖Āᱟ━䴚㘵ⲴӪāǄ㤕 x ᱟ c Ⲵ↓ֻˈ
ࡉ c(x)=1˗㤕 x Ѫ৽ֻˈࡉ c(x)=0Ǆ
‫ٷ‬ᇊᇎֻ᤹➗Ḁᾲ⦷࠶ᐳ D Ӿ X ѝ䲿ᵪӗ⭏Ǆֻྲ D ਟѪӾḀփ㛢⭘૱୶ᓇ䎠ࠪᶕⲴ

Ӫ䘉ṧањᇎֻ࠶ᐳǄа㡜ൠˈD ਟѪԫօ࠶ᐳˈ㘼фᆳሩᆖҐಘᱟᵚ⸕ⲴǄሩҾ D ᡰ㾱≲
ⲴᱟᆳⲴっᇊᙗˈণ䈕࠶ᐳнՊ䲿ᰦ䰤ਈॆǄ䇝㓳ṧֻⲴ⭏ᡀ᤹➗ D ࠶ᐳ䲿ᵪᣭਆᇎֻ xˈ
❦ਾ x ৺ަⴞḷ٬ c(x)㻛ᨀ‫׋‬㔉ᆖҐಘǄ
ᆖҐಘ L ൘ᆖҐⴞḷᾲᘥᰦ㘳㲁ਟ㜭‫ٷ‬䇮Ⲵ䳶ਸ HǄֻྲˈH ਟѪᡰᴹ㜭⭡኎ᙗ age ઼

height Ⲵਸਆ㺘⽪Ⲵ‫ٷ‬䇮䳶ਸǄ൘㿲ሏࡠҶа㌫ࡇ‫ޣ‬Ҿⴞḷᾲᘥ c ⲴⲴ䇝㓳ṧֻਾˈL ᗵ享
Ӿ H ѝ䗃ࠪḀ‫ٷ‬䇮 hˈᆳᱟሩ c Ⲵՠ䇑ǄѪ‫ޜ‬ᒣ䎧㿱ˈᡁԜ䙊䗷 h ൘Ӿ X ѝᣭਆⲴᯠᇎֻк
Ⲵᙗ㜭ᶕ䇴ՠ L ᱟ੖ᡀ࣏Ǆᣭਆ䗷〻᤹➗࠶ᐳ Dˈণоӗ⭏䇝㓳ᮠᦞ⴨਼Ⲵᾲ⦷࠶ᐳǄ
൘↔ṶᷦлˈᡁԜᝏ‫ޤ‬䏓Ⲵᱟ࡫⭫н਼ᆖҐಘ L Ⲵᙗ㜭ˈ䘉ӋᆖҐಘ֯⭘н਼‫ٷ‬䇮オ
䰤 HˈᒦᆖҐн਼㊫࡛Ⲵ C ѝⲴⴞḷᾲᘥǄ⭡ҾᡁԜ㾱≲ L 䏣ཏа㡜ˈԕӾ C ѝᆖࡠԫօ
ⴞḷᾲᘥˈᡰԕн䇪䇝㓳ṧֻⲴ࠶ᐳྲօˈᡁԜ㓿ᑨՊሩ C ѝᡰᴹਟ㜭Ⲵⴞḷᾲᘥ઼ᡰᴹ
ਟ㜭Ⲵᇎֻ࠶ᐳ D 䘋㹼ᴰᐞᛵߥⲴ࠶᷀Ǆ
7.2.2 ‫ٽ‬䇴Ⲻ䭏䈥⦽
ѪҶ᧿䘠ᆖҐಘ䗃ࠪⲴ‫ٷ‬䇮 h ሩⵏᇎⴞḷᾲᘥⲴ䙬䘁〻ᓖˈ俆‫ݸ‬㾱ᇊѹ‫ٷ‬䇮 h ሩᓄҾⴞ

ḷᾲᘥ c ઼ᇎֻ࠶ᐳ D Ⲵⵏᇎ䭉䈟⦷˄true error˅Ǆ䶎ᖒᔿⲴ᧿䘠ᱟ˖h Ⲵⵏᇎ䭉䈟⦷Ѫᓄ⭘
h ࡠሶᶕ᤹࠶ᐳ D ᣭਆⲴᇎֻᰦⲴᵏᵋⲴ䭉䈟⦷Ǆ ᇎ䱵кㅜ 5 ㄐᐢ㓿ᇊѹҶ h Ⲵⵏᇎ䭉䈟⦷Ǆ
Ѫᯩ‫ׯ‬䎧㿱ˈ䘉䟼䟽䘠ал䈕ᇊѹˈ֯⭘ c 㺘⽪ᐳቄⴞḷ࠭ᮠǄ
定义： ‫ٷ‬䇮 h ‫ޣ‬Ҿⴞḷᾲᘥ c ઼࠶ᐳ D Ⲵⵏ

ⵏᇎ䭉䈟⦷(true error)Ѫ h 䈟࠶㊫᤹➗ D 䲿
ᵪᣭਆⲴᇎֻⲴᾲ⦷Ǆ
errorD (h) { Pr [c( x) z h( x)]

xD
䘉䟼ㅖਧ Pr ԓ㺘൘ᇎֻ࠶ᐳ D к䇑㇇ᾲ⦷Ǆ

xD
മ 7-1 മ⽪Ҷ䈕䭉䈟⦷ⲴᇊѹǄᾲᘥ c ઼ h 㻛㺘⽪Ѫ X ѝḷѪ↓ֻⲴᇎֻ䳶ਸǄh ሩᓄ

Ҿ c Ⲵ䭉䈟⦷Ѫˈ䲿ᵪ䘹ਆⲴᇎֻ㩭‫ ޕ‬h ઼ c на㠤४䰤˄ণᆳԜⲴ䳶ਸᐞ˅Ⲵᾲ⦷Ǆ⌘᜿ˈ
䭉䈟⦷ᇊѹ൘ᮤњᇎֻ࠶ᐳѻкˈ㘼нਚᱟ䇝㓳ṧֻѻкˈഐѪᆳᱟ൘ᇎ䱵ᓄ⭘↔‫ٷ‬䇮 h
ࡠਾ㔝ᇎֻкᰦՊ䙷ࡠⲴⵏᇎ䭉䈟⦷Ǆ
Instance space: ᇎֻオ䰤
Where c and h disagree: c ઼ h на㠤Ⲵ४䰤
മ 7-1 ‫ޣ‬Ҿⴞḷᾲᘥ c ‫ٷ‬䇮 h Ⲵ䭉䈟⦷
‫ޣ‬Ҿ c Ⲵ h Ⲵ䭉䈟⦷Ѫањ䲿ᵪᣭਆⲴᇎֻ㩭‫ ޕ‬h ઼ c ሩᆳⲴ࠶㊫на㠤Ⲵ४䰤Ⲵᾲ⦷Ǆ+઼

-⛩㺘⽪↓৽䇝㓳ֻǄ⌘᜿ h ‫ޣ‬Ҿ c ᴹањ䶎䴦Ⲵ䭉䈟⦷ˈቭ㇑䗴ӺѪ→ h ઼ c ൘ᡰᴹ 5 њ䇝㓳ṧ
ֻк䜭а㠤Ǆ
⌘᜿ˈ↔䭉䈟⦷ᕪ⛸ൠ‫׍‬䎆Ҿᵚ⸕Ⲵᾲ⦷࠶ᐳ DǄֻྲˈྲ᷌ D ᱟањ൷रⲴᾲ⦷࠶ᐳˈ

ᆳሩ X ѝ⇿њᇎֻ䜭䍻Ҹ⴨਼Ⲵᾲ⦷ˈ䛓Ѹമ 7-1 ѝ‫ٷ‬䇮Ⲵ䭉䈟⦷ሶѪ h ઼ c на㠤Ⲵオ䰤
൘‫ޘ‬䜘ᇎֻオ䰤ѝⲴ∄ֻǄ❦㘼ˈྲ᷌ D ᚠྭᢺ h ઼ c на㠤४䰤ѝⲴᇎֻ䍻ҸҶᖸ儈Ⲵᾲ
⦷ˈ⴨਼Ⲵ h ઼ c ሶ䙐ᡀᴤ儈Ⲵ䭉䈟⦷Ǆᶱㄟᛵߥл㤕 D ሩ┑䏣 h(x)=c(x)Ⲵᡰᴹᇎֻ䍻Ҹ䴦
ᾲ⦷ˈമ 7-1 ѝ h Ⲵ䭉䈟⦷ሶѪ 1ˈ㘼н䇪 h ઼ c ൘ཊቁᇎֻк࠶㊫а㠤Ǆ
ᴰਾˈ⌘᜿hሩᓄҾcⲴ䭉䈟⦷н㜭ⴤ᧕⭡ᆖҐಘ㿲ሏࡠǄLਚ㜭㿲ሏࡠ൘䇝㓳ṧֻкh
Ⲵᙗ㜭ˈᆳҏਚ㜭൘↔ส⹰к䘹ᤙަ‫ٷ‬䇮䗃ࠪǄᡁԜሶ֯⭘ᵟ䈝䇝㓳䭉䈟⦷(training error)
ᶕᤷԓ䇝㓳ṧֻѝ㻛h䈟࠶㊫Ⲵṧֻᡰঐ∄ֻˈԕ४࠶к䶒ᇊѹⲴⵏᇎ䭉䈟⦷Ǆ䘉䟼‫ޣ‬Ҿᆖ
Ґ༽ᵲᓖⲴ࠶᷀ཊᮠത㔅⵰䘉ṧⲴ䰞仈˖Ā hⲴ㿲ሏࡠⲴ䇝㓳䭉䈟⦷ሩⵏᇎ䭉䈟⦷errorD(h)
ӗ⭏н↓⺞ՠ䇑Ⲵਟ㜭ᙗᴹཊབྷ˛ā
⌘᜿↔䰞仈оㅜ 5 ㄐ㘳㲁Ⲵ䰞仈ѻ䰤Ⲵᇶ࠷㚄㌫Ǆഎᗶ൘ㅜ 5 ㄐѝᇊѹҶ h ‫ޣ‬Ҿṧֻ䳶
ਸ S Ⲵṧᵜ䭉䈟⦷ (sample error)ˈѪṧֻ䳶ਸ S ѝ㻛 h 䈟࠶㊫Ⲵṧֻᡰঐ∄ֻǄк䶒ᇊѹ
Ⲵ䇝㓳䭉䈟⦷ቡᱟᖃ S Ѫ䇝㓳ṧֻ䳶ਸᰦⲴṧᵜ䭉䈟⦷Ǆ൘ㅜ 5 ㄐѝˈᡁԜ൘ᮠᦞṧᵜ S
⤜・Ҿ h ᣭਆⲴࡽᨀлˈ⺞ᇊṧᵜ䭉䈟⦷ሩՠ䇑ⵏᇎ䭉䈟⦷ӗ⭏䈟ሬⲴᾲ⦷Ǆ❦㘼ᖃ S ᱟ䇝
㓳ᮠᦞ䳶ਸᰦˈᆖࡠⲴ‫ٷ‬䇮䶎ᑨ‫׍‬䎆Ҿ SǄഐ↔ˈᵜㄐሶ㔉ࠪ䘉а䟽㾱Ⲵ⢩↺ᛵᖒлⲴ࠶᷀Ǆ
7.2.3 PAC ਥᆜҖᙝ
ᡁԜⲴⴞḷᱟ࡫⭫ࠪ䘉ṧⲴⴞḷᾲᘥˈᆳԜ㜭ཏӾਸ⨶ᮠ䟿Ⲵ䲿ᵪᣭਆ䇝㓳ṧֻѝ䙊䗷
ਸ⨶Ⲵ䇑㇇䟿ਟ䶐ൠᆖҐࡠǄ
ሩҾਟᆖҐᙗᘾṧ䘋㹼㺘䘠˛а⿽ਟ㜭Ⲵ䘹ᤙᱟ᧿䘠ѪҶᆖҐࡠањ֯errorD(h)=0 Ⲵ
‫ٷ‬䇮hˈᡰ䴰Ⲵ䇝㓳ṧֻᮠǄнᒨⲴᱟˈ䘉ṧⲴ䘹ᤙᱟнਟ㹼Ⲵˈ৏ഐᴹєњ˖俆‫ˈݸ‬䲔䶎
ሩXѝ⇿њਟ㜭Ⲵᇎֻ䜭ᨀ‫׋‬䇝㓳ṧֻ˄ањнᇎ䱵Ⲵ‫ٷ‬ᇊ˅ˈ੖ࡉՊᴹཊњ‫ٷ‬䇮о䇝㓳ṧ
ֻа㠤ˈ㘼фᆖҐಘᰐ⌅‫؍‬䇱䘹ᤙࡠⴞḷᾲᘥǄަ⅑ˈ⭡Ҿ䇝㓳ṧֻᱟ䲿ᵪᣭਆⲴˈᙫᴹа
њ䶎 0 Ⲵᾲ⦷֯ᗇᆖҐಘ䶒ѤⲴ䇝㓳ṧֻᴹ䈟ሬᙗǄ˄ֻྲˈ㲭❦ᡁԜ㓿ᑨਟ㿱ࡠн਼䓛儈
Ⲵ━䴚㘵ˈն൘Ḁаཙѝᙫᆈ൘䘉ṧⲴᵪՊˈᡰᴹ䇝㓳ṧֻ䜭ࡊྭᱟ 2 ㊣儈Ǆ˅
Ѫ䀓ߣ䘉єњഠ䳮ˈᡁԜ⭘є⿽ᯩ⌅ᕡॆҶሩᆖҐಘⲴ㾱≲Ǆ俆‫ˈݸ‬ᡁԜн㾱≲ᆖҐಘ
䗃ࠪ䴦䭉䈟⦷‫ٷ‬䇮ˈ㘼ਚ㾱≲ަ䭉䈟⦷㻛䲀ᇊ൘Ḁᑨᮠ¦Ⲵ㤳ത޵ˈ¦ਟѪԫ᜿ሿǄㅜҼˈ
н޽㾱≲ᆖҐಘሩᡰᴹⲴ䲿ᵪᣭਆṧֻᒿࡇ䜭㜭ᡀ࣏ˈਚ㾱≲ަཡ䍕Ⲵᾲ⦷㻛䲀ᇊ൘Ḁњᑨ
ᮠ¥Ⲵ㤳ത޵ˈ¥ҏਟਆԫ᜿ሿǄㆰ㘼䀰ѻˈᡁԜਚ㾱≲ᆖҐಘਟ㜭ᆖҐࡠањ䘁լ↓⺞Ⲵ
‫ٷ‬䇮ˈഐ↔ᗇࡠҶ䈕ᵟ䈝Āਟ㜭䘁լ↓⺞ᆖҐāˈᡆ PAC ᆖҐǄ
㘳㲁Ḁаⴞḷᾲᘥ㊫࡛C઼֯⭘‫ٷ‬䇮オ䰤HⲴᆖҐಘLǄ䶎ᖒᔿൠˈሩCѝԫ᜿ⴞḷᾲᘥ
cˈ㤕൘㿲ሏࡠਸ⨶ᮠⴞⲴ䇝㓳ṧֻᒦᢗ㹼Ҷਸ⨶Ⲵ䇑㇇䟿ਾˈLԕᾲ⦷(1- ¥ )䗃ࠪањ
errorD(h)<¦Ⲵ‫ٷ‬䇮hˈࡉᡁԜ〠ᾲᘥ㊫࡛Cᱟ֯⭘HⲴLਟPACᆖҐⲴǄᴤ㋮⺞Ⲵᇊѹྲл˖
定义：㘳㲁аᾲᘥ㊫࡛Cᇊѹ൘䮯ᓖѪnⲴᇎֻ䳶ਸXкˈᆖҐಘL֯⭘‫ٷ‬䇮オ䰤HǄ
ᖃሩᡰᴹcęCˈXкⲴ࠶ᐳDˈ¦┑䏣 0<¦<1/2ˈԕ৺¥┑䏣 0<¥<1/2ˈᆖҐಘL
ሶԕ㠣ቁ 1-¥Ⲵᾲ⦷䗃ࠪа‫ٷ‬䇮hęHˈ֯errorD(h)İ¦ˈ䘉ᰦ〠Cᱟ֯⭘HⲴLਟPAC
ᆖҐⲴǄᡰ֯⭘Ⲵᰦ䰤Ѫ 1/¦ˈ1/¥ˈnԕ৺size(c)Ⲵཊ亩ᔿ࠭ᮠǄ
䘉䟼Ⲵᇊѹ㾱≲ L ┑䏣єњᶑԦǄ俆‫ˈݸ‬L ᗵ享ԕԫ᜿儈ᾲ⦷(1-¥)䗃ࠪањ䭉䈟⦷ԫ

᜿վ˄¦˅Ⲵ‫ٷ‬䇮ǄㅜҼˈᆖҐ䗷〻ᗵ享ᱟ儈᭸Ⲵˈަᰦ䰤ᴰཊԕཊ亩ᔿᯩᔿ໎䮯ˈཊ亩ᔿ
ѝ 1/¦઼ 1/¥ᇊѹҶሩ䗃ࠪ‫ٷ‬䇮㾱≲Ⲵᕪᓖˈn ઼ size(c)ࡉᇊѹҶᇎֻオ䰤 X ઼ᾲᘥ㊫ C ѝ
പᴹⲴ༽ᵲᓖǄ䘉䟼ˈn Ѫ X ѝᇎֻⲴ䮯ᓖǄֻྲˈྲ᷌ᇎֻѪ k њᐳቄ٬Ⲵਸਆˈ䛓Ѹ
n=kǄsize(c)Ѫ‫ٷ‬ᇊሩ C 䟷⭘Ḁ⿽㺘⽪ᯩ⌅ᰦˈަѝⲴᾲᘥ c Ⲵ㕆⸱䮯ᓖǄֻྲˈ㤕 C ѝⲴ
ᾲᘥѪ㠣ཊ k њᐳቄ⢩ᖱⲴਸਆˈ⇿њᾲᘥ䙊䗷ࡇࠪਸਆᔿѝⲴ⢩ᖱⲴ㍒ᕅᶕ᧿䘠ˈ䛓Ѹ
size(c)Ѫᇎ䱵⭘ᶕ᧿䘠 c Ⲵᐳቄ⢩ᖱᮠ䟿Ǆ
䘉䟼ሩ PAC ᆖҐⲴᇊѹᔰ࿻ⴻᶕਚ‫ޣ‬ᗳᆖҐᡰ䴰Ⲵ䇑㇇䍴Ⓚˈ㘼൘ᇎ䐥ѝˈ䙊ᑨᴤ‫ޣ‬

ᗳᡰ䴰Ⲵ䇝㓳ṧֻᮠǄ❦㘼䘉є㘵ᱟ㍗ᇶ⴨‫ྲ˖Ⲵޣ‬᷌ L ሩ⇿њ䇝㓳ṧֻ䴰㾱Ḁᴰሿ༴⨶
ᰦ䰤ˈ䛓ѸѪҶ֯ c ᱟ L ਟ PAC ᆖҐⲴˈL ᗵ享Ӿཊ亩ᔿᮠ䟿Ⲵ䇝㓳ṧֻѝ䘋㹼ᆖҐǄᇎ
䱵кˈѪᱮ⽪Ḁⴞḷᾲᘥ㊫࡛ C ᱟਟ PAC ᆖҐⲴˈањިරⲴ䙄ᖴᱟ䇱᰾ C ѝ⇿њⴞḷᾲ
ᘥਟԕӾཊ亩ᔿᮠ䟿Ⲵ䇝㓳ṧֻѝᆖҐࡠˈ㘼ਾ䇱᰾⇿ṧֻ༴⨶ᰦ䰤ҏ䲀Ҿཊ亩ᔿ㓗Ǆ
൘㔗㔝䇘䇪ԕࡽˈᗵ享ᤷࠪ䳀ਜ਼൘ PAC ਟᆖҐᙗᇊѹѝⲴањѕṬⲴ䲀ࡦǄ䈕ᇊѹ䳀
ਜ਼‫ٷ‬ᇊҶᆖҐಘⲴ‫ٷ‬䇮オ䰤 H वਜ਼ањ‫ٷ‬䇮ˈᆳо C ѝ⇿њⴞḷᾲᘥਟᴹԫ᜿ሿⲴ䈟ᐞǄ
䘉а⛩ᶕⓀҾк䶒ᇊѹѝ㾱≲ᆖҐಘ䈟ᐞ⭼䲀¦ԫ᜿᧕䘁Ҿ 0 ᰦҏ㜭ᡀ࣏䘀㹼Ǆᖃ❦ˈྲ᷌
亴‫ݸ‬н⸕䚃 C ሶᖸ䳮‫؍‬䇱䘉а⛩˄ሩҾањӾമ‫ۿ‬ѝ䇶࡛ࠪӪ㝨Ⲵ〻ᒿᶕ䈤ˈC ᱟӰѸ˛˅ˈ
䲔䶎 H ਆѪ X Ⲵᑲ䳶Ǆྲㅜ 2 ㄐᤷࠪⲴˈ䘉ṧањᰐ‫ Ⲵٿ‬H ሶнՊӾਸ⨶ᮠ䟿Ⲵ䇝㓳ṧֻ
ѝ⌋ॆǄн䗷ˈสҾ PAC ᆖҐ⁑රⲴ㔃䇪ˈሩҾ亶Պн਼ᆖҐ䰞仈Ⲵ⴨ሩ༽ᵲᓖԕ৺⌋ॆ
㋮ᓖ䲿⵰䇝㓳ṧֻ㘼ᨀ儈Ⲵ∄⦷ॱ࠶ᴹ⳺Ǆᴤ䘋а↕ˈ7.3.1 㢲ѝሶ䀓䲔䘉аѕṬ‫ٷ‬ᇊˈԕ
㘳㲁ᆖҐಘн亴‫ٷݸ‬ᇊⴞḷᾲᘥᖒᔿⲴᛵߥǄ
7.3 ᴿ䲆‫ٽ‬䇴グ䰪Ⲻṭᵢགྷᵸᓜ
ྲкᡰ䘠ˈPAC ਟᆖҐᙗᖸབྷ〻ᓖк⭡ᡰ䴰Ⲵ䇝㓳ṧֻᮠ⺞ᇊǄ䲿⵰䰞仈㿴⁑Ⲵ໎䮯
ᡰᑖᶕⲴᡰ䴰䇝㓳ṧֻⲴ໎䮯〠Ѫ䈕ᆖҐ䰞仈Ⲵṧᵜ༽ᵲᓖ(sample complexity)ˈᆳᱟ䙊ᑨ
ᴰᝏ‫ޤ‬䏓Ⲵ⢩ᙗǄ৏ഐ൘Ҿˈ൘ཊᮠᇎ䱵䰞仈ѝˈᴰ䲀ࡦᆖҐಘᡀ࣏Ⲵഐ㍐ᱟᴹ䲀Ⲵਟ⭘䇝
㓳ᮠᦞǄ
䘉䟼ሶṧᵜ༽ᵲᓖⲴ䇘䇪䲀ᇊҾа㊫䶎ᑨᒯ⌋ⲴᆖҐಘˈ〠Ѫа㠤ᆖҐಘ(consistent
learner)ǄањᆖҐಘᱟа㠤Ⲵ(consistent)ˈᖃᆳਚ㾱൘ਟ㜭ᰦ䜭䗃ࠪ㜭ᆼ㖾ᤏਸ䇝㓳ᮠᦞⲴ
‫ٷ‬䇮Ǆ⭡ҾᡁԜ䙊ᑨ䜭ᴤௌ⅒㜭о䇝㓳ᮠᦞᤏਸ〻ᓖᴤ儈Ⲵ‫ٷ‬䇮ˈഐ↔㾱≲ᆖҐ㇇⌅Ⲵа㠤
ᙗᱟਸ⨶ⲴǄ⌘᜿ࡽ൘ㄐ㢲䇘䇪ⲴᖸཊᆖҐಘˈवᤜㅜ 2 ㄐѝⲴᡰᴹᆖҐ㇇⌅ˈ䜭ᱟа㠤ᆖ
ҐಘǄ
ᱟ੖㜭⤜・Ҿа⢩ᇊⲴ㇇⌅ˈ᧘ሬࠪԫ᜿а㠤ᆖҐಘᡰ䴰䇝㓳ṧֻᮠⲴ⭼䲀˛എㆄᱟ㛟
ᇊⲴǄѪ䘋㹼䈕᧘ሬˈ䴰㾱എ亮алㅜ 2 ㄐᇊѹⲴਈරオ䰤Ǆ൘䛓䟼ਈරオ䰤VSH, D㻛ᇊѹ
Ѫ㜭↓⺞࠶㊫䇝㓳ṧֻDⲴᡰᴹ‫ٷ‬䇮hęHⲴ䳶ਸ˖
VS H , D {h H | (¢ x, c( x)² D)(h( x) c( x)}
ਈරオ䰤Ⲵ䟽㾱᜿ѹ൘Ҿˈ⇿ња㠤ᆖҐಘ䜭䗃ࠪа኎Ҿਈරオ䰤Ⲵ‫ٷ‬䇮ˈ㘼н䇪ᴹᘾ
ṧⲴᇎֻオ䰤Xǃ‫ٷ‬䇮オ䰤Hᡆ䇝㓳ᮠᦞDǄ৏ഐᖸㆰঅˈ⭡ਈරオ䰤ⲴᇊѹˈVSH, Dवਜ਼H
ѝᡰᴹⲴа㠤‫ٷ‬䇮Ǆഐ↔ˈѪ⭼ᇊԫ᜿а㠤ᆖҐಘᡰ䴰Ⲵṧֻᮠ䟿ˈਚ䴰㾱⭼ᇊѪ‫؍‬䇱ਈර
オ䰤ѝ⋑ᴹнਟ᧕ਇ‫ٷ‬䇮ᡰ䴰Ⲵṧֻᮠ䟿Ǆл䶒Ⲵᇊѹ㋮⺞ൠ᧿䘠Ҷ䘉аᶑԦ˄㿱Haussler
1988˅˖
定义：㘳㲁а‫ٷ‬䇮オ䰤Hˈⴞḷᾲᘥcˈᇎֻ࠶ᐳDԕ৺cⲴа㓴䇝㓳ṧֻDǄᖃVSHˈ
Dѝ⇿њ‫ٷ‬䇮h‫ޣ‬Ҿc઼D䭉䈟⦷ሿҾ¦ᰦˈਈරオ䰤㻛〠Ѫ‫ޣ‬Ҿc઼Dᱟ¦-䈖ቭⲴ˄¦
-exhausted˅Ǆ
䈕ᇊѹ൘മ 7-2 ѝ⽪ࠪǄ¦-䈖ቭⲴਈරオ䰤㺘⽪о䇝㓳ṧֻа㠤Ⲵᡰᴹ‫ٷ‬䇮˄ণ䛓Ӌ

ᴹ 0 䇝㓳䭉䈟⦷Ⲵ‫ٷ‬䇮˅Ⲵⵏᇎ䭉䈟⦷ᚠྭ䜭ሿҾ¦Ǆᖃ❦ˈӾᆖҐಘⲴ䀂ᓖⴻˈᡰ㜭⸕䚃
Ⲵਚᱟ䘉Ӌ‫ٷ‬䇮㜭਼ㅹൠᤏਸ䇝㓳ᮠᦞˈᆳԜ䜭ᴹ䴦䇝㓳䭉䈟⦷Ǆਚᴹ⸕䚃⺞࠷Ⲵⴞḷᾲᘥ
Ⲵ㿲ሏ㘵᡽㜭⺞ᇊਈරオ䰤ᱟ੖Ѫ¦-䈖ቭⲴǄԔӪ᛺䇦Ⲵᱟˈণ֯н⸕䚃⺞࠷Ⲵⴞḷᾲᘥ
ᡆ䇝㓳ṧֻᣭਆⲴ࠶ᐳˈа⿽ᾲ⦷ᯩ⌅ਟ൘㔉ᇊᮠⴞⲴ䇝㓳ṧֻѻਾ⭼ᇊਈරオ䰤Ѫ¦䈖ቭ
Ⲵᾲ⦷ǄHaussler(1988)ԕл䶒Ⲵᇊ⨶ᖒᔿᨀ‫׋‬Ҷ䘉ṧⲴ⭼ᇊᯩ⌅Ǆ
Hypothesis space: ‫ٷ‬䇮オ䰤
മ 7-2 ֯ਈරオ䰤䈖ቭॆ
ਈරオ䰤VSHˈDѪ‫ٷ‬䇮hęHⲴᆀ䳶ˈަѝⲴ‫ٷ‬䇮䜭ᴹ䴦䇝㓳䭉䈟⦷˄൘മѝ㺘⽪Ѫr=0˅Ǆᖃ
❦ⵏᇎ䭉䈟⦷errorD(h)˄മѝ㺘⽪Ѫerror˅ਟ㜭䶎 0ˈণ֯䈕‫ٷ‬䇮൘ᡰᴹ䇝㓳ᮠᦞѝ䭉䈟Ѫ 0Ǆᖃ
ਈරオ䰤ѝᡰᴹ‫ٷ‬䇮h䜭┑䏣errorD(h)<¦ᰦˈਈරオ䰤᡽ᱟ¦-䈖ቭⲴǄ
ᇊ⨶ 7.1 ਈරオ䰤Ⲵ¦-䈖ቭॆ (¦-exhausting the version space)Ǆ㤕‫ٷ‬䇮オ䰤Hᴹ䲀ˈфDѪⴞ

ḷᾲᘥcⲴа㌫ࡇmı1 њ⤜・䲿ᵪᣭਆⲴṧֻˈ䛓ѸሩҾԫ᜿ 0İ¦İ1ˈਈරオ䰤VSHˈDнᱟ¦-䈖ቭ
˄‫ޣ‬Ҿc˅Ⲵᾲ⦷ሿҾᡆㅹҾ˖
H e Hm
䇱᰾˖Ԕh1ˈh2ĂĂhkѪHѝ‫ޣ‬ҾcⲴⵏᇎ䭉䈟⦷བྷҾ¦Ⲵᡰᴹ‫ٷ‬䇮Ǆᖃфӵᖃkњ‫ٷ‬䇮ѝ㠣ቁᴹ
ањᚠྭоᡰᴹmњ⤜・䲿ᵪᣭਆṧֻа㠤ᰦˈн㜭֯ਈරオ䰤¦-䈖ቭॆǄԫа‫ٷ‬䇮ⵏᇎ䭉䈟⦷བྷҾ
¦ˈфоањ䲿ᵪᣭਆṧֻа㠤Ⲵਟ㜭ᙗᴰཊѪ(1-¦)Ǆഐ↔ˈ䈕‫ٷ‬䇮оmњ⤜・ᣭਆṧֻа㠤Ⲵᾲ⦷
ᴰཊѪ(1-¦)mǄ⭡Ҿᐢ⸕ᴹkњ‫ٷ‬䇮䭉䈟⦷བྷҾ¦ˈ䛓Ѹ㠣ቁᴹањоᡰᴹmњ䇝㓳ṧֻ䜭на㠤Ⲵᾲ
⦷ᴰཊѪ
k (1 H ) m
ᒦфഐѪkİ|H|ˈкᔿᴰཊѪ|H|(1-¦)mǄᴰਾˈ֯⭘а䙊⭘нㅹᔿˈᖃ 0İ¦İ1 ࡉ(1-¦)İe-¦Ǆ

ഐ↔˖
k (1 H ) m d H 1 H d H e Hm
m
ᇊ⨶ᗇ䇱Ǆ
ࡊ᡽สҾ䇝㓳ṧֻⲴᮠⴞ mǃ‫ݱ‬䇨Ⲵ䭉䈟⦷¦઼ H ⲴབྷሿˈᗇࡠҶਈරオ䰤нᱟ¦-䈖

ቭⲴᾲ⦷Ⲵк⭼Ǆᦒ䀰ѻˈᆳሩҾԫ᜿֯⭘‫ٷ‬䇮オ䰤 H ⲴᆖҐಘ⭼ᇊҶ m њ䇝㓳ṧֻᵚ㜭
ሶᡰᴹĀൿāⲴ‫ٷ‬䇮˄ণ䭉䈟⦷བྷҾ¦Ⲵ‫ٷ‬䇮˅ࢄ䲔ࠪ৫Ⲵᾲ⦷Ǆ
ਟԕ⭘䘉а㔃䇪ᶕ⺞ᇊѪҶ߿ቁ↔Āᵚࢄ䲔āᾲ⦷ࡠаᐼᵋⲴ〻ᓖ¥ᡰ䴰Ⲵ䇝㓳ṧֻᮠǄ
⭡˖
H e Hm d G ˄7.1˅
Ӿѝ䀓ࠪ m ਟᗇ˖
1
mt (ln H ln(1 / G )) ˄7.2˅
H
ᾲᤜൠ䈤ˈᔿ 7.2 ѝⲴнㅹᔿᨀ‫׋‬Ҷ䇝㓳ṧֻᮠⴞⲴа㡜䗩⭼ˈ䈕ᮠⴞⲴṧֻ䏣ԕ൘ᡰ

ᵏᵋⲴ٬¥઼¦〻ᓖлˈ֯ԫօа㠤ᆖҐಘᡀ࣏ൠᆖҐࡠ H ѝⲴԫ᜿ⴞḷᾲᘥǄ䇝㓳ṧֻ
Ⲵᮠⴞ m 䏣ԕ‫؍‬䇱ԫ᜿а㠤‫ٷ‬䇮ᱟਟ㜭˄ਟ㜭ᙗѪ 1-¥˅䘁լ˄䭉䈟⦷Ѫ¦˅↓⺞ⲴǄ⌘
᜿ m 䲿⵰ 1/¦㓯ᙗ໎䮯ˈᒦ䲿 1/¥ሩᮠ໎䮯Ǆᆳ䘈䲿⵰‫ٷ‬䇮オ䰤 H Ⲵ㿴⁑ሩᮠ໎䮯Ǆ
⌘᜿к䶒Ⲵ⭼䲀ᴹਟ㜭ᱟ䗷儈Ⲵՠ䇑Ǆֻྲˈ㲭❦ᵚ㜭䈖ቭॆਈරオ䰤Ⲵᾲ⦷ᗵ享൘४
䰤[0, 1]޵ˈ❦㘼↔ᇊ⨶㔉ࠪⲴ䗩⭼䲿⵰|H|ሩᮠ໎䮯ǄሩҾ䏣ཏབྷⲴ‫ٷ‬䇮オ䰤ˈ䈕䗩⭼ᖸᇩ
᱃䎵䗷 1Ǆഐ↔ˈᔿ 7.2 ѝⲴнㅹᔿ㔉ࠪⲴ䗩⭼ਟ㜭䗷儈ՠ䇑Ҷᡰ䴰䇝㓳ṧֻⲴᮠ䟿Ǆ↔䗩
⭼Ⲵ㜶ᕡᙗѫ㾱ᶕⓀҾ|H|亩ˈᆳӗ⭏Ҿ䇱᰾䗷〻ѝ൘ᡰᴹਟ㜭‫ٷ‬䇮к䇑㇇䛓Ӌнਟ᧕ਇⲴ
‫ٷ‬䇮Ⲵᾲ⦷઼Ǆᇎ䱵кˈ൘䇨ཊᛵߥлਟԕᴹаᴤ㍗ࠁⲴ䗩⭼ˈԕ৺㜭ཏ㾶ⴆབྷⲴ‫ٷ‬䇮オ䰤
Ⲵ䗩⭼Ǆ䘉ᱟㅜ 7.4 㢲Ⲵѫ仈Ǆ
7.3.1 уਥ⸛ᆜҖૂуж㠪‫ٽ‬䇴
ᔿ 7.2 Ⲵ䟽㾱ᙗ൘Ҿˈᆳ੺䇹ᡁԜᴹཊቁ䇝㓳ṧֻ᡽䏣ԕ‫؍‬䇱˄ԕᾲ⦷ 1-¥˅H ѝ⇿њ

ᴹ䴦䇝㓳䭉䈟⦷Ⲵ‫ٷ‬䇮ˈަⵏᇎ䭉䈟⦷ᴰཊѪ¦ǄнᒨⲴᱟˈྲ᷌ H нवਜ਼ⴞḷᾲᘥ cˈ䛓
Ѹᒦнᙫ㜭᢮ࡠањ䴦䭉䈟⦷‫ٷ‬䇮Ǆ䘉ᰦˈᴰཊ㜭㾱≲ᆖҐಘ䗃ࠪⲴ‫ٷ‬䇮൘䇝㓳ṧֻкᴹᴰ
ሿⲴ䭉䈟⦷Ǆྲ᷌ᆖҐಘн‫ٷ‬ᇊⴞḷᾲᘥਟ൘ H ѝ㺘⽪ˈ㘼ਚㆰঅൠራ᢮ާᴹᴰሿ䇝䭉䈟
⦷Ⲵ‫ٷ‬䇮ˈ䘉ṧⲴᆖҐಘ〠Ѫнਟ⸕ᆖҐಘˈഐѪᆳн亴‫ݸ‬䇔ᇊ CHǄ
㲭❦ᔿ 7.2 สҾⲴ‫ٷ‬ᇊᱟᆖҐಘ䗃ࠪа䴦䭉䈟⦷‫ٷ‬䇮ˈሩҾᴤа⇥ⲴᛵᖒлᆖҐಘ㘳㲁

ࡠҶᴹ䶎䴦䇝㓳䭉䈟⦷Ⲵ‫ٷ‬䇮ᰦˈӽ㜭᢮ࡠањㆰঅⲴ䗩⭼Ǆ㋮⺞ൠ㺘䘠ྲлǄԔDԓ㺘ᆖ
Ґಘਟ㿲ሏࡠⲴ⢩ᇊ䇝㓳ṧֻ䳶ਸˈ㘼о↔н਼ⲴDԓ㺘൘ᮤњᇎֻ䳶ਸкⲴᾲ⦷࠶ᐳǄԔ
errorD(h)ԓ㺘‫ٷ‬䇮hⲴ䇝㓳䭉䈟⦷Ǆ⺞࠷ൠ䈤ˈerrorD(h)ᇊѹѪDѝ㻛h䈟࠶㊫Ⲵ䇝㓳ṧֻᡰ
ঐ∄ֻˈ⌘᜿errorD˄h˅ᱟ൘⢩ᇊ䇝㓳ᮠᦞṧᵜDкⲴˈᆳоⵏᇎ䭉䈟⦷errorD(h)н਼ˈਾ
㘵ᱟᇊѹ൘ᮤњᾲ⦷࠶ᐳDкⲴǄ⧠൘Ԕhbest ԓ㺘Hѝᴹᴰሿ䇝㓳䭉䈟⦷Ⲵ‫ٷ‬䇮Ǆཊቁ䇝㓳
ṧֻ᡽䏣ԕ˄ԕ䖳儈Ⲵᾲ⦷˅‫؍‬䇱ަⵏᇎ䭉䈟⦷errorD(hbest)нՊཊҾ¦+errorD(hbest)˛⌘᜿ࡽ
а㢲䇘䇪Ⲵ䰞仈ਚᱟ⧠൘䘉⿽ᛵߥⲴ⢩ֻˈަѝerrorD(hbest)ᚠྭѪ 0Ǆ
䈕䰞仈Ⲵഎㆄ˄㿱㓳Ґ 7.3˅֯⭘㊫լҾᇊ⨶ 7.1 Ⲵ䇱᰾ᯩ⌅Ǆ䘉䟼ᴹᗵ㾱ᕅ‫ޕ‬а㡜Ⲵ

Hoeffding 䗩⭼˄ᴹᰦ৸〠Ѫ䱴࣐ Chernoff 䗩⭼˅
ǄHoeffding 䗩⭼࡫⭫ⲴᱟḀһԦⲴⵏᇎᾲ
⦷৺ަ m њ⤜・䈅傼ѝ㿲ሏࡠⲴ仁⦷ѻ䰤ⲴᐞᔲǄᴤ㋮⺞ൠ䇢ˈ䘉Ӌ䗩⭼ᓄ⭘Ҿ m њн਼
Ⲵ Bernoulli 䈅傼˄ֻྲˈm ⅑ᣋ᧧а⺜ᐱˈ䈕⺜ᐱԕḀᾲ⦷ᱮ⽪Ѫ↓䶒˅Ǆ䘉⿽ᛵߥ䶎ᑨ㊫
լҾㅜ 5 ㄐ㘳㲁Ⲵ‫ٷ‬䇮䭉䈟⦷ՠ䇑䰞仈˖ণ⺜ᐱᱮ⽪Ѫ↓䶒Ⲵᾲ⦷ሩᓄࡠа䲿ᵪᣭਆᇎֻ㻛
‫ٷ‬䇮䈟࠶㊫Ⲵᾲ⦷Ǆm ⅑⤜・Ⲵ⺜ᐱᣋ᧧ሩᓄ m њ⤜・ᣭਆⲴᇎֻǄm ⅑ᇎ傼ࠪ⧠↓䶒Ⲵ
仁⦷ሩᓄҾ m њᇎֻѝ䈟࠶㊫Ⲵ仁⦷Ǆ
Hoeffding䗩⭼㺘᰾ˈᖃ䇝㓳䭉䈟⦷errorD(H)൘वਜ਼mњ䲿ᵪᣭਆṧֻⲴ䳶ਸDк⍻䟿ᰦˈ
䛓Ѹ˖
Pr[errorD (h) ! errorD (h) H ] d e 2 mH

2
ᆳ㔉ࠪҶањᾲ⦷䗩⭼ˈ䈤᰾ԫ᜿䘹ᤙⲴ‫ٷ‬䇮䇝㓳䭉䈟⦷н㜭ԓ㺘ⵏᇎᛵߥǄѪ‫؍‬䇱 L
ራ᢮ࡠⲴᴰ֣Ⲵ‫ٷ‬䇮Ⲵ䭉䈟⦷ᴹԕкⲴ䗩⭼ˈᡁԜᗵ享㘳㲁䘉|H|њ‫ٷ‬䇮ѝԫањᴹ䖳བྷ䭉
䈟⦷Ⲵᾲ⦷˖
Pr[(h H )errorD (h) ! errorD (h) H ] d H e 2 mH

2
ྲ᷌ሶ↔ᾲ⦷〠Ѫ¥ˈᒦф䰞˖ཊቁњ䇝㓳ṧֻ m ᡽䏣ԕ֯¥㔤ᤱ൘аᤷᇊⲴ٬޵˛
ਟᗇлᔿ˖
1
mt (ln H ln(1 / G )) ˄7.3˅
2H 2
䘉ᱟᔿ 7.2 Ⲵа㡜ॆᛵᖒˈ䘲⭘Ҿᖃᴰ֣‫ٷ‬䇮ਟ㜭ᴹ䶎䴦䇝㓳䭉䈟⦷ᰦˈᆖҐಘӽ㜭䘹

ᤙࡠᴰ֣‫ٷ‬䇮 hęH ⲴᛵᖒǄ⌘᜿ m ‫׍‬䎆Ҿ H ઼ 1/¥Ⲵሩᮠˈྲ൘ᔿ 7.2 ѝаṧǄ❦㘼൘䘉
њਇ䲀䖳ቁⲴᛵᖒлˈm 䲿 1/¦Ⲵᒣᯩ໎䮯ˈ㘼нᱟ 1/¦Ⲵ㓯ᙗ໎䮯Ǆ
7.3.2 ᐹቊᮽᆍⲺਾ਌ᱥ PAC ਥᆜҖⲺ
⧠൘ᡁԜᴹҶањ䇝㓳ṧֻᮠⴞⲴ䗩⭼ˈԕ㺘⽪䈕ᮠⴞѪཊቁᰦ᡽䏣ԕਟ㜭䘁լᆖҐࡠ
ⴞḷᾲᘥǄ❦ਾቡਟ⭘ᆳᶕ⺞ᇊḀӋ⢩ᇊᾲᘥ㊫Ⲵṧᵜ༽ᵲᓖ઼ PAC ਟᆖҐᙗǄ
㘳㲁ⴞḷᾲᘥ㊫ Cˈᆳ⭡ᐳቄ᮷ᆇⲴਸਆ㺘⽪Ǆᐳቄ᮷ᆇ˄literal˅ᱟԫ᜿Ⲵᐳቄਈ䟿(ྲ
Old)ˈᡆᆳⲴ੖ᇊ(ྲOld)Ǆഐ↔ˈᐳቄ᮷ᆇⲴਸਆᖒᔿਟ㜭ѪĀOldġTallāǄC ᱟ੖Ѫਟ
PAC ᆖҐⲴ˛ਟԕ䇱᰾ˈഎㆄᱟ㛟ᇊⲴǄ䇱᰾䗷〻俆‫ݸ‬ᱮ⽪ԫ᜿а㠤ᆖҐಘਚ䴰㾱ཊ亩ᔿ
ᮠⴞⲴ䇝㓳ṧֻԕᆖҐࡠ C ѝԫ᜿ cˈ❦ਾᗇࡠа⢩ᇊ㇇⌅㜭ሩ⇿䇝㓳ṧֻ֯⭘ཊ亩ᔿᰦ䰤Ǆ
㘳㲁ԫ᜿ᆖҐಘ Lˈᆳ֯⭘Ⲵ‫ٷ‬䇮オ䰤 H ㅹҾ CǄᡁԜਟԕ⭘ᔿ 7.2 䇑㇇ࠪ䏣ԕ‫؍‬䇱˄ԕ

ᾲ⦷ 1-¥˅䗃ࠪаᴰབྷ䭉䈟⦷Ѫ¦Ⲵ‫ٷ‬䇮ˈᡰ䴰Ⲵ䲿ᵪ䇝㓳ṧֻᮠⴞᱟ mǄѪ䗮ࡠ↔ⴞḷˈ
ਚ䴰㾱⺞ᇊ‫ٷ‬䇮オ䰤Ⲵ㿴⁑|H|Ǆ
㤕‫ٷ‬䇮オ䰤HᇊѹѪnњᐳቄ᮷ᆇⲴਸਆˈࡉ‫ٷ‬䇮オ䰤|H|ⲴབྷሿѪ 3nǄ৏ഐ൘Ҿˈԫа
㔉ᇊⲴ‫ٷ‬䇮ѝ⇿њਈ䟿ਟᴹй⿽ਟ㜭˖वਜ਼䈕ਈ䟿֌Ѫ᮷ᆇ˗वਜ਼䈕ਈ䟿Ⲵ੖ᇊ֌Ѫ᮷ᆇ˗
ᡆнवਜ਼䈕ਈ䟿Ǆ⭡Ҿᴹnњ䘉ṧⲴਈ䟿ˈᡰԕ‫ޡ‬ᴹ 3nњн਼Ⲵ‫ٷ‬䇮Ǆ
ሶ|H|=3nԓ‫ࡠޕ‬ᔿ 7.2 ѝˈᗇࡠԕл‫ޣ‬Ҿnᐳቄ᮷ᆇਸਆᆖҐ䰞仈Ⲵṧᵜ༽ᵲᓖ˖
1
mt (n ln 3 ln(1 / G )) ˄7.4˅
H
㤕ања㠤ᆖҐಘ㾱ᆖҐⲴⴞḷᾲᘥਟ⭡㠣ཊ 10 њᐳቄ᮷ᆇᶕ᧿䘠ˈ䛓Ѹਟᴹ 95%Ⲵ

ᾲ ⦷ ᆳ ሶ ᆖ Ґ ࡠ а њ 䭉䈟 ⦷ ሿ Ҿ 0.1 Ⲵ ‫ ٷ‬䇮 ˈ 㘼 ф ᡰ 䴰 Ⲵ 䇝㓳 ṧ ֻ ᮠ 䟿
1
m= (10ln3+ln(1/0.05))=140Ǆ
0.1
⌘᜿ m ᤹᮷ᆇᮠ䟿 n ઼ 1/¦㓯ᙗ໎䮯ˈᒦ᤹ 1/¥ሩᮠ໎䮯ǄᙫⲴ䘀㇇䟿ᱟཊቁ˛䘉ᖃ

❦‫׍‬䎆Ҿ⢩ᇊⲴᆖҐ㇇⌅Ǆ❦㘼ˈਚ㾱ᆖҐ㇇⌅Ⲵ⇿䇝㓳ṧֻ䇑㇇䟿н䎵䗷ཊ亩ᔿ㓗ˈᒦф
н䎵䗷䇝㓳ṧֻᮠⴞⲴཊ亩ᔿ㓗ˈ䛓ѸᮤփⲴ䘀㇇ҏѪཊ亩ᔿ㓗Ǆ
൘ᐳቄ᮷ᆇⲴᆖҐѝˈањ㜭ཏㅖਸ䈕㾱≲Ⲵ㇇⌅ᐢ㓿൘ㅜ 2 ㄐӻ㓽ҶǄ䘉ቡᱟ Find-S

㇇⌅ˈᆳ໎䟿ൠ䇑㇇о䇝㓳ṧֻа㠤Ⲵᴰ⢩↺‫ٷ‬䇮Ǆሩ⇿њᯠⲴ↓ֻˈ䈕㇇⌅䇑㇇Ҷᖃࡽ‫ٷ‬
䇮઼ᯠṧֻ䰤‫ޡ‬ӛⲴ᮷ᆇⲴӔ䳶ˈ֯⭘Ⲵᰦ䰤ҏ᤹ n 㓯ᙗ໎䮯Ǆഐ↔ˈFind-S ㇇⌅ਟ㜭䘁լ
↓⺞˄PAC˅ᆖҐа㊫ᑖ੖ᇊⲴ n њᐳቄ᮷ᆇਸਆⲴᾲᘥǄ
ᇊ⨶ 7.2˖ᐳቄਸਆᔿⲴ PAC ਟᆖҐᙗǄᐳቄ᮷ᆇਸਆⲴ㊫ C ᱟ⭘ Find-S ㇇⌅˄֯⭘ H=C˅PAC-

ਟᆖҐⲴǄ
䇱᰾˖ᔿ 7.4 ᱮ⽪Ҷ䈕ᾲᘥ㊫Ⲵṧᵜ༽ᵲᓖᱟ nǃ1/¥઼ 1/¦Ⲵཊ亩ᔿ㓗ˈ㘼ф⤜・Ҿ size(c)Ǆ

Ѫ໎䟿ᔿൠ༴⨶⇿њ䇝㓳ṧֻˈFind-S ㇇⌅㾱≲Ⲵ䘀㇇䟿ṩᦞ n 㓯ᙗ໎䮯ˈᒦ⤜・Ҿ 1/¥ˈ1/¦઼
size(c)Ǆഐ↔ˈ䘉аᾲᘥ㊫ᱟ Find-S ㇇⌅ PAC ਟᆖҐⲴǄ
7.3.3 ެԌᾸᘫ㊱Ⲻ PAC-ਥᆜҖᙝ
ྲࡽᡰ⽪ˈ൘ᆖҐ㔉ᇊ㊫ C ѝⲴⴞḷᾲᘥᰦˈᔿ 7.2 Ѫ⭼ᇊަṧᵜ༽ᵲᓖᨀ‫׋‬Ҷа㡜Ⲵ

ส⹰Ǆкֻሶަᓄ⭘ࡠᐳቄ᮷ᆇⲴਸਆ䘉ṧⲴ㊫࡛ѝǄᆳ䘈ਟ⭘Ҿ䇱᰾䇨ཊަԆᾲᘥ‫ޡ‬ᴹཊ
亩ᔿ㓗Ⲵṧᵜ༽ᵲᓖ˄ֻྲˈ㿱Ґ仈 7.2˅Ǆ
7.3.3.1 ᰖ‫څ‬ᆜҖಞ
ᒦ䶎ᡰᴹᾲᘥ㊫䜭ᴹྲᔿ 7.2 䛓ṧⲴཊ亩ᔿ㓗ṧᵜ༽ᵲᓖ䗩⭼Ǆֻྲˈ㘳㲁аᰐ‫ٿ‬

˄unbiased˅ᾲᘥ㊫Cˈᆳवਜ਼оX⴨‫Ⲵޣ‬ᡰᴹਟᮉᦸᾲᘥǄ䈕䳶ਸCሩᓄҾXⲴᑲ䳶ˈণXⲴ
ᴹᆀ䳶Ⲵ䳶ਸˈ‫ޡ‬वਜ਼|C|=2|X|њᾲᘥǄ㤕;ѝⲴᇎֻᇊѹѪQњᐳቄ٬⢩ᖱˈሶᴹ_;_ њн
Q
਼ᾲᘥǄᖃ❦ѪᆖҐ䘉ṧⲴᰐ‫ٿ‬ᾲᘥ㊫ˈᆖҐಘᵜ䓛ҏᗵ享֯⭘аᰐ‫ٷⲴٿ‬䇮オ䰤+ &Ǆሶ
Q

_+_ ԓ‫ࡠޕ‬ᔿ ѝˈᗇࡠѪᆖҐሩᓄҾ;Ⲵᰐ‫ٿ‬ᾲᘥ㊫Ⲵṧᵜ༽ᵲᓖǄ
1
mt (2 n ln 2 ln(1 / G )) ˄˅
H
䘉ṧˈ䈕ᰐ‫ⴞⲴٿ‬ḷᾲᘥ㊫൘ 3$& ⁑රлᴹᤷᮠ㓗Ⲵṧᵜ༽ᵲᓖǄ㲭❦ᔿ ઼

ѝᒦ䶎㍗ࠁⲴк⭼ˈᇎ䱵кਟ䇱᰾䈕ᰐ‫ٿ‬ᾲᘥ㊫Ⲵṧᵜ༽ᵲᓖ⺞Ѫ Q Ⲵᤷᮠ㓗Ǆ
7.3.3.2 K 亯 DNF ૂ K-CNF Ᾰᘫ
ᆈ൘䘉⿽ਟ㜭ˈণḀᾲᘥ㊫ᴹཊ亩ᔿ㓗Ⲵṧᵜ༽ᵲᓖˈնн㜭ཏ൘ཊ亩ᔿᰦ䰤޵㻛ᆖҐ
ࡠǄањᴹ䏓Ⲵֻᆀᱟᾲᘥ㊫&ѪN亩᷀ਆ㤳ᔿ˄N亩'1)˅ⲴᖒᔿǄN亩'1)㺘䗮ᔿᖒᔿѪ7Ģ
7ĢĂĂĢ7Nˈަѝ⇿а7L亩ѪQњᐳቄ኎ᙗ઼ᆳԜⲴ੖ᇊⲴਸਆǄ‫ٷ‬ᇊ+ &ˈᖸᇩ᱃䇱᰾_+_
QN Q QN
ᴰཊѪ ˄ഐѪᴹNњ亩ˈ⇿亩ਟᴹ њਟ㜭٬˅Ǆ⌘᜿ 䗷儈ՠ䇑Ҷ_+_ˈഐѪᆳ䟽༽䇑
㇇Ҷ7L 7Mԕ৺7L∄7Mᴤа㡜ⲴᛵᖒǄ↔к⭼ӽ❦ਟ⭘Ҿ㧧ᗇṧᵜ༽ᵲᓖⲴк⭼ˈሶަԓ‫ࡠޕ‬
ᔿ ѝ˖
1
mt (nk ln 3 ln(1 / G )) ˄˅
H
ᆳ㺘⽪ N亩 '1) Ⲵṧᵜ༽ᵲᓖѪ 1/¥ǃ1/¦ǃQ ઼ N Ⲵཊ亩ᔿ㓗Ǆ㲭❦ṧᵜ༽ᵲᓖᱟཊ

亩ᔿ㓗Ⲵˈ䇑㇇༽ᵲᓖতнᱟཊ亩ᔿ㓗ⲴˈഐѪ䈕㇇⌅ㅹ᭸ҾަԆᐢ⸕Ⲵн㜭൘ཊ亩ᔿᰦ䰤
޵䀓ߣⲴ䰞仈˄䲔䶎 53 13˅Ǆഐ↔ˈ㲭❦ N 亩 '1) ᴹཊ亩ᔿ㓗Ⲵṧᵜ༽ᵲᓖˈᆳሩҾ֯⭘
+ & ⲴᆖҐಘ⋑ᴹཊ亩ᔿ㓗Ⲵ䇑㇇༽ᵲᓖǄ
‫ޣ‬ҾN亩'1)ⲴԔӪਲ਼᛺Ⲵһᇎ൘Ҿˈ㲭❦ᆳнᱟ3$&ਟᆖҐⲴˈতᆈ൘ањᴤབྷⲴᾲᘥ
㊫ᱟ3$&ਟᆖҐⲴǄ䘉њᴤབྷⲴᾲᘥ㊫ᴹ⇿ṧֻⲴཊ亩ᔿ㓗ᰦ䰤༽ᵲᓖˈ਼ᰦᴹཊ亩ᔿ㓗Ⲵ
ṧᵜ༽ᵲᓖǄ䘉аᴤབྷⲴ㊫ѪN&1)㺘䗮ᔿ˖ԫ᜿䮯ᓖⲴਸਆᔿ7ġ7ġĂĂġ7Mˈަѝ⇿њ
7LѪᴰཊNњᐳቄਈ䟿Ⲵ᷀ਆǄᖸᇩ᱃䇱᰾N&1)वਜ਼ҶN'1)ˈഐѪԫ᜿N亩'1)ਟԕᖸᇩ᱃
ൠ䟽߉ѪN&1)㺘䗮ᔿ˄৽ѻতн❦˅Ǆ㲭❦N&1)∄N亩'1)㺘䗮࣋ᴤᕪˈնᆳᴹཊ亩ᔿ㓗ṧ
ᵜ༽ᵲᓖ઼ཊ亩ᔿ㓗ᰦ䰤༽ᵲᓖǄഐ↔ˈᾲᘥ㊫N亩'1)ᱟ֯⭘+ N&1)Ⲵањᴹ᭸㇇⌅ਟ3$&
ᆖҐⲴǄ㿱.HDUQV 9D]LUDQLѝᴤ䈖㓶Ⲵ䇘䇪Ǆ
7.4 ᰖ䲆‫ٽ‬䇴グ䰪Ⲻṭᵢགྷᵸᓜ
൘ка㢲ѝᡁԜ䇱᰾Ҷ 3$& ᆖҐⲴṧᵜ༽ᵲᓖ䲿‫ٷ‬䇮オ䰤Ⲵሩᮠ໎䮯Ǆ㲭❦ᔿ ᱟа

ᖸᴹ⭘Ⲵнㅹᔿˈնԕ_+_亩ᶕ࡫⭫ṧᵜ༽ᵲᓖᴹєњ㕪⛩Ǆ俆‫ˈݸ‬ᆳਟ㜭ሬ㠤䶎ᑨᕡⲴ䗩
⭼˄എᗶалሩҾབྷⲴ_+_൘¥кⲴ䗩⭼ਟ㜭䎵ࠪ ᖸཊ˅Ǆަ⅑ˈሩҾᰐ䲀‫ٷ‬䇮オ䰤Ⲵᛵᖒˈ
ᔿ ṩᵜᰐ⌅ᓄ⭘Ǆ
䘉䟼ᡁԜ㘳㲁 + Ⲵ༽ᵲᓖⲴਖа⿽ᓖ䟿ˈ〠Ѫ + Ⲵ 9DSQLN&KHUYRQHQNLV㔤ᓖ˄ㆰ〠

9& 㔤ᡆ 9&+˅Ǆਟԕⴻࡠˈ֯⭘ 9&+ԓᴯ_+_ҏਟԕᗇࡠṧᵜ༽ᵲᓖⲴ䗩⭼Ǆ൘䇨ཊᛵᖒ
лˈสҾ 9&+Ⲵṧᵜ༽ᵲᓖՊ∄ ᔿᗇࡠⲴᴤ㍗ࠁǄਖཆˈ䘉Ӌ䗩⭼ਟԕ࡫⭫䇨ཊᰐ䲀
‫ٷ‬䇮オ䰤Ⲵṧᵜ༽ᵲᓖˈ㘼фਟ䇱᰾⴨ᖃ㍗ࠁǄ
7.4.1 ᣼ᮙжѠᇔׁ䳼ਾ
9& 㔤㺑䟿‫ٷ‬䇮オ䰤༽ᵲᓖⲴᯩ⌅нᱟ⭘н਼‫ٷ‬䇮Ⲵᮠ䟿_+_ˈ㘼ᱟ⭘ ; ѝ㜭㻛 + ᖫᓅ४

࠶Ⲵн਼ᇎֻⲴᮠ䟿Ǆ
Ѫ㋮⺞ൠ᧿䘠䘉а⛩ˈ俆‫ݸ‬ᇊѹሩаᇎֻ䳶ਸⲴ᣶ᮓ6KDWWHULQJ᫽֌Ǆ㘳㲁ᇎֻⲴḀ

ᆀ䳶6;Ǆֻྲˈമ ᱮ⽪Ҷ;ѝањवਜ਼ њᇎֻⲴᆀ䳶Ǆ+ѝⲴ⇿њKሬ㠤6ѝⲴḀњࡂ
࠶GLFKRWRP\ˈণKሶ6࠶ࢢѪєњᆀ䳶^[ę6_K[ `ԕ৺^[ę6_K[ `Ǆ㔉ᇊḀᇎֻ䳶
_6_
ਸ6ˈᴹ ⿽ਟ㜭Ⲵࡂ࠶ˈ㲭❦ަѝⲴаӋн㜭⭡+ᶕ㺘䗮Ǆᖃ6Ⲵ⇿њਟ㜭Ⲵࡂ࠶ਟ⭡+ѝ
ⲴḀ‫ٷ‬䇮ᶕ㺘䗮ᰦˈᡁԜ〠+᣶ᮓ6Ǆ
定义： аᇎֻ䳶 S 㻛‫ٷ‬䇮オ䰤 H ᣶ᮓ(shatter) ˈᖃфӵᖃሩ S Ⲵ⇿њࡂ࠶ˈᆈ൘ H

ѝⲴḀ‫ٷ‬䇮о↔ࡂ࠶а㠤Ǆ

മ ᱮ⽪Ҷаवਜ਼ њᇎֻⲴ䳶ਸ6㻛‫ٷ‬䇮オ䰤ࡂ࠶Ⲵ㔃᷌Ǆ⌘᜿䘉 њᇎֻⲴ ⿽
ࡂ࠶ѝ⇿ањ䜭ਟ⭡Ḁ‫ٷ‬䇮㾶ⴆǄ
⌘᜿ˈྲ᷌аᇎֻ䳶ਸ⋑ᴹ㻛‫ٷ‬䇮オ䰤᣶ᮓˈ䛓Ѹᗵ❦ᆈ൘Ḁᾲᘥ˄ࡂ࠶˅ˈᆳᇊѹ൘
ᇎֻ䳶ѻкˈնн㜭⭡‫ٷ‬䇮オ䰤㺘⽪Ǆഐ↔ˈ+ Ⲵ䘉⿽᣶ᮓᇎֻ䳶ਸⲴ㜭࣋ᱟަ㺘⽪䘉Ӌᇎ
ֻкᇊѹⲴⴞḷᾲᘥⲴ㜭࣋Ⲵᓖ䟿Ǆ
Instance space X˖ᇎֻオ䰤 X
മ 7-3 㻛 8 њ‫ٷ‬䇮᣶ᮓⲴवਜ਼ 3 ᇎֻⲴ䳶ਸ
ሩ⇿⿽ਟ㜭Ⲵᇎֻࡂ࠶ˈᆈ൘ањሩᓄⲴ‫ٷ‬䇮Ǆ
7.4.2 Vapnik-Chervonenkis 㔪ᓜ
᣶ᮓаᇎֻ䳶ਸⲴ㜭࣋о‫ٷ‬䇮オ䰤Ⲵᖂ㓣‫ٿ‬㖞㍗ᇶ⴨‫ޣ‬Ǆഎᗶㅜ ㄐѝˈањᰐ‫ٷⲴٿ‬
䇮オ䰤ᱟ㜭ཏ㺘⽪ᇊѹ൘ᇎֻオ䰤 ; к⇿њਟ㜭ᾲᘥ˄ࡂ࠶˅Ⲵ‫ٷ‬䇮オ䰤Ǆㆰ⸝ൠ䇢ˈањ
ᰐ‫ٷٿ‬䇮オ䰤㜭ཏ᣶ᮓᇎֻオ䰤Ǆ䛓Ѹྲ᷌ + н㜭᣶ᮓ ;ˈնᆳਟ᣶ᮓ ; ⲴḀњབྷⲴᆀ䳶 6
Պᘾṧ˛ⴤ㿹кਟԕ䈤㻛᣶ᮓⲴ ; Ⲵᆀ䳶䎺བྷˈ+ Ⲵ㺘⽪㜭࣋䎺ᕪǄ+ Ⲵ 9& 㔤↓ᱟ䘉ṧа⿽
ᓖ䟿ḷ߶Ǆ
定义： ᇊѹ൘ᇎֻオ䰤 X кⲴ‫ٷ‬䇮オ䰤 H Ⲵ Vapnik-Chervonenkis 㔤ˈᡆ VC(H)ˈᱟ

ਟ㻛 H ᣶ᮓⲴ X Ⲵᴰབྷᴹ䲀ᆀ䳶ⲴབྷሿǄྲ᷌ X Ⲵԫ᜿ᴹ䲀བྷⲴᆀ䳶ਟ㻛 H ᣶ᮓˈ
䛓Ѹ VC(H)įĞǄ
G
⌘᜿ሩҾԫ᜿ᴹ䲀Ⲵ+ˈ9&+İORJ_+_ǄѪ䇱᰾䘉а⛩ˈ‫ٷ‬ᇊ9&+ GǄ䛓Ѹ+䴰㾱 њ
G
н਼‫ٷ‬䇮ᶕ᣶ᮓGњᇎֻǄഐ↔ İ_+_ˈᡰԕG 9&+İORJ_+_Ǆ
7.4.2.1 ⽰ׁ
ѪҶ㧧ᗇVC(H)Ⲵⴤ㿹Ⲵ᜿ѹˈ㘳㲁л䶒аӋ‫ٷ‬䇮オ䰤ⲴֻᆀǄ俆‫ٷˈݸ‬ᇊᇎֻオ䰤X
Ѫᇎᮠ䳶ਸX=˄ֻྲˈ᧿䘠ӪⲴ䓛儈height˅ˈ㘼фHѪᇎᮠ䖤кⲴ४䰤Ⲵ䳶ਸǄᦒ䀰ѻˈ
HѝⲴ‫ٷ‬䇮ᖒᔿѪa<x<bˈަѝaǃbѪԫ᜿ᇎᮠǄᆳⲴVC(H)ᱟཊቁ˛Ѫഎㆄ䘉а䰞仈ˈᗵ享
᢮ࡠ㜭㻛H᣶ᮓⲴXⲴᴰབྷᆀ䳶Ǆ㘳㲁а⢩ᇊⲴᆀ䳶ˈवਜ਼єњн਼ᇎֻˈྲS={3.1, 5.7}Ǆ
䘉њS㜭㻛H᣶ᮓੇ˛എㆄᱟ㛟ᇊⲴǄֻྲˈԕлഋњ‫ٷ‬䇮˄1<x<2˅ˈ˄1<x<4˅ ˈ˄4<x<7˅઼
˄1<x<7˅ˈਟ‫ࡠڊ‬䘉а⛩ǄᆳԜ㺘⽪ҶSкⲴഋ⿽ࡂ࠶ˈণнवਜ਼ԫօᇎֻǃਚवਜ਼ᇎֻѝ
Ⲵањǃԕ৺वਜ਼єњᇎֻǄഐѪᡁԜ᢮ࡠҶањབྷሿѪ 2 Ⲵ䳶ਸˈᆳਟ㻛H᣶ᮓˈᡰԕH
ⲴVC㔤㠣ቁѪ 2ǄབྷሿѪ 3 Ⲵ䳶ਸᱟ੖ਟ㻛᣶ᮓ˛㘳㲁а䳶ਸS={x0, x1, x2}वਜ਼ 3 њԫ᜿ᇎ
ֻǄнཡа㡜ᙗˈਟ‫ٷ‬ᇊx0<x1<x2Ǆᱮ❦ˈ↔䳶ਸн㜭㻛᣶ᮓˈഐѪवਜ਼x0઼x2նнवਜ਼x1Ⲵ
ࡂ࠶ሶн㜭⭡অњⲴ䰝४䰤ᶕ㺘⽪Ǆഐ↔ˈSѝ⋑ᴹབྷሿѪ 3 Ⲵᆀ䳶ਟ㻛᣶ᮓˈഐ↔VC(H)=2Ǆ
⌘᜿䘉䟼Hᱟᰐ䲀ⲴˈնVC(H)ᴹ䲀Ǆ
л䶒㘳㲁Ⲵᇎֻ䳶ਸSሩᓄxǃyᒣ䶒кⲴ⛩˄㿱മ 7-4˅
ǄԔHѪ↔ᒣ䶒޵ᡰᴹ㓯ᙗߣㆆ
䶒Ⲵ䳶ਸǄᦒ䀰ѻˈHሩᓄᴹৼ䗃‫Ⲵޕ‬অњᝏ⸕ಘঅ‫ٷⲴݳ‬䇮オ䰤˄㿱ㅜ 4 ㄐѝሩᝏ⸕ಘⲴ
䇘䇪˅ǄHⲴVC㔤ᱟཊቁ˛ᖸᇩ᱃ਟⴻࠪ䈕ᒣ䶒޵ԫ᜿єњн਼⛩ਟ㻛H᣶ᮓˈ䘉ᱟഐѪᡁ
Ԝਟԕ᢮ࡠ 4 њ㓯ᙗ㺘䶒ˈᆳԜवਜ਼⋑ᴹ⛩ǃަѝа⛩ᡆє⛩Ǆ3 њ⛩Ⲵ䳶ਸՊᘾѸṧ˛ਚ
㾱 3 њ⛩н‫ޡ‬㓯ˈቡਟԕ᢮ࡠ 23њ㓯ᙗ㺘䶒ᶕ᣶ᮓᆳԜǄᖃ❦ 3 њ‫ޡ‬㓯Ⲵ⛩ᰐ⌅㻛᣶ᮓ˄о
ࡽֻѝᇎ䖤к 3 њ⛩ᰐ⌅㻛᣶ᮓ਼ṧⲴ⨶⭡˅Ǆ൘↔VC㔤ᱟཊቁ˛2 䘈ᱟ 3˛㠣ቁᓄ䈕ᱟ 3Ǆ
Ѫ䇱᰾VC(H)<dˈᗵ享䇱᰾བྷሿѪdⲴ䳶ਸ䜭н㜭㻛᣶ᮓǄ൘↔ֻѝˈབྷሿѪ 4 Ⲵ䳶ਸ䜭н
㜭㻛᣶ᮓˈഐ↔VC(H)=3ǄVC㔤Ⲵᇊѹ㺘⽪ˈྲ᷌㜭᢮ࡠԫ᜿ањབྷሿѪdⲴᇎֻ䳶ਸˈᆳ
ਟ㻛᣶ᮓˈ䛓ѸVC(H)ıdǄᴤа㡜ൠˈਟ䇱᰾ˈ൘r㔤オ䰤ѝ˄ྲᴹrњ䗃‫Ⲵޕ‬ᝏ⸕ಘ˅ˈ㓯
ᙗߣㆆ䶒ⲴVC㔤Ѫr+1Ǆ
മ 7-4 ൘ xˈy ᒣ䶒ѝ㓯ᙗߣㆆ䶒Ⲵ VC 㔤Ѫ 3
a˅ањ 3 ⛩䳶ਸਟ㻛㓯ᙗߣㆆ䶒᣶ᮓǄb˅а 3 ⛩䳶ਸн㜭㻛᣶ᮓǄ
ᴰਾањֻᆀˈ‫ٷ‬ᇊXк⇿њᇎֻ⭡ᚠྭ 3 њᐳቄ᮷ᆇⲴਸਆ㺘⽪ˈ㘼ф‫ٷ‬ᇊHѝ⇿њ
‫ٷ‬䇮⭡㠣ཊ 3 њᐳቄ᮷ᆇ᧿䘠ǄVC(H)ᱟཊቁ˛ਟ䇱᰾䘉њ٬㠣ቁѪ 3Ǆሶ⇿њᇎֻ㺘⽪Ѫ
а 3 սᆇѢˈሩᓄ⇿њᇎֻⲴйњ᮷ᆇl1ˈl2઼l3Ǆ㘳㲁л䶒 3 њᇎֻ䳶ਸ˖
instance1˖100
instance2˖010
instance3˖001
䘉йњᇎֻⲴ䳶ਸਟ㻛H᣶ᮓˈᱟഐѪਟሩԫ᜿ᡰᐼᵋⲴࡂ࠶ᔪ・а‫ٷ‬䇮ˈᯩ⌅ྲл˖
ྲ᷌䈕ࡂ࠶㾱ᧂ䲔instancei ˈቡሶ᮷ᆇli ࣐‫ٷࡠޕ‬䇮ѝǄֻྲˈ㾱वਜ਼instance2 ˈфᧂ䲔
instance1઼instance3Ǆ䛓Ѹਟ֯⭘‫ٷ‬䇮l1ġl3Ǆ↔䇘䇪ਟᖸᇩ᱃ൠᢙኅࡠ⢩ᖱᮠѪnⲴᛵߥǄ
䘉ṧˈnњᐳቄ᮷ᆇਸਆⲴVC㔤㠣ቁѪnǄᇎ䱵кҏ⺞ᇎѪnˈᇎ䱵Ⲵ䇱᰾∄䖳ഠ䳮ˈഐѪᆳ
䴰㾱䈤᰾n+1 њᇎֻⲴ䳶ਸнਟ㜭㻛᣶ᮓǄ
7.4.3 ṭᵢགྷᵸᓜૂ VC 㔪
ࡽ䶒㘳㲁ҶĀᴹཊቁ䲿ᵪᣭਆ䇝㓳ṧֻ᡽䏣ԕਟ㜭䘁լ↓⺞˄PAC˅ൠᆖҐࡠ C ѝԫ᜿
ⴞḷᾲᘥā䘉њ䰞仈˄ণᴹཊቁṧֻ䏣ԕԕ 1-¥Ⲵᾲ⦷¦-䈖ቭਈරオ䰤˛˅Ǆ֯⭘ VC(H)
֌Ѫ H ༽ᵲᓖⲴᓖ䟿ˈቡᴹਟ㜭᧘ሬࠪ䈕䰞仈Ⲵਖа⿽䀓ㆄˈ㊫լҾࡽ䶒ᔿ 7.2 ѝⲴ䗩⭼Ǆ
ᯠሬࠪⲴ䗩⭼˄㿱 Blumer et al. 1989˅Ѫ˖
1
mt (4 log 2 (2 / G ) 8VC ( H ) log 2 (13 / H )) ˄7.7˅
H
↓ྲᔿ 7.2 ѝⲴ䗩⭼аṧˈᡰ䴰䇝㓳ṧֻⲴᮠⴞmԕ 1/¥Ⲵሩᮠ໎䮯Ǆն䈕䗩⭼⧠൘䲿

⵰ 1/¦Ⲵሩᮠ҈ԕ㓯ᙗ໎䮯ˈ㘼нਚᱟ㓯ᙗǄ⢩࡛㾱ᤷࠪˈࡽ䶒䗩⭼ѝⲴln|H|亩㻛ᴯᦒѪਖ
а⿽‫ٷ‬䇮オ䰤༽ᵲᓖⲴᓖ䟿ˈণVC(H)˄㘼VC(H)İlog2|H|˅Ǆ
ᔿ 7.7 ሩҾ䏣ԕਟ㜭䘁լᆖҐࡠ C ѝԫ᜿ⴞḷᾲᘥᡰ䴰Ⲵ䇝㓳ṧֻ㔉ࠪҶањк⭼ˈሩ
ᓄҾԫ᜿ᐼᵋⲴ¦઼¥Ǆ䘈ਟ㜭ᗇࡠањл⭼ˈྲл䶒ᇊ⨶ᡰᾲᤜⲴ˄㿱 Ehrenfeucht et al.
1989˅Ǆ
ᇊ⨶ 7.3 ṧᵜ༽ᵲᓖл⭼Ǆ㘳㲁ԫ᜿ᾲᘥ㊫ Cˈф VC(C)ı2ˈԫ᜿ᆖҐಘ Lˈԕ৺ԫ᜿ 0<¦<1/8ˈ

0<¥<1/100Ǆᆈ൘ањ࠶ᐳ D ԕ৺ C ѝањⴞḷᾲᘥˈᖃ L 㿲ሏࡠⲴṧֻᮠⴞሿҾлᔿᰦ˖
ª1 VC (C ) 1º
max « log(1 / G ),
¬H 32H »¼
Lሶԕ㠣ቁ¥Ⲵᾲ⦷䗃ࠪа‫ٷ‬䇮hˈ֯errorD(h)>¦Ǆ
䈕ᇊ⨶䈤᰾ˈ㤕䇝㓳ṧֻⲴᮠⴞཚቁˈ䛓Ѹнᆈ൘ᆖҐಘ㜭ཏPACᆖҐࡠԫ᜿䶎ᒣࠑⲴ
Cѝ⇿њⴞḷᾲᘥǄഐ↔ˈ䈕ᇊ⨶ᨀ‫׋‬Ҷᡀ࣏ⲴᆖҐᡰᗵ㾱Ⲵ䇝㓳ṧֻⲴᮠⴞⲴл⭼ˈᆳሩ
Ҿࡽ䶒к⭼㔉ࠪⲴ‫؍‬䇱‫ݵ‬䏣Ⲵᮠ䟿Ⲵк⭼ᱟа㺕‫ݵ‬Ǆ⌘᜿䈕л⭼ᱟ⭡ᾲᘥ㊫CⲴ༽ᵲᓖ⺞ᇊ
Ⲵˈ㘼ࡽ䶒Ⲵк⭼⭡H⺞ᇊǄѪӰѸ˛1
䈕л⭼䈤᰾ᔿ 7.7 㔉ࠪⲴк⭼⴨ᖃ㍗ࠁǄഐѪєњ䗩⭼䜭ᱟ 1/¥Ⲵሩᮠ઼ VC(H)Ⲵ㓯ᙗ

ᮠ䟿㓗Ǆ൘䘉єњ䗩⭼ѝᜏаⲴ४࡛ᱟк⭼ѝཊࠪⲴ log(1/¦)‫׍‬䎆ᙗǄ
7.4.4 ⾔㔅㖇㔒Ⲻ VC 㔪
ሩҾㅜ 4 ㄐ䇘䇪ⲴӪᐕ⾎㓿㖁㔌ˈᡁԜᴹ‫ޤ‬䏓㘳㲁ᘾṧ䇑㇇ањӂ㚄অ‫Ⲵݳ‬㖁㔌Ⲵ VC
㔤ˈྲ⭡৽ੁՐ᫝䗷〻䇝㓳Ⲵࡽ侸㖁㔌Ǆᵜ㢲㔉ࠪҶа㡜ᙗⲴ㔃䇪ˈԕ䇑㇇࠶ቲᰐ⧟㖁㔌Ⲵ
VC 㔤Ǆ䘉а VC 㔤ਟ㻛⭘Ҿ⭼ᇊ䇝㓳ṧֻⲴᮠ䟿ˈ䈕ᮠ䗮ࡠཊབྷⲴ᡽䏣ԕ᤹➗ᐼᵋⲴ¦઼
¥٬䘁լਟ㜭↓⺞ൠᆖҐࡠањࡽ侸㖁㔌Ǆᵜ㢲൘ㅜа⅑䰵䈫ᰦਟᘭ⮕ᦹˈ㘼нཡ䘎㔝ᙗǄ
㘳㲁ањ⭡অ‫ݳ‬㓴ᡀⲴ㖁㔌 Gˈᆳᖒᡀањ࠶ቲᴹੁᰐ⧟മǄᴹੁᰐ⧟(directed acyclic)

മᱟᕗᑖᴹᯩੁ˄ྲঅ‫ݳ‬ᴹ䗃‫઼ޕ‬䗃ࠪ˅ˈնнᆈ൘ᴹੁ⧟ⲴമǄ࠶ቲ˄layered˅മѝ㢲⛩
ਟ㻛ࡂ࠶Ѫቲˈ䘉ṧᡰᴹㅜ l ቲࠪᶕⲴᴹੁ䗩䘋‫ࡠޕ‬ㅜ l+1 ቲ㢲⛩Ǆㅜ 4 ㄐӻ㓽Ⲵ࠶ቲࡽ侸
㖁㔌ቡᱟ䘉ṧⲴ࠶ቲᴹੁᰐ⧟മⲴֻᆀǄ
ਟԕⴻࠪˈ䘉ṧⲴ㖁㔌ⲴVC㔤Ⲵ⭼ᇊਟԕสҾަമⲴ㔃ᶴ઼ᶴ䙐䈕മⲴสᵜঅ‫Ⲵݳ‬VC
㔤ǄѪᖒᔿॆൠ᧿䘠ˈ俆‫ݸ‬ᇊѹаӋᵟ䈝ǄԔnѪ㖁㔌GⲴ䗃‫ޕ‬ᮠⴞˈᒦф‫ٷ‬ᇊਚᴹ 1 њ䗃
ࠪ㔃⛩ǄԔGⲴ⇿њ޵䜘অ‫ݳ‬Ni˄ণ⇿њ䶎䗃‫ޕ‬㢲⛩˅ᴹᴰཊr њ䗃‫ˈޕ‬ᒦᇎ⧠аᐳቄ࠭ᮠ
ci˖rė{0,1}ᖒᡀа࠭ᮠ㊫CǄֻྲˈ㤕޵䜘㢲⛩Ѫᝏ⸕ಘˈ䛓ѸCѪᇊѹ൘rкⲴ㓯ᙗ䰸
٬࠭ᮠ㊫Ǆ
⧠൘ਟᇊѹ C Ⲵ G-ਸᡀ˄G-composition˅Ѫˈ൘ G ѝ⤜・অ‫ݳ‬䜭ਆ㊫ C ѝⲴ࠭ᮠᰦˈ

⭡㖁㔌 G 㜭ᇎ⧠ᡰᴹ࠭ᮠⲴ㊫Ǆㆰঅൠ䈤ˈC Ⲵ G-ਸᡀᱟਟ⭡㖁㔌 G 㺘⽪Ⲵ‫ٷ‬䇮オ䰤Ǆ
л䶒Ⲵᇊ⨶⭼ᇊҶ C Ⲵ G ਸᡀⲴ VC 㔤ˈสҾ C Ⲵ VC 㔤઼ G Ⲵ㔃ᶴǄ
1
ᨀ⽪˖ྲ᷌ᡁԜ൘л⭼ѝ⭘HԓᴯCˈᖃHCᰦՊᗇࡠmⲴањᴤ㍗ࠁⲴ⭼䲀
ᇊ⨶ 7.4 ࠶ቲᴹੁᰐ⧟㖁㔌ⲴVC㔤Ǆ˄㿱Kearns & Vazirani 1994˅ԔGѪа࠶ቲᴹੁᰐ⧟മˈᴹ
nњ䗃‫ޕ‬㢲⛩઼sı2 њ޵䜘㢲⛩ˈ⇿њਟᴹ㠣ቁrњ䗃‫ޕ‬ǄԔCѪVC㔤ѪdⲴrкⲴᾲᘥ㊫ˈሩᓄҾਟ
⭡⇿њ޵䜘㢲⛩᧿䘠Ⲵ࠭ᮠ䳶ਸǄԔCGѪCⲴGਸᡀˈሩᓄҾਟ⭡G㺘⽪Ⲵ࠭ᮠ䳶ਸǄ䛓ѸVC(CG)İ
dslog(es)ˈަѝeѪ㠚❦ሩᮠᓅǄ
⌘᜿䘉а㖁㔌 G Ⲵ VC 㔤䗩⭼䲿অњঅ‫ Ⲵݳ‬VC 㔤 d 㓯ᙗ໎䮯ˈᒦ䲿 s˄ণ㖁㔌ѝ䰸٬

অ‫Ⲵݳ‬ᮠⴞ˅Ⲵሩᮠ҈㓯ᙗ໎䮯Ǆ
‫ٷ‬ᇊ㾱㘳㲁Ⲵ࠶ቲᴹੁᰐ⧟㖁㔌ѝঅњ㢲⛩䜭ᱟᝏ⸕ಘǄഎᗶㅜ 4 ㄐѝᨀࡠⲴˈr䗃‫ޕ‬
ᝏ⸕ಘ֯⭘㓯ᙗߣㆆ䶒ᶕ㺘⽪rкⲴᐳቄ࠭ᮠǄྲ 7.4.2.1 㢲ᤷࠪⲴ䛓ṧˈ൘rкⲴ㓯ᙗߣ
ㆆ䶒ⲴVC㔤Ѫr+1Ǆഐ↔ˈঅ⤜Ⲵr䗃‫ޕ‬ᝏ⸕ಘVC㔤Ѫr+1Ǆਟ֯⭘䘉а㔃᷌৺к䶒Ⲵᇊ⨶
ᶕ䇑㇇वਜ਼sњr 䗃‫ޕ‬ᝏ⸕ಘⲴ࠶ቲᰐ⧟㖁㔌ⲴVC㔤䗩⭼ˈྲл˖
VC(CGperceptrons ) d 2(r 1) s log(es)
⧠൘ਟԕ䇑㇇ˈѪҶ൘䈟ᐞ¦㤳ത޵ԕ㠣ቁ 1-¥Ⲵᾲ⦷ᆖҐࡠᶕ㠚CGperceptronsⲴⴞḷᾲ
ᘥˈ䏣ཏⲴ䇝㓳ṧֻᮠⴞmⲴ䗩⭼Ǆሶк䶒㖁㔌VC㔤Ⲵ㺘䗮ᔿԓ‫ࡠޕ‬ᔿ 7.7ˈਟᴹ˖
1
mt (4 log(2 / G ) 8VC ( H ) log(13 / H ))
H
1
t (4 log(2 / G ) 16(r 1) s log(es) log(13 / H )) ˄7.8˅
H
ྲ↔ᝏ⸕ಘ㖁㔌ֻᆀᡰ⽪ˈк䶒Ⲵᇊ⨶Ⲵ֌⭘൘Ҿˈᆳᨀ‫׋‬Ҷања㡜ᙗᯩ⌅ˈสҾ㖁
㔌㔃ᶴ઼অњঅ‫ Ⲵݳ‬VC 㔤⭼ᇊ࠶ቲᰐ⧟অ‫ݳ‬㖁㔌Ⲵ VC 㔤Ǆн䗷ˈк䶒Ⲵ㔃᷌н㜭ⴤ᧕ᓄ
⭘ҾਾੁՐ᫝Ⲵ㖁㔌ˈ৏ഐᴹєњǄ俆‫↔ˈݸ‬㔃᷌ᓄ⭘Ҿᝏ⸕ಘ㖁㔌ˈ㘼нᱟ sigmoid অ‫ݳ‬
㖁㔌ˈਾ㘵ᱟਾੁՐ᫝㇇⌅ᓄ⭘Ⲵ㤳തǄ❦㘼ˈ⌘᜿ࡠ sigmoid অ‫ Ⲵݳ‬VC 㔤㠣ቁՊоᝏ⸕
ಘঅ‫ Ⲵݳ‬VC 㔤аṧབྷǄഐѪ䙊䗷֯⭘䏣ཏⲴᵳ٬ˈsigmoid অ‫ݳ‬ਟԕԫ᜿㋮ᓖ䙬䘁ᝏ⸕ಘǄ
ഐ↔ˈк䶒Ⲵ m 䗩⭼㠣ቁՊо sigmoid অ‫ݳ‬㓴ᡀⲴ࠶ቲᰐ⧟㖁㔌ѝⲴаṧབྷǄк䘠㔃䇪Ⲵㅜ
Ҽњн䏣൘Ҿˈᆳн㜭༴⨶ਾੁՐ᫝ѝⲴ䇝㓳䗷〻ˈণᔰ࿻ԕ㓖ㅹҾ 0 Ⲵᵳ٬ˈ❦ਾ৽༽ൠ
ᴤᯠ䈕ᵳ٬ˈⴤࡠ᢮ࡠаਟ᧕ਇⲴ‫ٷ‬䇮Ǆഐ↔ˈਾੁՐ᫝ᑖᴹӔ৹傼䇱㓸→ḷ߶ˈᆳӗ⭏а
њᴤ‫ྭٿ‬ሿᵳ٬㖁㔌Ⲵᖂ㓣‫ٿ‬㖞Ǆ䘉аᖂ㓣‫ٿ‬㖞ˈ䱽վҶᴹ᭸Ⲵ VC 㔤ˈᱟк䶒Ⲵ࠶᷀ᡰн
㜭⏥ⴆⲴǄ
7.5 ᆜҖⲺ࠰䭏⮂䲆⁗ශ
䲔Ҷ PAC ᆖҐ⁑රԕཆˈ䇑㇇ᆖҐ⨶䇪䘈㘳㲁Ҷཊ⿽н਼Ⲵ䰞仈ṶᷦǄᐢ㓿⹄ウⲴᆖ
Ґ䰞仈Ṷᷦѝн਼ѻ༴൘Ҿ䇝㓳ṧֻⲴ⭏ᡀᯩᔿ˄㻛ࣘ㿲ሏᆖҐṧֻ䘈ᱟѫࣘᨀࠪḕ䈒˅ˈ
ᮠᦞѝⲴಚ༠˄ᴹಚ༠ᮠᦞ䘈ᱟᰐᐞ䭉ᮠᦞ˅ˈᡀ࣏ᆖҐⲴᇊѹ˄ᗵ享ᆖࡠ↓⺞Ⲵⴞḷᾲᘥˈ
䘈ᱟᴹаᇊⲴਟ㜭ᙗ઼䘁լᙗ˅ˈᆖҐಘᡰ‫ٷⲴڊ‬ᇊ˄ᇎֻⲴ࠶ᐳᛵߥԕ৺ᱟ੖ CH˅ˈ઼
䇴ՠᆖҐಘⲴᓖ䟿ḷ߶˄䇝㓳ṧֻᮠ䟿ǃࠪ䭉ᮠ䟿ǃ䇑㇇ᰦ䰤˅Ǆ
ᵜ㢲ሶ㘳㲁ᵪಘᆖҐⲴࠪ䭉⭼䲀˄mistake bound˅⁑රˈަѝᆖҐಘ䇴ՠḷ߶ᱟᆳ൘᭦
ᮋࡠ↓⺞‫ٷ‬䇮ࡽᙫⲴࠪ䭉ᮠǄྲ൘ PAC 䰞仈Ṷᷦѝаṧˈ䘉䟼‫ٷ‬ᇊᆖҐಘ᧕᭦ࡠа㌫ࡇⲴ
䇝㓳ṧֻǄ❦㘼ˈ䘉䟼ᡁԜᐼᵋ⇿᧕ਇࡠањṧֻ xˈᆖҐಘᗵ享‫ݸ‬亴⍻ⴞḷ٬ c(x)ˈѻਾ
޽⭡ᯭᮉ㘵㔉ࠪ↓⺞Ⲵⴞḷ٬Ǆ䘉䟼㘳㲁Ⲵ䰞仈ᱟ˖Ā൘ᆖҐಘᆖҐࡠⴞḷᾲᘥࡽˈᆳⲴ亴
⍻Պᴹཊቁ⅑ࠪ䭉āǄ䘉а䰞仈൘ᇎ䱵⧟ຳлॱ࠶䟽㾱ˈަѝᆖҐ䗷〻о㌫㔏䘀㹼਼ᰦ䘋㹼ˈ
㘼нᱟ㓿䗷а⇥⿫㓯Ⲵ䇝㓳䗷〻Ǆֻྲˈྲ᷌㌫㔏㾱ᆖ⵰亴⍻ଚӋؑ⭘঑䍝⢙ਟ㻛‫ݱ‬䇨ˈଚ
ӋᴹⅪ䇸㹼Ѫˈᗵ享สҾ൘֯⭘ѝᩌ䳶Ⲵᮠᦞˈ❦ਾᡁԜቡ㾱൘ަ᭦ᮋࡠ↓⺞ⴞḷ࠭ᮠࡽ֯
ަࠪ䭉ⲴᮠⴞᴰሿॆǄ䘉䟼ࠪ䭉Ⲵᙫᮠਟ㜭∄䇝㓳ṧֻⲴᙫᮠᴤ䟽㾱Ǆ
䘉⿽ࠪ䭉⭼䲀ᆖҐ䰞仈ਟԕ൘䇨ཊ⢩↺Ⲵ㛼Ჟѝ䘋㹼⹄ウǄֻྲˈᡁԜਟԕ䇑㇇ᆖҐಘ
൘ PAC ᆖҐࡠⴞḷᾲᘥࡽࠪ䭉Ⲵ⅑ᮠǄ൘л䶒ⲴֻᆀѝˈᡁԜਚ㘳㲁൘ᆖҐಘ⺞࠷ᆖࡠⴞ
ḷᾲᘥࡽࠪ䭉Ⲵ⅑ᮠǄަѝ⺞࠷ᆖࡠⴞḷᾲᘥ᜿ણ⵰(x)h(x)=c(x)Ǆ
7.5.1 Find-S ㇍⌋Ⲻ࠰䭏⮂䲆
޽⅑㘳㲁‫ٷ‬䇮オ䰤Hवਜ਼㠣ཊnњᐳቄ᮷ᆇ˄l1ĂlnᡆᆳԜⲴ੖ᇊ˅ⲴਸਆⲴᛵߥ˄ֻྲ˖
RichġHandsome˅Ǆഎᗶㅜ 2 ㄐѝⲴFind-S㇇⌅ˈᆳ໎䟿ᔿൠ䇑㇇о䇝㓳ṧֻа㠤Ⲵᶱབྷ⢩
↺‫ٷ‬䇮Ǆሩ‫ٷ‬䇮オ䰤HⲴFind-S㇇⌅Ⲵањㆰ⌱ᇎ⧠ྲл˖
Find-S˖
x ሶh㻛࿻ॆѪᴰ⢩↺‫ٷ‬䇮l1ġl1ġl2ġl2…lnġln
x ሩ⇿њ↓ֻ x
x Ӿ h ѝ〫৫ԫօн┑䏣 x Ⲵ᮷ᆇ
x 䗃ࠪ‫ٷ‬䇮 h
ྲ᷌ CH ф䇝㓳ᮠᦞᰐಚ༠ˈFind-S ᶱ䲀ᰦ᭦ᮋࡠањᰐᐞ䭉Ⲵ‫ٷ‬䇮ǄFind-S ᔰ࿻Ҿ

ᴰ⢩↺Ⲵ‫ٷ‬䇮˄ᆳሶ⇿њᇎֻ࠶Ѫ৽ֻ˅ˈ❦ਾ໎䟿ᔿൠ⌋ॆ䈕‫ٷ‬䇮ˈԕ㾶ⴆ㿲ሏࡠⲴ↓ֻǄ
ሩҾ䘉䟼֯⭘Ⲵ‫ٷ‬䇮㺘⽪ˈ⌋ॆ䗷〻⭡ࡐ䲔н┑䏣Ⲵ᮷ᆇ᫽֌ᶴᡀǄ
ᱟ੖ਟԕ䇑㇇ࠪањ䗩⭼ˈԕ᧿䘠 Find-S ൘⺞࠷ᆖࡠⴞḷᾲᘥ c ࡽ‫ޘ‬䜘Ⲵࠪ䭉⅑ᮠ˛

എㆄᱟ㛟ᇊⲴǄѪ䇱᰾ѻˈ俆‫⌘ݸ‬᜿ྲ᷌ cHˈ䛓Ѹ Find-S ≨䘌нՊሶа৽ֻ䭉䈟ൠࡂ࠶
Ѫ↓ֻǄ৏ഐѪᖃࡽ‫ٷ‬䇮 h ᙫ∄ⴞḷᾲᘥ c ᴤ⢩↺ǄᡁԜਚ䴰㾱䇑㇇ሶ↓ֻࡂ࠶Ѫ৽ֻⲴࠪ
䭉⅑ᮠǄ൘ Find-S ⺞࠷ᗇࡠ c ࡽˈ䘉ṧⲴࠪ䭉ᴹཊቁ⅑˛㘳㲁 Find-S ㇇⌅䙷ࡠⲴㅜањ↓
ֻǄᆖҐಘᖃ❦Պ൘࠶㊫∄ֻᰦࠪ䭉ˈഐѪᆳⲴࡍ࿻‫ٷ‬䇮ሶ‫ޘ‬䜘ᇎֻ䜭࠶Ѫ৽ֻǄ❦㘼ˈ㔃
᷌ሶᱟࡍ࿻‫ٷ‬䇮ѝ 2n њ亩ѝॺᮠሶ㻛ࡐ৫ˈਚ⮉л n њ亩Ǆሩ⇿њਾ㔝Ⲵ↓ֻˈ㤕ᆳ㻛ᖃ
ࡽ‫ٷ‬䇮䈟࠶㊫ˈ࢙։ n њ亩ѝ㠣ቁᴹа亩ᗵ享Ӿ‫ٷ‬䇮ѝࡐ৫Ǆഐ↔ˈࠪ䭉Ⲵᙫᮠ㠣ཊѪ n+1Ǆ
䈕ࠪ䭉⅑ᮠᱟᴰൿᛵߥлᡰ䴰Ⲵ⅑ᮠˈሩᓄҾᆖҐᴰа㡜Ⲵⴞḷᾲᘥ˖˄x˅c(x)=1ˈᒦф
ᇎֻᒿࡇҏᱟᴰൿᛵߥлⲴ⇿⅑ࠪ䭉ਚ㜭〫৫ањ᮷ᆇǄ
7.5.2 Halving ㇍⌋Ⲻ࠰䭏⮂䲆
ㅜҼњֻᆀˈ㘳㲁ањ㇇⌅ˈᆳⲴᢗ㹼䗷〻ᱟ㔤ᣔањਈරオ䰤ˈᒦ൘䙷ࡠᯠṧֻᰦ㋮
ॆ䈕ਈරオ䰤Ǆㅜ 2 ㄐⲴ‫ى‬䘹⎸䲔㇇⌅઼ࡇ㺘ਾ⎸䲔㇇⌅䜭ᱟ䘉ṧⲴ㇇⌅Ǆᵜ㢲ᡁԜ᧘ሬ䘉
ṧⲴᆖҐಘ䪸ሩԫ᜿ᴹ䲀‫ٷ‬䇮オ䰤 H ᴰൿᛵߥлࠪ䭉ᮠ䟿Ⲵ䗩⭼ˈᒦ޽⅑‫ٷ‬ᇊⴞḷᾲᘥ㜭
㻛⺞࠷ᆖҐࡠǄ
Ѫ࠶᷀ᆖҐ䗷〻ѝࠪ䭉Ⲵᮠ䟿ˈᗵ享俆‫ݸ‬㋮⺞ᤷᇊᆖҐಘሩ⇿њᯠᇎֻՊ֌ࠪᘾṧⲴ亴
⍻Ǆ‫ٷ‬ᇊ䈕亴⍻ᱟ൘ᖃࡽਈරオ䰤Ⲵᡰᴹ‫ٷ‬䇮ѝ֌ཊᮠᣅ⾘ᗇᶕǄྲ᷌ਈරオ䰤ѝཊᮠ‫ٷ‬䇮
ሶᯠᇎֻࡂ࠶Ѫ↓ֻˈ䛓Ѹ䈕亴⍻⭡ᆖҐಘ䗃ࠪǄ੖ࡉ䗃ࠪ৽ֻⲴ亴⍻Ǆ
䘉⿽ሶਈරオ䰤ᆖҐ઼⭘ཊᮠᣅ⾘ᶕ䘋㹼ਾ㔝亴⍻є㘵㔃ਸ䎧ᶕⲴ㇇⌅䙊ᑨ㻛〠Ѫ
Halving ㇇⌅Ǆሩԫ᜿ᴹ䲀 HˈHalving ㇇⌅൘⺞࠷ᆖҐࡠⴞḷᾲᘥࡽࠪ䭉Ⲵᴰབྷ⅑ᮠᱟཊቁ˛
⌘᜿Ā⺞࠷āൠᆖҐࡠⴞḷᾲᘥㅹҾ䈤ࡠ䗮ањ⣦ᘱˈਈරオ䰤ѝਚवਜ਼ањ‫ٷ‬䇮˄ྲᖰᑨ
䛓ṧ‫ٷ‬ᇊⴞḷᾲᘥ c ൘ H ѝ˅Ǆ
Ѫ᧘ሬ䈕ࠪ䭉⭼䲀ˈ⌘᜿Halving㇇⌅ਚ൘ᖃࡽਈරオ䰤Ⲵཊᮠ‫ٷ‬䇮н㜭↓⺞࠶㊫ᯠṧ
ֻᰦࠪ䭉Ǆ൘䘉⿽ᛵߥлˈаն↓⺞࠶㊫㔃᷌ᨀ‫׋‬㔉ᆖҐಘਾˈਈරオ䰤ਟ߿ሿࡠᆳⲴᴰཊ
аॺབྷሿ˄ণਚᴹᣅቁᮠ⾘Ⲵ‫ٷ‬䇮㻛‫˅⮉؍‬Ǆ⭡Ҿ⇿⅑ࠪ䭉ሶਈරオ䰤㠣ቁ߿ሿаॺˈ㘼ф
ࡍ࿻ਈරオ䰤वਜ਼|H|њᡀઈˈᡰԕਈරオ䰤ࡠਚवਜ਼ањᡀઈࡽࠪ䭉⅑ᮠᴰབྷѪlog2|H|Ǆᇎ
䱵кਟ䇱᰾䈕䗩⭼Ѫ¬log2|H|¼Ǆֻྲˈ㘳㲁|H|=7 ⲴᛵߥǄㅜањࠪ䭉ਟሶ|H|߿ሿࡠᴰཊѪ 3ˈ
ㅜҼ⅑ࠪ䭉ቡਟሶަ߿ሿࡠ 1Ǆ
⌘᜿¬log2|h|¼ѪᴰൿᛵߥлⲴ䗩⭼ˈᒦфᴹਟ㜭Halving㇇⌅нࠪԫօᐞ䭉ቡ⺞࠷ᆖҐࡠ
ⴞḷᾲᘥǄഐѪণ֯ཊᮠ⾘㔃᷌ᱟ↓⺞Ⲵˈ㇇⌅ӽሶ〫৫䛓Ӌн↓⺞Ⲵǃቁᮠ⾘‫ٷ‬䇮Ǆ㤕↔
ᛵߥ൘ᮤњ䇝㓳䗷〻ѝਁ⭏ˈ䛓Ѹਈරオ䰤ਟ൘нࠪᐞ䭉Ⲵᛵߥл߿ሿࡠঅњᡀઈǄ
ሩ Halving ㇇⌅Ⲵањᴹ䏓Ⲵᢙኅᱟ‫ݱ‬䇨‫ٷ‬䇮ԕн਼Ⲵᵳ٬䘋㹼ᣅ⾘Ǆㅜ 6 ㄐ᧿䘠Ҷ䍍

ਦᯟᴰՈ࠶㊫ಘˈᆳቡ൘‫ٷ‬䇮ѝ䘋㹼࣐ᵳᣅ⾘Ǆ൘䍍ਦᯟᴰՈ࠶㊫ಘѝˈѪ⇿њ‫ٷ‬䇮䍻ҸⲴ
ᵳ٬Ѫަ᧿䘠ⴞḷᾲᘥⲴՠ䇑ਾ傼ᾲ⦷˄㔉ᇊ䇝㓳ᮠᦞл˅Ǆᵜ㢲Ⲵਾ䶒ሶ᧿䘠ਖаสҾ࣐
ᵳᣅ⾘Ⲵ㇇⌅ˈ〠Ѫ࣐ᵳཊᮠ㇇⌅Ǆ
7.5.3 ᴶՎ࠰䭏⮂䲆
к䶒Ⲵ࠶᷀㔉ࠪҶєњ⢩ᇊ㇇⌅˖Find-S઼‫ى‬䘹⎸䲔㇇⌅൘ᴰൿᛵߥлⲴࠪ䭉⭼䲀Ǆа
њᖸᴹ䏓Ⲵ䰞仈ᱟˈሩҾԫ᜿ᾲᘥ㊫Cˈ‫ٷ‬ᇊH=CˈᴰՈⲴࠪ䭉䗩⭼ᱟӰѸ˛ᴰՈࠪ䭉䗩⭼
ᱟᤷ൘ᡰᴹਟ㜭ⲴᆖҐ㇇⌅ѝˈᴰൿᛵߥлࠪ䭉䗩⭼ѝᴰሿⲴ䛓ањǄᴤ㋮⺞ൠ䈤ˈሩԫ᜿
ᆖҐ㇇⌅A઼ԫ᜿ⴞḷᾲᘥcˈԔMA(c)ԓ㺘AѪҶ⺞࠷ᆖࡠcˈ൘ᡰᴹਟ㜭䇝㓳ṧֻᒿࡇѝࠪ
䭉Ⲵᴰབྷ٬Ǆ⧠൘ሩҾԫ᜿䶎オᾲᘥ㊫CˈԔMA(C)į max cC M A (c) Ǆ⌘᜿к䶒ᡁԜ䇱᰾
ҶᖃCᱟ㠣ཊnњᐳቄ᮷ᆇ᧿䘠Ⲵᾲᘥ㊫ᰦˈ M Find S (C ) n 1 Ǆ਼ᰦˈሩԫ᜿ᾲᘥ㊫Cˈ
ᡁԜᴹ M Halving (C ) d log 2 (| C |) Ǆ
л䶒ᇊѹᾲᘥ㊫ C ⲴᴰՈࠪ䭉䗩⭼Ǆ
定义： ԔCѪԫ᜿䶎オᾲᘥ㊫ǄCⲴᴰ ᴰՈࠪ䭉⭼䲀˄optimal mistake bound˅ᇊѹѪ

Opt(C)ˈᱟᡰᴹਟ㜭ᆖҐ㇇⌅AѝMA(C)Ⲵᴰሿ٬Ǆ
Opt (C ) { min M A (C )
Alearning _a lg orithms
⭘䶎ᖒᔿⲴ䈝䀰ᶕ䇢ˈ䈕ᇊѹ㺘᰾ Opt(C)ᱟ C ѝᴰഠ䳮Ⲵ䛓њⴞḷᾲᘥ֯⭘ᴰн࡙Ⲵ

䇝㓳ṧֻᒿࡇˈ⭘ᴰྭⲴ㇇⌅Ⲵࠪ䭉⅑ᮠǄLittlestone˄1987˅䇱᰾ሩԫ᜿ᾲᘥ㊫ Cˈ൘ C
ⲴᴰՈࠪ䭉䗩⭼ˈHalving ㇇⌅䗩⭼઼ C Ⲵ VC 㔤ѻ䰤ᆈ൘аᴹ䏓Ⲵ㚄㌫ˈྲл˖
VC (C ) d Opt (C ) d M Halving (C ) d log 2 (| C |)
ᴤ䘋а↕ˈᆈ൘䘉ṧⲴᾲᘥ㊫֯к䶒Ⲵ 4 њ䟿ᚠྭ⴨ㅹǄ䘉ṧⲴᾲᘥ㊫ަѝѻаᱟԫ᜿
ᴹ䲀䳶ਸⲴᑲ䳶CPǄ൘↔ᛵߥлˈVC(CP)=|X|=log2(|CP|)ˈഐ↔ᡰᴹ䘉 4 њ䟿⴨ㅹǄLittlestone
˄1987˅ᨀ‫׋‬ҶަԆᾲᘥ㊫ⲴֻᆀˈަѝVC(C)ѕṬሿҾOpt(C)ˈOpt(C)ѕṬሿҾMHalving(C)Ǆ
7.5.4 ࣖᵹཐᮦ㇍⌋
ᵜ㢲䇘䇪 Halving ㇇⌅Ⲵᴤа㡜Ⲵᖒᔿˈ〠Ѫ࣐ᵳཊᮠ㇇⌅Ǆ࣐ᵳཊᮠ㇇⌅䙊䗷൘а亴

⍻㇇⌅⊐ѝ䘋㹼࣐ᵳᣅ⾘ᶕ֌ࠪ亴⍻ˈᒦ䙊䗷᭩ਈ⇿њ亴⍻㇇⌅Ⲵᵳ䟽ᶕᆖҐǄ䘉Ӌ亴⍻㇇
⌅ਟ㻛ⴻ֌ᱟ H ѝⲴн਼‫ٷ‬䇮ˈᡆ㻛ⴻ֌ᵜ䓛䲿ᰦ䰤ਈॆⲴн਼ᆖҐ㇇⌅ǄሩҾ䘉Ӌ亴⍻
㇇⌅ˈᡰ䴰㾱Ⲵਚᱟ൘㔉ᇊаᇎֻᰦ亴⍻ⴞḷᾲᘥⲴ٬Ǆ࣐ᵳཊᮠ㇇⌅Ⲵањᴹ䏓኎ᙗᱟᆳ
ਟԕ༴⨶на㠤Ⲵ䇝㓳ᮠᦞǄ䘉ᱟഐѪᆳнՊ⎸䲔ᦹоṧֻна㠤Ⲵ‫ٷ‬䇮ˈ㘼ਚᱟ䱽վަᵳ
䟽ǄᆳⲴㅜҼњᴹ䏓኎ᙗᱟˈ㾱䇑㇇↔㇇⌅Ⲵࠪ䭉ᮠ䟿䗩⭼ˈਟԕ⭘亴⍻㇇⌅⊐ѝᴰྭⲴ䛓
њ㇇⌅Ⲵࠪ䭉ᮠ䟿ᶕ䇑㇇Ǆ
࣐ᵳཊᮠ㇇⌅аᔰ࿻ሶ⇿њ亴⍻㇇⌅䍻ԕᵳ䟽 1ˈ❦ਾ㘳㲁䇝㓳ṧֻǄᰐ䇪օᰦа亴⍻
㇇⌅䈟࠶㊫аᯠ䇝㓳ṧֻᰦˈᆳⲴᵳ䟽㻛҈ԕḀњ㌫ᮠ£ˈ0İ£Ǆ࣐ᵳཊᮠ㇇⌅Ⲵ⺞࠷
ᇊѹ㿱㺘 Ǆ⌘᜿ྲ᷌£ ˈ䛓Ѹ࣐ᵳཊᮠ㇇⌅ㅹҾ Halving ㇇⌅Ǆਖаᯩ䶒ˈྲ᷌Ѫ£
䘹ᤙަԆⲴ٬ˈ⋑ᴹањ亴⍻㇇⌅Պ㻛ᆼ‫ޘ‬৫䲔Ǆྲ᷌а㇇⌅䈟࠶㊫ањṧֻˈᆳ൘ሶᶕՊ
ঐ䖳ቁⲴ⾘ᮠ∄ֻǄ
㺘 7-1 ࣐ᵳཊᮠ㇇⌅
aiԓ㺘㇇⌅⊐Aѝㅜiњ亴⍻㇇⌅Ǆwiԓ㺘оai⴨‫ޣ‬㚄Ⲵᵳ٬Ǆ
ሩᡰᴹiˈࡍ࿻ॆwiѪ 1
ሩ⇿њ䇝㓳ṧֻ<x, c(x)>
ࡍ࿻ॆq0઼q1Ѫ 0
ሩ⇿њ亴⍻㇇⌅ai
ྲ᷌ai(x)=0ˈ䛓Ѹq0mq0+wi
ྲ᷌ai(x)=1ˈ䛓Ѹq1mq1+wi
ྲ᷌q1>q0ˈ䛓Ѹ亴⍻c(x)=1
ྲ᷌q0>q1ˈ䛓Ѹ亴⍻c(x)=0
ྲ᷌q1=q0ˈ䛓Ѹሩc(x)䲿ᵪ亴⍻ 0 ᡆ 1
ሩAѝ⇿њ亴⍻㇇⌅aiˈ‫˖ڊ‬
ྲ᷌ai(x) zc(x)ˈ䛓ѸwimEwi
⧠൘䇱᰾ˈ࣐ᵳཊᮠ㇇⌅Ⲵࠪ䭉ᮠ䟿䗩⭼ਟԕ⭡ᣅ⾘⊐ѝᴰ֣亴⍻㇇⌅Ⲵࠪ䭉ᮠᶕ㺘
⽪Ǆ
ᇊ⨶ 7.5 ࣐ᵳཊᮠ㇇⌅Ⲵ⴨ሩ䈟ᐞ⭼䲀ǄԔ D Ѫԫ᜿Ⲵ䇝㓳ṧֻᒿࡇˈԔ A Ѫԫ᜿ n њ亴⍻㇇

⌅Ⲵ䳶ਸˈԔ k Ѫ A ѝԫ᜿㇇⌅ሩṧֻᒿࡇ D Ⲵࠪ䭉⅑ᮠⲴᴰሿ٬Ǆ䛓Ѹ֯⭘£=1/2 Ⲵ࣐ᵳཊᮠ㇇
⌅൘ D кࠪ䭉⅑ᮠᴰཊѪ˖
2.4(k+log2n)
䇱᰾˖ሩᇊ⨶Ⲵ䇱᰾ਟ䙊䗷∄䖳ᴰ֣亴⍻㇇⌅Ⲵᴰ㓸ᵳ䟽઼ᡰᴹ㇇⌅Ⲵᵳ䟽ѻ઼ǄԔajԓ㺘Aѝ
а㇇⌅ˈᒦфᆳࠪ䭉Ⲵ⅑ᮠѪᴰՈⲴk⅑Ǆоaj⴨㚄㌫Ⲵᵳ䟽wjሶѪ(1/2)kˈഐѪᆳⲴࡍ࿻ᵳ䟽Ѫ 1ˈᒦ
¦
n
൘⇿⅑ࠪ䭉ᰦ҈ԕ 1/2Ǆ⧠൘㘳㲁Aѝᡰᴹnњ㇇⌅Ⲵᵳ䟽઼W=
i 1
wi ǄWࡍ࿻ѪnǄሩ࣐ᵳཊᮠ㇇
⌅Ⲵ⇿⅑ࠪ䭉ˈW㻛߿ሿѪᴰཊ 3/4 WǄަ৏ഐᱟ࣐ᵳᣅ⾘ঐᴹཊᮠⲴ㇇⌅ᴰቁᤕᴹᮤњᵳ䟽WⲴаॺ

٬ˈ㘼䘉а䜘࠶ሶ㻛҈ԕഐᆀ 1/2ǄԔMԓ㺘࣐ᵳཊᮠ㇇⌅ሩ䇝㓳ᒿࡇDⲴᙫࠪ䭉⅑ᮠˈଚѸᴰ㓸Ⲵᙫ
M
§3·
ᵳ䟽WᴰཊѪ n¨ ¸ ǄഐѪᴰ㓸Ⲵᵳ䟽wjнՊ∄ᴰ㓸ᙫᵳ䟽བྷˈഐ↔ᴹ˖
©4¹
k M
§1· §3·
¨ ¸ d n¨ ¸
©2¹ ©4¹
䟽ᯠᆹᧂ਴亩ᗇࡠ˖
(k log 2 n)
Md d 2.4(k log 2 n)
§3·
log 2 ¨ ¸
©4¹
ᇊ⨶ᗇ䇱Ǆ
ᾲᤜൠ䈤ˈк䶒Ⲵᇊ⨶䈤᰾࣐ᵳཊᮠ㇇⌅Ⲵࠪ䭉ᮠ䟿нՊབྷҾ㇇⌅⊐ѝᴰ֣㇇⌅ࠪ䭉ᮠ
䟿ˈ࣐ка䲿⵰㇇⌅⊐བྷሿሩᮠ໎䮯Ⲵ亩ˈ޽҈ԕаᑨᮠഐᆀǄ
䈕ᇊ⨶⭡ Littlestone Warmuth˄˅䘋а↕а㡜ॆˈ䇱᰾Ҷሩԫ᜿ İ£ˈк䘠

䗩⭼Ѫ˖
1
k log 2 log 2 n
E

2
log 2
1 E
x ਟ㜭䘁լ↓⺞⁑ර˄PAC˅䪸ሩⲴ㇇⌅ӾḀᾲᘥ㊫ C ѝᆖҐⴞḷᾲᘥˈ֯⭘᤹
аᵚ⸕նപᇊⲴᾲᘥ࠶ᐳѝ䲿ᵪᣭਆⲴ䇝㓳ṧֻǄᆳ㾱≲ᆖҐಘਟ㜭˄ԕ㠣ቁ
1-¥Ⲵᾲ⦷˅ᆖҐࡠа䘁լ↓⺞˄䭉䈟⦷ሿҾ¦˅Ⲵ‫ٷ‬䇮ˈ㘼䇑㇇䟿઼䇝㓳ṧֻ
ᮠ䜭ਚ䲿⵰ 1/¥ˈ1/¦ˈᇎֻ䮯ᓖ઼ⴞḷᾲᘥ䮯ᓖⲴཊ亩ᔿ㓗㓯ᙗ໎䮯Ǆ
x ൘ PAC ᆖҐ⁑රⲴṶᷦлˈԫօ֯⭘аᴹ䲀‫ٷ‬䇮オ䰤 H˄ަѝ CH˅Ⲵа㠤ᆖ
Ґಘˈሶԕᾲ⦷ 1-¥䗃ࠪањⴞḷᾲᘥѝ䈟ᐞ൘¦㤳ത޵Ⲵ‫ٷ‬䇮ˈᡰ䴰䲿ᵪᣭ
ਆ䇝㓳ṧֻᮠⴞѪ mˈф m ┑䏣
1
mt (ln(1 / G ) ln H )
H
䈕ᔿ㔉ࠪҶ PAC ⁑රлᡀ࣏ⲴᆖҐᡰ䴰Ⲵ䏣ཏⲴ䇝㓳ṧֻᮠⴞⲴ䗩⭼Ǆ
x PAC ᆖҐ⁑රⲴањᴹ㓖ᶏⲴ‫ٷ‬ᇊᱟˈᆖҐಘ亴‫⸕ݸ‬䚃Ḁਇ䲀Ⲵᾲᘥ㊫ Cˈᆳ
वਜ਼㾱ᆖҐⲴⴞḷᾲᘥǄ⴨৽ˈнਟ⸕ᆖҐ˄agnostic learning˅㘳㲁ᴤа㡜Ⲵ䰞
仈ṶᷦˈަѝᆖҐಘн‫ٷ‬ᇊⴞḷᾲᘥᡰ൘Ⲵ㊫࡛ǄᆖҐಘਚӾ䇝㓳ᮠᦞѝ䗃ࠪ H
ѝᴹᴰሿ䈟ᐞ⦷˄ਟ㜭䶎 0˅Ⲵ‫ٷ‬䇮Ǆ൘䘉њਇ䲀䖳ቁⲴнਟ⸕ᆖҐ⁑රѝˈᆖ
Ґಘ‫؍‬䇱ԕᾲ⦷ 1-¥Ӿ H ѝᴰਟ㜭ᴹ‫ٷ‬䇮ѝ䗃ࠪ䭉䈟⦷ሿҾ¦Ⲵ‫ٷ‬䇮ˈ㾱㿲ሏ
Ⲵ䲿ᵪᣭਆ䇝㓳ṧֻᮠⴞ m ┑䏣˖
1
mt (ln(1 / G ) ln H )
2H 2
x ᡀ࣏ⲴᆖҐᡰ䴰Ⲵ䇝㓳ֻᮠⴞᖸᕪ⛸ൠਇࡠᆖҐಘᡰ㘳㲁Ⲵ‫ٷ‬䇮オ䰤༽ᵲᓖⲴ
ᖡ૽ǄሩҾ‫ٷ‬䇮オ䰤 H ༽ᵲᓖⲴањᴹ⭘Ⲵᓖ䟿ᱟ VC 㔤ˈVC(H)ǄVC(H)ᱟਟ
㻛 H ᣶ᮓ˄ԕᡰᴹਟ㜭ᯩᔿ࠶ࢢ˅ᴰབྷᇎֻᆀ䳶ⲴབྷሿǄ
x ൘ PAC ⁑රлԕ VC(H)㺘⽪Ⲵ䏣ԕሬ㠤ᡀ࣏ᆖҐⲴ䇝㓳ṧֻᮠⴞⲴк⭼Ѫ
1
mt (4 log 2 (2 / G ) 8VC ( H ) log 2 (13 / H ))
H
л⭼Ѫ
ª1 VC (C ) 1º
m t max « log(1 / G ),
¬H 32H »¼
x ਖа⿽ᆖҐ⁑ᔿ〠Ѫࠪ䭉⭼䲀⁑ᔿˈᆳ⭘Ҿ࠶᷀ањᆖҐಘ൘⺞࠷ᆖҐࡠⴞḷ
ᾲᘥѻࡽՊӗ⭏Ⲵ䈟࠶㊫⅑ᮠǄֻྲˈHalving㇇⌅൘ᆖҐࡠHѝⲴԫ᜿ⴞḷᾲ
ᘥࡽՊᴹ㠣ཊ¬log2|H|¼⅑ࠪ䭉Ǆሩԫ᜿ᾲᘥ㊫Cˈᴰൿᛵߥлᴰ֣㇇⌅ሶᴹOpt(C)
⅑ࠪ䭉ˈަѝ˖
VC(C)İOpt(C)İ log2(|C|)
x ࣐ᵳཊᮠ㇇⌅㔃ਸҶཊњ亴⍻㇇⌅Ⲵ࣐ᵳᣅ⾘ᶕ࠶㊫ᯠⲴᇎֻǄᆳสҾ䘉Ӌ亴
⍻㇇⌅൘ṧֻᒿࡇѝⲴࠪ䭉ᶕᆖҐ⇿њ㇇⌅Ⲵᵳ٬Ǆᴹ䏓Ⲵᱟˈ࣐ᵳཊᮠ㇇⌅
ӗ⭏Ⲵ䭉䈟⭼䲀ਟ⭘㇇⌅⊐ѝᴰ֣亴⍻㇇⌅Ⲵࠪ䭉ᮠᶕ䇑㇇Ǆ
䇑㇇ᆖҐ⨶䇪ѝ䇨ཊᰙᵏⲴᐕ֌䪸ሩⲴ䰞仈ᱟˈ൘㔉ᇊањнᇊ䮯Ⲵ䇝㓳ṧֻᒿֻᰦˈ
ᆖҐಘᱟ੖㜭൘ᶱ䲀ᰦ⺞ᇊⴞḷᾲᘥǄ൘ᶱ䲀⁑රлⲴ⺞ᇊᯩ⌅⭡ Gold㔉ࠪǄ‫ޣ‬Ҿ
↔亶ฏⲴањྭⲴ㔬䘠㿱AngluinǄVapnik䈖㓶㘳ḕҶа㠤᭦ᮋ˄uniform
convergence˅Ⲵ䰞仈ˈ㘼ᇶ࠷⴨‫ Ⲵޣ‬PAC ᆖҐ⁑ර⭡ ValiantᨀࠪǄᵜㄐѝ¦䈖ቭਈ
රオ䰤Ⲵ䇘䇪สҾ HausslerⲴ䱀䘠Ǆ൘ PAC ⁑රлⲴа㓴ᴹ⭘Ⲵ㔃䇪ਟ൘ Blueretal
˄˅ѝ᢮ࡠǄKearns Vaziraniᨀ‫׋‬Ҷ䇑㇇ᆖҐ⨶䇪ѝ䇨ཊ㔃䇪ⲴањՈ⿰Ⲵ䱀
䘠Ǆ↔亶ฏаӋᰙᵏⲴ᮷ㄐवᤜ Anthsny Biggs઼ NatarajanǄ
ⴞࡽ䇑㇇ᆖҐ⨶䇪Ⲵ⹄ウ㾶ⴆҶ䇨䇨ཊཊⲴᆖҐ⁑ර઼ᆖҐ㇇⌅Ǆ䇨ཊ䘉ᯩ䶒Ⲵ⹄ウਟ
ԕ൘䇑㇇ᆖҐ⨶䇪˄COLT˅ⲴᒤᓖՊ䇞Ⲵ䇪᮷䳶ѝ᢮ࡠǄᵏ࠺ǉᵪಘᆖҐǊMachineLearning
ѝаӋ⢩↺Ⲵḿⴞҏ⎹৺䘉аѫ仈Ǆ
Ґ仈
㘳㲁䇝㓳ањє䗃‫ޕ‬ᝏ⸕ಘǄ㔉ࠪ䇝㓳ṧֻᮠⴞⲴк⭼ˈԕ‫؍‬䇱ᆖҐࡠⲴᝏ⸕ಘᴹ
Ⲵ㖞ؑᓖˈަⵏᇎ䭉䈟⦷н䎵䗷 Ǆ䘉а䗩⭼ᱟ੖ᇎ䱵˛
㘳㲁ᾲᘥ㊫ C ⲴᖒᔿѪ˄aİxİb˅ġ˄cİyİd˅ަѝ abcd Ѫ४䰤˄˅

䰤ⲴᮤᮠǄ⌘᜿䈕㊫ѝⲴ⇿њᾲᘥሩᓄањ⸙ᖒˈᆳⲴ䗩⭼ᱟ xy ᒣ䶒Ⲵа䜘࠶кⲴᮤᮠ٬Ǆ
ᨀ⽪˖㔉ᇊањ䈕ᒣ䶒кⲴ४䰤ˈަ䗩⭼Ѫ⛩઼nnǄ൘↔४䰤޵н਼Ⲵᇎ䗩
2
§ n(n 1) ·
⭼⸙ᖒⲴᮠ䟿Ѫ ¨ ¸ Ǆ
© 2 ¹
a㔉ࠪ䲿ᵪᣭਆ䇝㓳ṧֻⲴᮠ䟿Ⲵк⭼ˈ֯䏣ԕ‫؍‬䇱ሩ C ѝԫ᜿ⴞḷᾲᘥ cˈԫа֯

⭘ H C ⲴᆖҐಘሶԕ Ⲵᾲ⦷䗃ࠪањ䭉䈟⦷ᴰཊѪ Ⲵ‫ٷ‬䇮Ǆ
b⧠‫ٷ‬ᇊ⸙ᖒ䗩⭼ abcd ਆᇎᮠ٬Ǆ䟽ᯠഎㆄㅜањ䰞仈Ǆ
൘ᵜㄐѝᡁԜ᧘ሬҶ䇝㓳ṧֻᮠ䟿Ⲵ㺘䗮ᔿˈ֯䏣ԕ‫؍‬䇱⇿њ‫ٷ‬䇮Ⲵⵏᇎ䭉䈟⦷н
ՊᐞҾަ㿲ሏࡠⲴ䇝㓳䭉䈟⦷errorDh࣐к¦Ǆ⢩࡛ൠˈᡁԜ֯⭘ҶHoeffding⭼䲀ᶕ᧘ሬ
ᔿ Ǆ䈅᧘ሬ䇝㓳ṧֻᮠⴞⲴਖа㺘䗮ᔿˈ֯䏣ԕ‫؍‬䇱⇿њ‫ٷ‬䇮Ⲵⵏᇎ䭉䈟⦷нՊᐞҾ
¤errorDhǄ᧘ሬⲴ䗷〻ਟ֯⭘л䶒Ⲵ䙊⭘ⲴChernoff⭼䲀Ǆ
Chernoff⭼䲀˖‫ٷ‬ᇊxĂxmˈѪmњ⤜・⺜ᐱᣅ᧧˄Bernonlliᇎ傼˅Ⲵ䗃ࠪˈަѝ⇿
⅑ᇎ傼↓䶒Ⲵᾲ⦷ѪPr>Xi @ pˈ㘼৽䶒ᾲ⦷Pr>Xi @ pǄᇊѹS XXĂXmѪ䘉m⅑ᇎ
傼䗃ࠪⲴ઼ǄSmⲴᵏᵋ٬ѪE>Sm@ pǄChernoff⭼䲀᧿䘠ҶSmԕḀഐᆀ İ¤İ н਼Ҿ
pⲴᾲ⦷˖
Pr[ S / m ! (1 J ) p ] d e mpJ
2
/3

Pr[ S / m ! (1 J ) p ] d e mpJ
2
/2

㘳㲁аᆖҐ䰞仈ˈަѝ X Ѫᇎᮠ䳶ਸˈᒦф C H ѪᇎᮠкⲴ४䰤䳶ਸˈ

H ^axb_abę`Ǆ㤕а‫ٷ‬䇮о↔ⴞḷᾲᘥⲴ m њṧֻа㠤ˈ䛓Ѹᆳ䭉䈟⦷㠣ቁѪ¦
Ⲵᾲ⦷ᱟཊቁ˛֯⭘ VC 㔤䀓ߣ↔䰞仈Ǆᱟ੖㜭᢮ࡠਖа⿽ᯩ⌅สҾᴰสᵜⲴ৏⨶ᒦфн⭘
VC 㔤ᶕ䀓ߣ↔䰞仈˛
㘳㲁ሩᓄ xy ᒣ䶒кᡰᴹ⛩Ⲵᇎֻオ䰤 Xˈ㔉ࠪлࡇ‫ٷ‬䇮オ䰤Ⲵ VC 㔤˖
aHr xyᒣ䶒кᡰᴹ⸙ᖒⲴ䳶ਸǄণH {((a x b) (c y d )) | a, b, c, d }
bHc xyᒣ䶒ⲴശǄ൘ശ޵Ⲵ⛩㻛࠶㊫Ѫ↓ֻǄ
cHt xyᒣ䶒޵Ⲵй䀂ᖒǄ൘й䀂ᖒ޵Ⲵ⛩㻛࠶㊫Ѫ↓ֻǄ
߉ࠪҐ仈 ѝሩHrⲴања㠤ᆖҐಘǄ䲿ᵪ⭏ᡀа㓴н਼Ⲵⴞḷᾲᘥˈሩᓄᒣ䶒
кн਼Ⲵ⸙ᖒǄѪ⇿ањⴞḷᾲᘥ䲿ᵪ⭏ᡀṧֻˈަѝⲴᇎֻ࠶ᐳѪ⸙ᖒ!ࡠ!
޵Ⲵ൷र࠶ᐳǄ൘മк⭫ࠪሩᓄ䇝㓳ṧֻᮠⴞⲴmⲴ⌋ॆ䭉䈟⦷Ǆ൘਼аമкˈ⭫ࠪ¥
ᰦ¦઼mѻ䰤⨶䇪кⲴ‫ޣ‬㌫ᴢ㓯Ǆ䈕⨶䇪ᱟ੖оᇎ傼⴨ㅖਸ˛
㘳㲁‫ٷ‬䇮㊫HrdѪnњᐳቄਈ䟿кⲴĀ㿴ࡉⲴˈ␡ᓖ ⲴߣㆆṁāǄ䘉ṧⲴߣㆆṁᱟ

ᤷ␡ᓖѪ ˄ণᴹഋњਦ㔃⛩ˈоṩⲴ⸙⿫䜭Ѫ ˅ˈфṩⲴᐖᆀ㔃⛩઼ਣᆀ㔃⛩㾱≲वਜ਼਼
ṧⲴਈ䟿Ǆֻྲˈл䶒ⲴṁѪHrdѝⲴањᇎֻǄ
aԕnⲴ࠭ᮠᖒᔿ㺘⽪ࠪHrd2ѝᴹཊቁ䈝⌅н਼ⲴṁǄ
(b)㔉ࠪPAC⁑රлᡰ䴰Ⲵṧֻᮠⴞк⭼ˈ֯ᆖҐࡠHrd2䭉䈟⦷Ѫ¦ˈ㖞ؑᓖѪ¥Ǆ
(c)㘳㲁л䶒ⲴሩHrd2㊫Ⲵ࣐ᵳཊᮠ㇇⌅Ǆᔰ࿻ˈHrd2ѝᡰᴹ‫ٷ‬䇮ࡍ࿻ᵳ٬䜭Ѫ Ǆ⇿⅑
䙷ࡠᯠṧֻˈ㾱สҾHrd2ѝᡰᴹ‫ٷ‬䇮Ⲵ࣐ᵳᣅ⾘䘋㹼亴⍻Ǆ❦ਾˈнᱟ⎸䲔ᦹна㠤Ⲵṁˈ
㘼ᱟሶᆳԜⲴᵳ٬ԕഐᆀ 䘋㹼ࡺ߿Ǆ↔䗷〻ᴰཊՊᴹཊቁ⅑ࠪ䭉˛ԕHrd2ѝᴰ֣ṁⲴࠪ䭉
ᮠ઼nᶕ㺘⽪Ǆ
7.8 ᵜ䰞仈ѫ㾱㘳㲁㾱ᵜㄐѝⲴ PAC ࠶઼᷀ㅜ 5 ㄐ䇘䇪Ⲵ‫ٷ‬䇮䇴ՠѻ䰤Ⲵ㚄㌫Ǆ㘳㲁а
ᆖҐԫ࣑ˈަѝᇎֻ䜭⭡ n њᐳቄਈ䟿᧿䘠˄ྲ˖ x1 x 2 x3 ... x n ˅ˈᒦфަᣭਆ᤹➗
Ḁപᇊնᵚ⸕Ⲵᾲ⦷࠶ᐳ DǄⴞḷᾲᘥᐢ⸕ਟ⭡ᐳቄ኎ᙗᡆᆳԜⲴ੖ᇊⲴਸਆᶕ㺘⽪˄ྲ
x 2 x5 ˅ˈᒦфᆖҐ㇇⌅֯⭘䈕ᾲᘥ㊫֌ѪᆳⲴ‫ٷ‬䇮オ䰤 HǄања㠤ᆖҐಘ㻛㔉Ҹ 100
њ᤹ D ᣭਆⲴ䇝㓳ṧֻǄᆳӾ H ѝ䗃ࠪањ‫ٷ‬䇮 hˈᱟоᡰᴹ 100 њṧֻа㠤Ⲵ˄ণ൘䘉

Ӌ䇝㓳ṧֻк h Ⲵ䭉䈟⦷Ѫ 0Ǆ˅
(a)ᡁԜᝏ‫ޤ‬䏓Ⲵᱟ h Ⲵⵏᇎ䭉䈟⦷ˈণሶᶕ᤹ D ᣭਆⲴᇎֻ㻛䈟࠶㊫Ⲵᾲ⦷ᱟཊቁǄ

สҾк䶒Ⲵؑ᚟ˈ㜭੖㔉ࠪањ४䰤ˈ֯ⵏᇎ䭉䈟⦷㩭‫ަޕ‬ѝⲴᾲ⦷㠣ቁѪ 95%˛ྲ᷌㜭ˈ
䈧᧿䘠䈕४䰤ᒦㆰ䘠⨶⭡Ǆ੖ࡉˈ䀓䟺ഠ䳮ᡰ൘Ǆ
(b)⧠൘ᣭਆ 100 њᯠⲴᇎֻˈᣭਆ᤹➗࠶ᐳ D ᒦ⴨ӂ⤜・Ǆ㔃᷌ਁ⧠ h ሶ 100 њᯠṧ

ֻѝⲴ 30 њ䈟࠶㊫ҶǄ㜭੖㔉ࠪањ४䰤֯ⵏᇎ䭉䈟⦷㩭‫ަޕ‬ѝᾲ⦷㓖Ѫ 95%˛˄൘䘉䟼
ᘭ⮕ԕࡽሩ䇝㓳ᮠᦞⲴᙗ㜭Ǆ˅ྲ᷌㜭ཏˈ䈧᧿䘠䈕४䰤ᒦㆰ䘠⨶⭡Ǆ੖ࡉ䀓䟺ഠ䳮ᡰ൘Ǆ
(c)ণ֯ h 㜭ཏᆼழൠ࠶㊫䇝㓳ṧֻˈᆳӽ❦ᢺᯠṧֻ 30%䈟࠶㊫ҶǄࡔᯝ䘉⿽ᛵߥᱟ

ሩ䖳བྷⲴ n 䘈ᱟ䖳ሿⲴ n ᴤᴹਟ㜭ࠪ⧠Ǆ⭘аਕ䈍䈤᰾֐ⲴഎㆄⲴ⨶⭡Ǆ
ㅢ8ㄖ ะӄᇔׁⲺᆜҖ
已知一系列的训练样例，很多学习方法为目标函数建立起明确的一般化描述；
但与此不同，基于实例的学习方法只是简单地把训练样例存储起来。从这些实例中
泛化的工作被推迟到必须分类新的实例时。每当学习器遇到一个新的查询实例，它
分析这个新实例与以前存储的实例的关系，并据此把一个目标函数值赋给新实例。
基于实例的学习方法包括最近邻（nearest neighbor）法和局部加权回归（locally
weighted regression）法，它们都假定实例可以被表示为欧氏空间中的点。基于实
例的学习方法还包括基于案例的推理（case-based reasoning），它对实例采用更
复杂的符号表示。基于实例的学习方法有时被称为消极（lazy）学习法，因为它们
把处理工作延迟到必须分类新的实例时。这种延迟的或消极的学习方法有一个关键
的优点，即它们不是在整个实例空间上一次性地估计目标函数，而是针对每个待分
类新实例作出局部的和相异的估计。
8.1 ㆶԁ
สҾᇎֻⲴᆖҐᯩ⌅ѝˈᴰ䘁䛫⌅઼ተ䜘࣐ᵳഎᖂ⌅⭘Ҿ䙬䘁ᇎ٬ᡆ⿫ᮓⴞḷ࠭ᮠˈᆳ
Ԝ൘ᾲᘥк䜭ᖸㆰ᰾ǄሩҾ䘉Ӌ㇇⌅ˈᆖҐ䗷〻ਚᱟㆰঅൠᆈۘᐢ⸕Ⲵ䇝㓳ᮠᦞǄᖃ䙷ࡠᯠ
Ⲵḕ䈒ᇎֻᰦˈа㌫ࡇ⴨լⲴᇎֻ㻛Ӿᆈۘಘѝਆࠪˈᒦ⭘ᶕ࠶㊫ᯠⲴḕ䈒ᇎֻǄ䘉Ӌᯩ⌅
оަԆㄐ䇘䇪Ⲵᯩ⌅⴨∄ˈањ‫ޣ‬䭞ᐞᔲᱟ˖สҾᇎֻⲴᯩ⌅ਟԕѪн਼Ⲵᖵ࠶㊫ḕ䈒ᇎֻ
ᔪ・н਼Ⲵⴞḷ࠭ᮠ䙬䘁Ǆһᇎкˈᖸཊᢰᵟਚᔪ・ⴞḷ࠭ᮠⲴተ䜘䙬䘁ˈሶަᓄ⭘Ҿоᯠ
ḕ䈒ᇎֻ䛫䘁Ⲵᇎֻˈ㘼Ӿнᔪ・൘ᮤњᇎֻオ䰤к䜭㺘⧠㢟ྭⲴ䙬䘁Ǆᖃⴞḷ࠭ᮠᖸ༽ᵲˈ
նᆳਟ⭘нཚ༽ᵲⲴተ䜘䙬䘁᧿䘠ᰦˈ䘉ṧ‫ڊ‬ᴹᱮ㪇ⲴՈ࣯Ǆ
สҾᇎֻⲴᯩ⌅ҏਟԕ֯⭘ᴤ༽ᵲⲴㅖਧ㺘⽪⌅ᶕ᧿䘠ᇎֻǄ൘สҾṸֻⲴᆖҐѝˈᇎ
ֻণԕ䘉⿽ᯩᔿ㺘⽪ˈ㘼фҏ᤹➗䘉⿽ᯩᔿᶕ⺞ᇊ䛫䘁ᇎֻǄสҾṸֻⲴ᧘⨶ᐢ㓿㻛ᓄ⭘ࡠ
ᖸཊԫ࣑ѝˈ∄ྲˈ൘૘䈒ਠкᆈ઼ۘ༽⭘䗷৫Ⲵ㓿傼˗ṩᦞԕࡽⲴ⌅ᖻṸԦ䘋㹼᧘⨶˗䙊
䗷༽⭘ԕࡽ≲䀓Ⲵ䰞仈Ⲵ⴨‫ޣ‬䜘࠶ᶕ䀓ߣ༽ᵲⲴ䈳ᓖ䰞仈Ǆ
สҾᇎֻᯩ⌅Ⲵањн䏣ᱟˈ࠶㊫ᯠᇎֻⲴᔰ䬰ਟ㜭ᖸབྷǄ䘉ᱟഐѪࠐѾᡰᴹⲴ䇑㇇䜭
ਁ⭏൘࠶㊫ᰦˈ㘼нᱟ൘ㅜа⅑䙷ࡠ䇝㓳ṧֻᰦǄᡰԕˈྲօᴹ᭸ൠ㍒ᕅ䇝㓳ṧֻˈԕ߿ቁ
ḕ䈒ᰦᡰ䴰䇑㇇ᱟањ䟽㾱Ⲵᇎ䐥䰞仈Ǆ↔㊫ᯩ⌅ⲴㅜҼњн䏣ᱟ˄ቔަሩҾᴰ䘁䛫⌅˅ˈ
ᖃӾᆈۘಘѝỰ㍒⴨լⲴ䇝㓳ṧֻᰦˈᆳԜа㡜㘳㲁ᇎֻⲴᡰᴹ኎ᙗǄྲ᷌ⴞḷᾲᘥӵ‫׍‬䎆
Ҿᖸཊ኎ᙗѝⲴࠐњᰦˈ䛓Ѹⵏ↓ᴰĀ⴨լāⲴᇎֻѻ䰤ᖸਟ㜭⴨䐍⭊䘌Ǆ
൘ла㢲ᡁԜሶӻ㓽 k-䘁䛫˄k-Nearest Neighbor˅⌅ˈԕ৺䘉њᒯ⌋ᓄ⭘Ⲵᯩ⌅Ⲵࠐњ

ਈփǄ൘↔ѻਾᡁԜሶ䇘䇪ተ䜘࣐ᵳഎᖂ⌅ˈа⿽ᔪ・ⴞḷ࠭ᮠⲴተ䜘䙬䘁ⲴᆖҐᯩ⌅ˈ䘉
⿽ᯩ⌅ਟԕ㻛ⴻ֌ k-䘁䛫⌅Ⲵа㡜ᖒᔿǄ❦ਾᡁԜ᧿䘠ᖴੁส࠭ᮠ˄radial basis function˅
㖁㔌ˈ䘉⿽㖁㔌ѪสҾᇎֻⲴᆖҐ㇇⌅઼⾎㓿㖁㔌ᆖҐ㇇⌅ᨀ‫׋‬Ҷањᴹ䏓ⲴẕằǄ޽ла
㢲䇘䇪สҾṸֻⲴ᧘⨶ˈа⿽֯⭘ㅖਧ㺘⽪઼สҾ⸕䇶Ⲵ᧘⨶˄knowledge-based inference˅
Ⲵᯩ⌅Ǆ䘉а㢲वᤜҶањสҾṸֻⲴ᧘⨶ᓄ⭘ᇎֻˈ⭘Ҿ䀓ߣᐕ〻䇮䇑䰞仈ǄᴰਾˈᡁԜ
䇘䇪Ҷᵜㄐ䇢䘠Ⲵ⎸ᶱᆖҐᯩ⌅઼ᵜҖަԆ਴ㄐⲴ〟ᶱ˄eager˅ᆖҐᯩ⌅䰤ⲴᐞᔲǄ
8.2 k-䘇䛱⌋
สҾᇎֻⲴᆖҐᯩ⌅ѝᴰสᵜⲴᱟk-䘁䛫㇇⌅Ǆ䘉њ㇇⌅‫ٷ‬ᇊᡰᴹⲴᇎֻሩᓄҾn㔤⅗
∿オ䰤nѝⲴ⛩ǄањᇎֻⲴᴰ䘁䛫ᱟṩᦞḷ߶⅗∿䐍⿫ᇊѹⲴǄᴤ㋮⺞ൠ䇢ˈᢺԫ᜿Ⲵ
ᇎֻx㺘⽪Ѫл䶒Ⲵ⢩ᖱੁ䟿˖
<a1(x)ˈa2(x)ˈ}an(x)>
ަѝar(x)㺘⽪ᇎֻxⲴㅜrњ኎ᙗ٬Ǆ䛓Ѹєњᇎֻxi઼xj䰤Ⲵ䐍⿫ᇊѹѪd(xi, xj)ˈަѝ˖
n
d(xi, xj){ ¦ (a ( x ) a ( x
r 1
r i r j )) 2
൘ᴰ䘁䛫ᆖҐѝˈⴞḷ࠭ᮠ٬ਟԕѪ⿫ᮓ٬ҏਟԕѪᇎ٬ǄᡁԜ‫ݸ‬㘳㲁ᆖҐԕлᖒᔿⲴ
⿫ᮓⴞḷ࠭ᮠf : noVǄަѝVᱟᴹ䲀䳶ਸ{v1, }vs}Ǆ㺘 8-1 㔉ࠪҶ䙬䘁⿫ᮓⴞḷ࠭ᮠⲴk-
䘁䛫㇇⌅Ǆ↓ྲ㺘ѝᡰᤷࠪⲴˈ䘉њ㇇⌅Ⲵ䘄എ٬ fˆ (xq)Ѫሩf(xq)Ⲵՠ䇑ˈᆳቡᱟ䐍⿫xqᴰ
䘁Ⲵkњ䇝㓳ṧֻѝᴰᲞ䙽Ⲵf٬Ǆྲ᷌ᡁԜ䘹ᤙk=1ˈ䛓ѸĀ1-䘁䛫㇇⌅āቡᢺf(xi)䍻㔉 fˆ (xq)ˈ
ަѝxiᱟᴰ䶐䘁xqⲴ䇝㓳ᇎֻǄሩҾ䖳བྷⲴk٬ˈ䘉њ㇇⌅䘄എࡽkњᴰ䶐䘁Ⲵ䇝㓳ᇎֻѝᴰ
Პ䙽Ⲵf٬Ǆ
㺘 8-1 䙬䘁⿫ᮓ٬࠭ᮠf : noVⲴk-䘁䛫㇇⌅
䇝㓳㇇⌅˖K-NN(k nearest neighbor learning)⺞ᇎᱟ࠶㊫㇇⌅ˈᱟ᤹➗оᮠᦞ⛩ᴰ᧕䘁

Ⲵ⛩Ⲵᣅ⾘ᶕߣᇊ䈕⛩኎Ҿଚњ㊫Ǆ
zሩҾ⇿њ䇝㓳ṧֻ<x, f(x)>ˈᢺ䘉њṧֻ࣐‫ࡇޕ‬㺘 training_examples
࠶㊫㇇⌅˖
z㔉ᇊањ㾱࠶㊫Ⲵḕ䈒ᇎֻxq
z൘training_examplesѝ䘹ࠪᴰ䶐䘁xqⲴkњᇎֻˈᒦ⭘x1}xk㺘⽪
z䘄എ
k
fˆ ( x q ) m arg max ¦ G (v, f ( xi ))
vV i 1
ަѝྲ᷌ a=b 䛓ѸG(a, b)=1ˈ੖ࡉG(a, b)=0Ǆ
മ 8-1 മ䀓Ҷа⿽ㆰঅᛵߥлⲴk-䘁䛫㇇⌅ˈ൘䘉䟼ᇎֻᱟҼ㔤オ䰤ѝⲴ⛩ˈⴞḷ࠭ᮠ
ާᴹᐳቄ٬Ǆ↓৽䇝㓳ṧֻ⭘Ā+ā઼Ā-ā࠶࡛㺘⽪Ǆമѝҏ⭫ࠪҶањḕ䈒⛩xqǄ⌘᜿൘
䘉ᑵമѝˈ1-䘁䛫㇇⌅ᢺxq࠶㊫Ѫ↓ֻˈ❦㘼 5-䘁䛫㇇⌅ᢺxq࠶㊫Ѫ৽ֻǄ
മ 8-1 k-䘁䛫㇇⌅
ᐖമ⭫ࠪҶа㌫ࡇⲴ↓৽䇝㓳ṧ઼ֻањ㾱࠶㊫Ⲵḕ䈒ᇎֻxqǄ1-䘁䛫㇇⌅ᢺxq࠶㊫Ѫ↓ֻˈ❦㘼
5-䘁䛫㇇⌅ᢺxq࠶㊫Ѫ৽ֻǄਣമᱟሩҾањިරⲴ䇝㓳ṧֻ䳶ਸ 1-䘁䛫㇇⌅ሬ㠤Ⲵߣㆆ䶒Ǆത㔅
⇿њ䇝㓳ṧֻⲴࠨཊ䗩ᖒ㺘⽪ᴰ䶐䘁䘉њ⛩Ⲵᇎֻオ䰤˄ণ䘉њオ䰤ѝⲴᇎֻՊ㻛 1-䘁䛫㇇⌅䍻Ҹ
䈕䇝㓳ṧֻᡰ኎Ⲵ࠶㊫˅Ǆ
k-䘁䛫⌅䳀ਜ਼㘳㲁Ⲵ‫ٷ‬䇮オ䰤HⲴ⢩ᙗᱟӰѸ઒˛⌘᜿k-䘁䛫㇇⌅Ӿᶕнᖒᡀ‫ޣ‬Ҿⴞḷ
࠭ᮠfⲴ᰾⺞Ⲵа㡜‫ٷ‬䇮 fˆ Ǆᆳӵ൘䴰㾱ᰦ䇑㇇⇿њᯠḕ䈒ᇎֻⲴ࠶㊫Ǆ❦㘼ˈᡁԜ‫❦׍‬ਟ
ԕ䰞˖䳀ਜ਼Ⲵа㡜࠭ᮠᱟӰѸ˛ᡆ㘵䈤ˈྲ᷌‫؍‬ᤱ䇝㓳ṧֻнਈˈᒦ⭘XѝⲴ⇿њਟ㜭ᇎֻ
ḕ䈒㇇⌅ˈՊᗇࡠӰѸṧⲴ࠶㊫˛മ 8-1 ѝⲴਣമ⭫ࠪҶ 1-䘁䛫㇇⌅൘ᮤњᇎֻオ䰤кሬ㠤
Ⲵߣㆆ䶒ᖒ⣦Ǆߣㆆ䶒ᱟത㔅⇿њ䇝㓳ṧֻⲴࠨཊ䗩ᖒⲴਸᒦǄሩҾ⇿њ䇝㓳ṧֻˈཊ䗩ᖒ
ᤷࠪҶањḕ䈒⛩䳶ਸˈᆳⲴ࠶㊫ᆼ‫⴨⭡ޘ‬ᓄ䇝㓳ṧֻߣᇊǄ൘䘉њཊ䗩ᖒཆⲴḕ䈒⛩ᴤ᧕
䘁ަԆⲴ䇝㓳ṧֻǄ䘉⿽㊫රⲴമ㓿ᑨ㻛〠Ѫ䘉њ䇝㓳ṧֻ䳶ਸⲴVoronoiമ ķ (Voronoi
diagram)Ǆ
ሩࡽ䶒Ⲵk-䘁䛫㇇⌅֌ㆰঅⲴ‫؞‬᭩ਾˈᆳቡਟ㻛⭘Ҿ䙬䘁䘎㔝٬Ⲵⴞḷ࠭ᮠǄѪҶᇎ⧠
䘉а⛩ˈᡁԜ䇙㇇⌅䇑㇇kњᴰ᧕䘁ṧֻⲴᒣ൷٬ˈ㘼нᱟ䇑㇇ަѝⲴᴰᲞ䙽Ⲵ٬Ǆᴤ㋮⺞
ൠ䇢ˈѪҶ䙬䘁ањᇎ٬ⴞḷ࠭ᮠf:noˈᡁԜਚ㾱ᢺ㇇⌅ѝⲴ‫ޜ‬ᔿᴯᦒѪ˖
¦
k
f ( xi )
fˆ ( x q ) m i 1
˄8.1˅
k
8.2.1 䐓⿱ࣖᵹᴶ䘇䛱㇍⌋
ሩk-䘁䛫㇇⌅Ⲵањᱮ㘼᱃㿱Ⲵ᭩䘋ᱟሩkњ䘁䛫Ⲵ䍑⥞࣐ᵳˈṩᦞᆳԜ⴨ሩḕ䈒⛩xqⲴ
䐍⿫ˈሶ䖳བྷⲴᵳ٬䍻㔉䖳䘁Ⲵ䘁䛫Ǆֻྲˈ൘㺘 8-1 䙬䘁⿫ᮓⴞḷ࠭ᮠⲴ㇇⌅ѝˈᡁԜਟ
ԕṩᦞ⇿њ䘁䛫оxqⲴ䐍⿫ᒣᯩⲴ‫ق‬ᮠ࣐ᵳ䘉њ䘁䛫ⲴĀ䘹ѮᵳāǄᯩ⌅ᱟ䙊䗷⭘лᔿਆԓ
㺘 8-1 ㇇⌅ѝⲴ‫ޜ‬ᔿᶕᇎ⧠˖
k
fˆ ( x q ) m arg max ¦ wi G (v, f ( xi )) ˄8.2˅
vV i 1
1
ަѝ˖ wi {
d ( xq , xi ) 2
˄8.3˅
ķ
䈁⌘˖৸〠ở἞ཊ䗩ᖒ˄Thiessen Polygons˅ˈਟԕ⨶䀓Ѫሩオ䰤Ⲵа⿽࠶ࢢᯩᔿˈањở἞ཊ䗩ᖒ޵Ⲵ
ԫа⛩ࡠᵜở἞ཊਈᖒѝᗳ⛩Ⲵ䐍⿫䜭ሿҾࡠަԆở἞ཊ䗩ᖒѝᗳ⛩Ⲵ䐍⿫Ǆ
ѪҶ༴⨶ḕ䈒⛩xqᚠྭ३䝽Ḁњ䇝㓳ṧֻxiˈӾ㘼ሬ㠤࠶⇽Ѫ 0 ⲴᛵߥˈᡁԜԔ䘉⿽ᛵ
ߥлⲴ fˆ (xq)ㅹҾf(xi)Ǆྲ᷌ᴹཊњ䘉ṧⲴ䇝㓳ṧֻˈᡁԜ֯⭘ᆳԜѝঐཊᮠⲴ࠶㊫Ǆ
ᡁԜҏਟԕ⭘㊫լⲴᯩᔿሩᇎ٬ⴞḷ࠭ᮠ䘋㹼䐍⿫࣐ᵳˈਚ㾱⭘лᔿᴯᦒ㺘 8-1 ѝⲴ‫ޜ‬

ᔿ˖
¦
k
wi f ( xi )
fˆ ( x q ) m i 1
˄8.4˅
¦
k
i
w
1 i
ަѝwiⲴᇊѹо‫ޜ‬ᔿ˄8.3˅ѝ⴨਼Ǆ⌘᜿‫ޜ‬ᔿ˄8.4˅ѝⲴ࠶⇽ᱟањᑨ䟿ˈᆳሶн਼
ᵳ٬Ⲵ䍑⥞ᖂаॆ˄ֻྲˈᆳ‫؍‬䇱ྲ᷌ሩᡰᴹⲴ䇝㓳ṧֻxiˈf(xi)=cˈ䛓Ѹ fˆ (xq)mc˅Ǆ
⌘᜿ԕкk-䘁䛫㇇⌅Ⲵᡰᴹਈփ䜭ਚ㘳㲁kњ䘁䛫ԕ࠶㊫ḕ䈒⛩Ǆྲ᷌֯⭘᤹䐍⿫࣐ᵳˈ
䛓Ѹ‫ݱ‬䇨ᡰᴹⲴ䇝㓳ṧֻᖡ૽xqⲴ࠶㊫һᇎк⋑ᴹൿ༴ˈഐѪ䶎ᑨ䘌Ⲵᇎֻሩ fˆ (xq)Ⲵᖡ૽
ᖸሿǄ㘳㲁ᡰᴹṧֻⲴᜏан䏣ᱟՊ֯࠶㊫䘀㹼ᗇᴤធǄྲ᷌࠶㊫ањᯠⲴḕ䈒ᇎֻᰦ㘳㲁
ᡰᴹⲴ䇝㓳ṧֻˈᡁԜ〠↔Ѫ‫ޘ‬ተ˄global˅⌅Ǆྲ᷌ӵ㘳㲁ᴰ䶐䘁Ⲵ䇝㓳ṧֻˈᡁԜ〠↔
Ѫተ䜘˄local˅⌅Ǆᖃ‫ޜ‬ᔿ˄8.4˅Ⲵ⌅ࡉ㻛ᓄ⭘Ѫ‫ޘ‬ተ⌅ᰦˈᆳ㻛〠ѪShepard⌅˄Shepard
1968˅Ǆ
8.2.2 ሯ k-䘇䛱㇍⌋Ⲻ䈪᱄
᤹䐍⿫࣐ᵳⲴ k-䘁䛫㇇⌅ᱟа⿽䶎ᑨᴹ᭸Ⲵᖂ㓣᧘⨶ᯩ⌅Ǆᆳሩ䇝㓳ᮠᦞѝⲴಚ༠ᴹᖸ
ྭⲴ励ἂᙗˈ㘼фᖃ㔉ᇊ䏣ཏབྷⲴ䇝㓳䳶ਸᰦᆳҏ䶎ᑨᴹ᭸Ǆ⌘᜿䙊䗷ਆ k њ䘁䛫Ⲵ࣐ᵳᒣ
൷ˈਟԕ⎸䲔ᆔ・Ⲵಚ༠ṧֻⲴᖡ૽Ǆ
k-䘁䛫㇇⌅Ⲵᖂ㓣‫ٿ‬㖞(inductive bias)ᱟӰѸ઒˛䙊䗷࠶᷀മ 8-1 ѝⲴ⽪ֻˈਟԕᖸᇩ᱃

ൠ⨶䀓䘉⿽㇇⌅࠶㊫ᯠḕ䈒ᇎֻⲴṩᦞǄᆳⲴᖂ㓣‫ٿ‬㖞ሩᓄҾ‫ٷ‬ᇊ˖ањᇎֻⲴ࠶㊫xqᴰ⴨
լҾ൘⅗∿オ䰤ѝᆳ䱴䘁ⲴᇎֻⲴ࠶㊫Ǆ
ᓄ⭘ k-䘁䛫㇇⌅Ⲵањᇎ䐥䰞仈ᱟˈᇎֻ䰤Ⲵ䐍⿫ᱟṩᦞᇎֻⲴᡰᴹ኎ᙗ˄ҏቡᱟवਜ਼
ᇎֻⲴ⅗∿オ䰤Ⲵᡰᴹ඀ḷ䖤˅䇑㇇ⲴǄ䘉о䛓Ӌਚ䘹ᤙ‫ޘ‬䜘ᇎֻ኎ᙗⲴањᆀ䳶Ⲵᯩ⌅н
਼ˈֻྲߣㆆṁᆖҐ㌫㔏ǄѪҶ⨶䀓䘉⿽ㆆ⮕Ⲵᖡ૽ˈ㘳㲁ᢺ k-䘁䛫㇇⌅ᓄ⭘ࡠ䘉ṧањ䰞
仈˖⇿њᇎֻ⭡ 20 њ኎ᙗ᧿䘠ˈն൘䘉Ӌ኎ᙗѝӵᴹ 2 њоᆳⲴ࠶㊫ᱟᴹ‫ޣ‬Ǆ൘䘉⿽ᛵߥ
лˈ䘉єњ⴨‫ޣ‬኎ᙗⲴ٬а㠤Ⲵᇎֻਟ㜭൘䘉њ 20 㔤Ⲵᇎֻオ䰤ѝ⴨䐍ᖸ䘌Ǆ㔃᷌ˈ‫׍‬䎆
䘉 20 њ኎ᙗⲴ⴨լᙗᓖ䟿Պ䈟ሬ k-䘁䛫㇇⌅Ⲵ࠶㊫Ǆ䘁䛫䰤Ⲵ䐍⿫Պ㻛བྷ䟿Ⲵн⴨‫ޣ‬኎ᙗ
ᡰ᭟䝽Ǆ䘉⿽⭡Ҿᆈ൘ᖸཊн⴨‫ޣ‬኎ᙗᡰሬ㠤Ⲵ䳮仈ˈᴹᰦ㻛〠Ѫ㔤ᓖ⚮䳮˄curse of
dimensionality˅Ǆᴰ䘁䛫ᯩ⌅ሩ䘉њ䰞仈⢩࡛᭿ᝏǄ
䀓ߣ䈕䰞仈Ⲵањᴹ䏓Ⲵᯩ⌅ᱟˈ ˄1˅ᖃ䇑㇇єњᇎֻ䰤Ⲵ䐍⿫ᰦሩ⇿њ኎ᙗ࣐ᵳǄ䘉
⴨ᖃҾ᤹∄ֻ㕙᭮⅗∿オ䰤ѝⲴ඀ḷ䖤ˈ㕙⸝ሩᓄҾнཚ⴨‫ޣ‬኎ᙗⲴ඀ḷ䖤ˈ᣹䮯ሩᓄҾᴤ
⴨‫Ⲵޣ‬኎ᙗⲴ඀ḷ䖤Ǆ⇿њ඀ḷ䖤ᓄըኅⲴᮠ䟿ਟԕ䙊䗷Ӕ৹傼䇱Ⲵᯩ⌅㠚ࣘߣᇊǄާփ‫ڊ‬
⌅ྲлˈ俆‫ٷݸ‬ᇊ֯⭘ഐᆀzjըኅ˄҈˅ㅜjṩ඀ḷ䖤ˈ䘹ᤙzjⲴ਴њ٬z1}znԕ֯ᆖҐ㇇⌅Ⲵ
ⵏᇎ࠶㊫䭉䈟⦷ᴰሿॆǄަ⅑ˈ䘉њⵏᇎ䭉䈟⦷ਟԕ֯⭘Ӕ৹傼䇱ᶕՠ䇑Ǆᡰԕˈа⿽㇇⌅
ᱟ䲿ᵪ䘹ਆ⧠ᴹᮠᦞⲴањᆀ䳶֌Ѫ䇝㓳ṧֻˈ❦ਾߣᇊz1}znⲴ٬࢙֯։ṧֻⲴ࠶㊫䭉䈟
⦷ᴰሿॆǄ䙊䗷ཊ⅑䟽༽䘉њ༴⨶䗷〻ˈਟԕ࣐֯ᵳഐᆀⲴՠ䇑ᴤ࣐߶⺞Ǆ䘉⿽ըኅ඀ḷ䖤
ԕՈॆk-䘁䛫㇇⌅Ⲵ䗷〻ˈᨀ‫׋‬Ҷа⿽ᣁࡦᰐ‫ޣ‬኎ᙗᖡ૽ⲴᵪࡦǄ
˄2˅ਖཆа⿽ᴤᕪᴹ࣋Ⲵᯩ⌅ᱟӾᇎֻオ䰤ѝᆼ‫⎸ޘ‬䲔ᴰн⴨‫Ⲵޣ‬኎ᙗǄ䘉ㅹ᭸Ҿ䇮
㖞Ḁњ㕙᭮ഐᆀzjѪ 0ǄMoore & Lee˄1994˅䇘䇪Ҷᴹ᭸ⲴӔ৹傼䇱ᯩ⌅ˈѪk-䘁䛫㇇⌅䘹
ᤙ⴨‫Ⲵޣ‬኎ᙗᆀ䳶Ǆ⺞࠷ൠ䇢ˈԆԜ᧒㍒ҶสҾĀ⮉а⌅ā˄leave-one-out˅ⲴӔ৹傼䇱ˈ൘
䘉⿽ᯩ⌅ѝˈmњ䇝㓳ᇎֻⲴ䳶ਸԕ਴⿽ਟ㜭ᯩᔿ㻛࠶ᡀm-1 њᇎֻⲴ䇝㓳䳶ਸ઼ 1 њᇎֻ
Ⲵ⍻䈅䳶ਸǄ䘉⿽ᯩ⌅൘k-䘁䛫㇇⌅ѝᱟᇩ᱃ᇎ⧠ⲴˈഐѪ⇿а⅑䟽ᯠᇊѹ䇝㓳䳶ᰦн䴰㾱
仍ཆⲴ䇝㓳ᐕ֌Ǆ⌘᜿к䶒Ⲵє⿽ᯩ⌅䜭ਟԕ㻛ⴻ֌ԕḀњᑨ䟿ഐᆀըኅ඀ḷ䖤Ǆਖཆа⿽
ਟ䘹Ⲵ‫⌅ڊ‬ᱟ֯⭘ањ൘ᇎֻオ䰤кਈॆⲴ٬ըኅ඀ḷ䖤Ǆ䘉ṧ໎࣐Ҷ㇇⌅䟽ᯠᇊѹ䐍⿫ᓖ
䟿Ⲵ㠚⭡ᓖˈ❦㘼ᆳҏ໎࣐Ҷ䗷ᓖᤏਸⲴ仾䲙Ǆᡰԕˈተ䜘ըኅ඀ḷ䖤Ⲵᯩ⌅ᱟнཚᑨ㿱ⲴǄ
ᓄ⭘k-䘁䛫㇇⌅Ⲵਖཆањᇎ䐥䰞仈ᱟྲօᔪ・儈᭸Ⲵ㍒ᕅǄഐѪ䘉њ㇇⌅᧘䘏ᡰᴹⲴ
༴⨶ˈⴤࡠ᧕᭦ࡠањᯠⲴḕ䈒ˈᡰԕ༴⨶⇿њᯠḕ䈒ਟ㜭䴰㾱བྷ䟿Ⲵ䇑㇇Ǆⴞࡽᐢ㓿ᔰਁ
Ҷᖸཊᯩ⌅⭘ᶕሩᆈۘⲴ䇝㓳ṧֻ䘋㹼㍒ᕅˈԕ‫ׯ‬൘໎࣐аᇊᆈۘᔰ䬰ᛵߥлᴤ儈᭸ൠ⺞ᇊ
ᴰ䘁䛫Ǆа⿽㍒ᕅᯩ⌅ᱟkd-tree˄Bentley 1975˗Friedman et al. 1977˅ˈᆳᢺᇎֻᆈۘ൘ṁⲴ
ਦ㔃⛩޵ˈ䛫䘁Ⲵᇎֻᆈۘ൘਼ањᡆ䱴䘁Ⲵ㔃⛩޵Ǆ䙊䗷⍻䈅ᯠḕ䈒xqⲴ䘹ᇊ኎ᙗˈṁⲴ
޵䜘㔃⛩ᢺḕ䈒xqᧂࡇࡠ⴨‫Ⲵޣ‬ਦ㔃⛩Ǆ
8.2.3 ᵥ䈣⌞䀙
൘‫ޣ‬Ҿᴰ䘁䛫⌅઼ተ䜘࣐ᵳഎᖂ⌅Ⲵᖸཊ᮷⥞ѝˈ֯⭘ҶаӋᶕ㠚㔏䇑⁑ᔿ䇶࡛亶ฏⲴ
ᵟ䈝Ǆ൘䰵䈫䘉Ӌ᮷⥞ᰦˈ⸕䚃лࡇᵟ䈝ᱟᴹᑞࣙⲴ˖
z എᖂ˄Regression˅Ⲵਜ਼ѹᱟ䙬䘁ањᇎ٬ⴞḷ࠭ᮠǄ
z ↻ᐞ˄Residual˅ᱟ䙬䘁ⴞḷ࠭ᮠᰦⲴ䈟ᐞ fˆ (x)- f(x)Ǆ
z Ṩ࠭ᮠ˄Kernel function˅ᱟањ䐍⿫࠭ᮠˈᆳ⭘ᶕߣᇊ⇿њ䇝㓳ṧֻⲴ
ᵳ٬Ǆᦒਕ䈍䈤ˈṨ࠭ᮠቡᱟ֯wi=K(d(xi, xq))Ⲵ࠭ᮠKǄ
8.3 ቶ䜞ࣖᵹഔᖈ
ࡽа㢲᧿䘠Ⲵᴰ䘁䛫ᯩ⌅ਟԕ㻛ⴻ֌൘অаⲴḕ䈒⛩x=xqк䙬䘁ⴞḷ࠭ᮠf(x)Ǆተ䜘࣐
ᵳഎᖂᱟ䘉⿽ᯩ⌅Ⲵ᧘ᒯǄᆳ൘⧟㔅xqⲴተ䜘४ฏ޵Ѫⴞḷ࠭ᮠfᔪ・᰾⺞Ⲵ䙬䘁Ǆተ䜘࣐
ᵳഎᖂ֯⭘䱴䘁Ⲵᡆ䐍⿫࣐ᵳⲴ䇝㓳ṧֻᶕᖒᡀ䘉⿽ሩfⲴተ䜘䙬䘁ǄֻྲˈᡁԜਟԕ֯⭘
㓯ᙗ࠭ᮠǃҼ⅑࠭ᮠǃཊቲ⾎㓿㖁㔌ᡆ㘵ަԆ࠭ᮠᖒᔿ൘⧟㔅xqⲴ䛫ฏ޵䙬䘁ⴞḷ࠭ᮠǄ Āተ
䜘࣐ᵳഎᖂā਽〠ѝˈѻᡰԕਛĀተ䜘āᱟഐѪⴞḷ࠭ᮠⲴ䙬䘁ӵӵṩᦞḕ䈒⛩䱴䘁Ⲵᮠᦞˈ
ѻᡰԕਛĀ࣐ᵳāᱟഐѪ⇿ањ䇝㓳ṧֻⲴ䍑⥞ᱟ⭡ᆳоḕ䈒⛩䰤Ⲵ䐍⿫࣐ᵳⲴˈѻᡰԕਛ
ĀഎᖂāᱟഐѪ㔏䇑ᆖҐ⭼ᒯ⌋֯⭘䘉њᵟ䈝ᶕ㺘⽪䙬䘁ᇎᮠ٬࠭ᮠⲴ䰞仈Ǆ
㔉ᇊањᯠⲴḕ䈒ᇎֻxqˈተ䜘࣐ᵳഎᖂⲴа㡜ᯩ⌅ᱟᔪ・ањ䙬䘁 fˆ ˈ֯ fˆ ᤏਸ⧟㔅
xqⲴ䛫ฏ޵Ⲵ䇝㓳ṧֻǄ❦ਾ⭘䘉њ䙬䘁ᶕ䇑㇇ fˆ (xq)Ⲵ٬ˈҏቡᱟѪḕ䈒ᇎֻՠ䇑Ⲵⴞḷ
٬䗃ࠪǄ❦ਾ fˆ Ⲵ᧿䘠㻛ࡐ䲔ˈഐѪሩҾ⇿ањ⤜・Ⲵḕ䈒ᇎֻ䜭Պ䇑㇇н਼Ⲵተ䜘䙬䘁Ǆ
8.3.1 ቶ䜞ࣖᵹ㓵ᙝഔᖈ
л䶒ˈᡁԜ‫ݸ‬㘳㲁ተ䜘࣐ᵳഎᖂⲴа⿽ᛵߥˈণ֯⭘ྲлᖒᔿⲴ㓯ᙗ࠭ᮠᶕ䙬䘁xq䛫ฏ
Ⲵⴞḷ࠭ᮠf˖
fˆ (x)=w0+w1a1(x)+}+wnan(x)
઼ࡽ䶒аṧˈai(x)㺘⽪ᇎֻxⲴㅜiњ኎ᙗ٬Ǆ
എᗶㅜ 4 ㄐѝᡁԜ䇘䇪Ⲵởᓖл䱽ᯩ⌅ˈ൘ᤏਸԕкᖒᔿⲴ㓯ᙗ࠭ᮠࡠ㔉ᇊⲴ䇝㓳䳶ਸ
ᰦˈᆳ㻛⭘ᶕ᢮ࡠ֯䈟ᐞᴰሿॆⲴ㌫ᮠw0}wnǄ൘䛓аㄐѝᡁԜᝏ‫ޤ‬䏓Ⲵᱟⴞḷ࠭ᮠⲴ‫ޘ‬ተ
䙬䘁ǄᡰԕᖃᰦᡁԜ᧘ሬࠪⲴᵳ٬䘹ᤙᯩ⌅ᱟ֯䇝㓳䳶ਸDкⲴ䈟ᐞᒣᯩ઼ᴰሿॆˈণ˖
1
E{ ¦ ( f ( x) fˆ ( x)) 2
2 xD
˄8.5˅
ṩᦞ䘉њ䈟ᐞᇊѹˈᡁԜᗇࠪҶԕлởᓖл䱽䇝㓳⌅ࡉ˖
'wj=K ¦
xD
(f(x)- fˆ (x))aj(x) ˄8.6˅
ަѝKᱟањᑨᮠˈ〠ѪᆖҐ䙏⦷Ǆ㘼ф䘉њ⌅ࡉᐢ㓿㻛䟽ᯠ㺘⽪ˈ‫؞‬᭩Ҷަѝㅜ 4 ㄐ
ѝⲴ䇠ਧԕ३䝽ᖃࡽⲴ䇠ਧ˄ҏቡᱟˈto f(x)ˈoo fˆ (x)ˈxjo aj(x)˅Ǆ
ᡁԜᓄ䈕ྲօ‫؞‬᭩䘉њ䗷〻ᶕ᧘ሬࠪተ䜘䙬䘁઒˛ㆰঅⲴᯩ⌅ᱟ䟽ᯠᇊѹ䈟ᐞ߶ࡉE
ԕ⵰䟽Ҿᤏਸተ䜘䇝㓳ṧֻǄл䶒㔉ࠪҶй⿽ਟ㜭Ⲵ䈟ᐞ߶ࡉǄ⌘᜿ᡁԜᢺ䈟ᐞ߉ѪE(xq)ˈ
ⴞⲴᱟѪҶᕪ䈳ⴞࡽⲴ䈟ᐞ㻛ᇊѹѪḕ䈒⛩xqⲴ࠭ᮠǄ
1. ਚ൘ k њ䘁䛫кⲴ䈟ᐞᒣᯩ઼ᴰሿॆ˖
1
E1 ( xq ) { ¦ ( f ( x) fˆ ( x)) 2
2 xxqⲴkњ䘁䛫
2. ֯ᮤњ䇝㓳ṧֻ䳶ਸDкⲴ䈟ᐞᒣᯩ઼ᴰሿॆˈնሩ⇿њ䇝㓳ṧֻ࣐ᵳˈ
ᵳ٬Ѫ‫ޣ‬Ҿ⴨䐍xq䐍⿫ⲴḀњ䙂߿࠭ᮠK˖
1
E2 ( xq ) { ¦
2 xD
( f ( x) fˆ ( x)) 2 K (d ( x q , x))
3. 㔬ਸ 1 ઼ 2˖
1
E3 ( xq ) { ¦ ( f ( x) fˆ ( x)) 2 K (d ( xq , x))
2 xxqⲴkњ䘁䛫
߶ࡉ 2 ᡆ䇨ᱟᴰԔӪ┑᜿ⲴˈഐѪᆳ‫ݱ‬䇨⇿њ䇝㓳ṧֻ䜭ሩxqⲴ࠶㊫ӗ⭏ᖡ૽Ǆ❦㘼䘉
⿽ᯩ⌅ᡰ䴰Ⲵ䇑㇇䟿䲿⵰䇝㓳ṧֻᮠ䟿㓯ᙗ໎䮯Ǆ߶ࡉ 3 ᖸྭൠ䘁լҶ߶ࡉ 2 ᒦфާᴹྲл
Ո⛩˖䇑㇇ᔰ䬰⤜・Ҿ䇝㓳ṧֻᙫᮠˈ㘼ӵ‫׍‬䎆Ҿᡰ㘳㲁Ⲵᴰ䘁䛫ᮠkǄ
ྲ᷌֯⭘к䶒Ⲵ߶ࡉ 3ˈᒦ֯⭘оㅜ 4 ㄐ⴨਼Ⲵ᧘⨶ᯩᔿ䟽ᯠ᧘ሬởᓖл䱽⌅ࡉˈਟԕ

ᗇࡠԕл䇝㓳⌅ࡉ˖
'wi K ¦ K (d ( x , x))( f ( x) fˆ ( x))a ( x)

xxqⲴkњ䘁䛫
q j ˄8.7˅
////////// wi wi 'wi //////////
⌘᜿䘉њᯠⲴ⌅ࡉ઼‫ޜ‬ᔿ˄8.6˅㔉ࠪⲴ⌅ࡉⲴᐞᔲᱟˈᇎֻxሩᵳ٬ᴤᯠⲴ䍑⥞⧠൘҈
кҶањ䐍⿫᜙㖊亩K(d(xq, x))ˈᒦфӵሩkњᴰ䛫䘁Ⲵ䇝㓳ᇎֻⲴ䈟ᐞ≲઼Ǆһᇎкˈྲ᷌
㾱֯ањ㓯ᙗ࠭ᮠᤏਸപᇊⲴ䇝㓳ṧֻ䳶ਸˈ䛓ѸᴹаӋ∄ởᓖл䱽ᴤ儈᭸Ⲵᯩ⌅ˈᆳԜⴤ
᧕≲䀓ᡰ䴰㾱Ⲵ㌫ᮠw0}wnǄAtkeson et al.˄1997a˅઼Bishop˄1995˅䈳ḕҶࠐњ䘉ṧⲴᯩ
⌅Ǆ
8.3.2 ቶ䜞ࣖᵹഔᖈⲺ䈪᱄
к䶒ᡁԜ㘳㲁Ҷ֯⭘ањ㓯ᙗ࠭ᮠ൘ḕ䈒ᇎֻxq䛫ฏ޵䙬䘁fǄ‫ޣ‬Ҿተ䜘࣐ᵳഎᖂⲴ᮷
⥞ѝˈ൘ሩ䇝㓳ṧֻ䐍⿫࣐ᵳᯩ䶒वਜ਼བྷ䟿Ⲵਟ䘹ᯩ⌅ˈ䘈वਜ਼བྷ䟿Ⲵⴞḷ࠭ᮠተ䜘䙬䘁ᯩ
⌅Ǆ൘བྷཊᮠᛵߥлᱟ䙊䗷ањᑨ䟿ǃ㓯ᙗ࠭ᮠᡆҼ⅑࠭ᮠᶕተ䜘䙬䘁ⴞḷ࠭ᮠǄᴤ༽ᵲⲴ
࠭ᮠᖒᔿнཚᑨ㿱ˈ৏ഐᱟ˄1˅ሩ⇿њḕ䈒ᇎֻ⭘ᴤ༽ᵲⲴ࠭ᮠᶕᤏਸˈަԓԧॱ࠶儈ᰲ˗
˄2˅൘䏣ཏሿⲴᇎֻオ䰤ᆀฏкˈ֯⭘䘉ӋㆰঅⲴ䘁լᐢ㜭⴨ᖃྭൠ⁑ᤏⴞḷ࠭ᮠǄ
8.4 ᖺੇะ࠳ᮦ
ਖа⿽࠭ᮠ䙬䘁Ⲵᯩ⌅ᱟ֯⭘ᖴੁส࠭ᮠ˄radial basis function˅ˈ䘉⿽ᯩ⌅о䐍⿫࣐ᵳ

എᖂ઼Ӫᐕ⾎㓿㖁㔌䜭ᴹ⵰㍗ᇶ㚄㌫˄Powell 1987˗Broomhead & Lowe 1988˗Moody &
Ǆ൘䘉⿽ᯩ⌅ѝˈᖵᆖҐⲴ‫ٷ‬䇮ᱟањԕлᖒᔿⲴ࠭ᮠˈ
Darken 1989˅
k
fˆ ( x) w0 ¦ wu K u (d ( xu , x)) ˄8.8˅
u 1
ަѝ⇿њxuᱟXѝањᇎֻˈṨ࠭ᮠKu(d(xu, x))㻛ᇊѹѪ䲿䐍⿫d(xu, x)Ⲵ໎བྷ㘼߿ሿǄ䘉
䟼Ⲵkᱟ⭘ᡧᨀ‫Ⲵ׋‬ᑨ䟿ˈ⭘ᶕᤷᇊ㾱वਜ਼ⲴṨ࠭ᮠⲴᮠ䟿Ǆቭ㇑ fˆ (x)ᱟሩf(x)Ⲵ‫ޘ‬ተ䙬䘁ˈ
նᶕ㠚⇿њKu(d(xu, x)) 亩Ⲵ䍑⥞㻛ተ䜘ॆࡠ⛩xu䱴䘁Ⲵ४ฏǄа⿽ᖸᑨ㿱Ⲵ‫⌅ڊ‬ᱟ䘹ᤙ儈ᯟ

࠭ᮠ˄Gaussian function˅
˄㿱㺘 5-4˅֌Ѫ⇿њṨ࠭ᮠKu(d(xu, x))ˈ儈ᯟ࠭ᮠⲴѝᗳ⛩Ѫxuˈ
ᯩᐞᱟ V u Ǆ
2
1 2
d ( xu , x )
2V u2
Ku(d(xu, x)) = e
л䶒ᡁԜᶕ䳶ѝ䇘䇪䘉њᑨ㿱Ⲵ儈ᯟṨ࠭ᮠǄṩᦞHartman et al.˄1990˅ᡰᤷࠪⲴˈ‫ޜ‬
ᔿ˄8.8˅䘉ṧⲴ࠭ᮠᖒᔿ㜭ཏԕԫ᜿ሿⲴ䈟ᐞ䙬䘁ԫօ࠭ᮠˈਚ㾱ԕк儈ᯟṨⲴᮠ䟿k䏣ཏ
བྷˈᒦфਟԕ࠶࡛ᤷᇊ⇿њṨⲴᇭᓖV2Ǆ
‫ޜ‬ᔿ˄8.8˅㔉ࠪⲴ࠭ᮠਟԕ㻛ⴻ֌ᱟ᧿䘠ҶањєቲⲴ㖁㔌ˈㅜаቲ䇑㇇н਼ⲴKu(d(xu,
x))ˈㅜҼቲ䇑㇇ㅜаቲঅ‫ݳ‬٬Ⲵ㓯ᙗ㓴ਸǄമ 8-2 ⭫ࠪҶањᖴੁส࠭ᮠ㖁㔌ⲴֻᆀǄ
മ 8-2 ањᖴੁส࠭ᮠ㖁㔌
⇿њ䳀㯿অ‫ݳ‬ӗ⭏ањ◰ਁ˄activation˅ˈᆳ⭡ԕḀњᇎֻxuѪѝᗳⲴ儈ᯟ࠭ᮠߣᇊǄᡰԕˈ䲔䶎
x䶐䘁xuˈ੖ࡉᆳⲴ◰ਁ᧕䘁Ҿ 0Ǆ䗃ࠪঅ‫ݳ‬ӗ⭏Ⲵ䗃ࠪᱟ䳀㯿অ‫Ⲵਁ◰ݳ‬㓯ᙗ㓴ਸǄቭ㇑䘉䟼⭫ࠪ
Ⲵ㖁㔌ӵᴹањ䗃ࠪˈնᱟҏਟԕवਜ਼ཊњ䗃ࠪǄ
㔉ᇊҶⴞḷ࠭ᮠⲴ䇝㓳ṧֻ䳶ਸˈа㡜࠶єњ䱦⇥ᶕ䇝㓳RBF(radial basis function)㖁
㔌Ǆ俆‫ߣˈݸ‬ᇊ䳀㯿অ‫Ⲵݳ‬ᮠ䟿kˈᒦ䙊䗷䘹ਆ⭘ᶕᇊѹṨ࠭ᮠKu(d(xu, x))Ⲵxuǃ V u ٬ᇊѹ

2
⇿њ䳀㯿অ‫ݳ‬ǄㅜҼˈ֯⭘ᔿ˄8.5˅㔉ࠪⲴ‫ޘ‬ተ䈟ᐞ߶ࡉᶕ䇝㓳ᵳ٬wuˈ֯㖁㔌ᤏਸ䇝㓳
ᮠᦞ〻ᓖᴰབྷॆǄഐѪṨ࠭ᮠ൘ㅜҼ䱦⇥ᱟ‫؍‬ᤱнਈⲴˈᡰԕ㓯ᙗᵳ٬wuਟԕ㻛䶎ᑨ儈᭸ൠ
䇝㓳ᗇࡠǄ
ӪԜᐢ㓿ᨀࠪҶࠐ⿽ᯩ⌅ᶕ䘹ਆ䘲ᖃⲴ䳀㯿অ‫ݳ‬ᡆ㘵䈤Ṩ࠭ᮠⲴᮠ䟿Ǆа⿽ᯩ⌅ᱟѪ⇿
ањ䇝㓳ṧֻ<xi, f(xi)>࠶䝽ањ儈ᯟṨ࠭ᮠˈ↔儈ᯟ࠭ᮠⲴѝᗳ⛩㻛䇮ѪxiǄᡰᴹ儈ᯟ࠭ᮠ
ⲴᇭᓖV2ਟ㻛䍻Ѫ਼ṧⲴ٬Ǆ䙊䗷䘉⿽ᯩ⌅ˈRBF㖁㔌ᆖҐⴞḷ࠭ᮠⲴ‫ޘ‬ተ䙬䘁ˈަѝ⇿њ
䇝㓳ṧֻ<xi, f(xi)>䜭ਚ൘xiⲴ䛫ฏ޵ᖡ૽ fˆ Ⲵ٬Ǆ䘉ṧ䘹ᤙṨ࠭ᮠⲴањՈ⛩ᱟᆳ‫ݱ‬䇨RBF
㖁㔌㋮⺞ൠᤏਸ䇝㓳ᮠᦞǄҏቡᱟ䈤ˈሩҾԫ᜿mњ䇝㓳ṧֻ䳶ਸˈѪҶਸᒦmњ儈ᯟṨ࠭
ᮠⲴᵳ٬w0}wmਟԕ㻛䇮㖞Ѫ֯ᗇሩҾ⇿ањ䇝㓳ṧֻ<xi, f(xi)>䜭┑䏣 fˆ (xi)=f(xi)Ǆ

ㅜҼ⿽ᯩ⌅ᱟ䘹ਆа㓴ᮠ䟿ቁҾ䇝㓳ṧֻᮠ䟿ⲴṨ࠭ᮠǄ䘉⿽ᯩ⌅ਟԕ∄ㅜа⿽ᯩ⌅ᴤ
ᴹ᭸ˈ⢩࡛ᱟ൘䇝㓳ṧֻⲴᮠ䟿ᐘབྷⲴᰦ‫ى‬ǄṨ࠭ᮠ㻛࠶ᐳ൘ᮤњᇎֻオ䰤XкˈᆳԜⲴѝ
ᗳѻ䰤ᴹ൷रⲴ䰤䳄Ǆᡆ㘵ҏਟԕ䶎൷रൠ࠶ᐳṨ࠭ᮠѝᗳˈ⢩࡛ᱟ൘ᇎֻᵜ䓛൘Xк䶎൷
र࠶ᐳⲴᰦ‫ى‬Ǆ൘ਾа⿽ᛵߥлˈਟԕ䲿ᵪ䘹ਆ䇝㓳ṧֻⲴањᆀ䳶֌ѪṨ࠭ᮠⲴѝᗳˈӾ
㘼ሩᇎֻⲴส߶࠶ᐳ䘋㹼䟷ṧǄᡆ㘵ˈᡁԜਟԕḷ䇶ࠪᇎֻⲴ৏࿻㚊㊫˄prototypical cluster˅ˈ
❦ਾԕ⇿њ㚊㊫Ѫѝᗳ࣐‫ޕ‬ањṨ࠭ᮠǄ䘉⿽ᯩᔿⲴṨ࠭ᮠᐳ㖞ਟԕ䙊䗷䶎ⴁⶓⲴ㚊㊫㇇⌅
ᶕᇎ⧠ˈަѝᢺ䇝㓳ᇎֻ˄нवਜ਼ⴞḷ٬˅ᤏਸࡠ␧ਸ儈ᯟǄ6.12.1 㢲䇘䇪ⲴEM㇇⌅ᨀ‫׋‬
Ҷа⿽Ӿkњ儈ᯟ࠭ᮠⲴ␧ਸѝ䘹ᤙ൷٬ˈԕᴰ֣ᤏਸ㿲⍻ࡠᇎֻⲴᯩ⌅Ǆ൘EM㇇⌅ѝˈ൷
٬Ⲵ䘹ਆᯩ⌅ᱟ˖ሩ㔉ᇊⲴkњՠ䇑Ⲵ൷٬ˈ֯㿲⍻ࡠᇎֻxiⲴᾲ⦷ᴰབྷॆǄ⌘᜿൘ᰐⴁⶓ
Ⲵ㚊㊫ᯩ⌅ѝˈᇎֻⲴⴞḷ࠭ᮠ٬f(xi)н৲оṨ࠭ᮠѝᗳⲴ䇑㇇Ǆⴞḷ٬Ⲵᜏа֌⭘ᱟߣᇊ
䗃ࠪቲⲴᵳ٬wuǄ
ᾲ㘼䀰ѻˈ⭘ཊњተ䜘Ṩ࠭ᮠⲴ㓯ᙗ㓴ਸ㺘⽪Ⲵᖴੁส࠭ᮠ㖁㔌ᨀ‫׋‬Ҷа⿽ⴞḷ࠭ᮠⲴ
‫ޘ‬ተ䙬䘁Ǆӵᖃ䗃‫ ޕ‬x 㩭‫ޕ‬ḀњṨ࠭ᮠⲴѝᗳ઼ᇭᓖᡰᇊѹⲴ४ฏ޵ᰦˈ䘉њṨ࠭ᮠⲴ٬᡽
ᱟнਟᘭ⮕ⲴǄഐ↔ˈRBF 㖁㔌ਟԕ㻛ⴻ֌ⴞḷ࠭ᮠⲴཊњተ䜘䙬䘁Ⲵᒣ━㓯ᙗ㓴ਸǄRBF
㖁㔌Ⲵањ‫ޣ‬䭞Ո⛩ᱟˈо৽ੁՐ᫝㇇⌅䇝㓳Ⲵࡽ侸㖁㔌⴨∄ˈᆳⲴ䇝㓳ᴤ࣐儈᭸Ǆ䘉ᱟഐ
Ѫ RBF 㖁㔌Ⲵ䗃‫ޕ‬ቲ઼䗃ࠪቲਟԕ㻛࠶࡛䇝㓳Ǆ
8.5 ะӄṾׁⲺ᧞⨼
k-䘁䛫㇇⌅઼ተ䜘࣐ᵳഎᖂ䜭ᱟสҾᇎֻⲴᯩ⌅ˈᆳԜާᴹйњ‫ޣⲴ਼ޡ‬䭞⢩ᙗǄㅜ 1ˈ
ᆳԜᱟ⎸ᶱᆖҐᯩ⌅ˈ䜭ᢺ൘䇝㓳ᮠᦞкⲴ⌋ॆ᧘䘏㠣䙷ࡠањᯠⲴḕ䈒ᇎֻᰦǄㅜ 2ˈᆳ
Ԝ䙊䗷࠶᷀⴨լⲴᇎֻᶕ࠶㊫ᯠⲴḕ䈒ᇎֻˈ㘼ᘭ⮕оḕ䈒ᶱަн਼ⲴᇎֻǄㅜ 3ˈᆳԜᢺ
ᇎֻ㺘⽪Ѫ n 㔤⅗∿オ䰤ѝⲴᇎᮠ⛩ǄสҾṸֻⲴ᧘⨶˄Case-based reasoningˈCBR˅䘉⿽
ᆖҐ㤳රสҾࡽєњ৏ࡉˈնнवᤜㅜ 3 њǄ൘ CBR ѝˈа㡜֯⭘ᴤѠᇼⲴㅖਧ᧿䘠ᶕ㺘
⽪ᇎֻ˗⴨ᓄൠˈ⭘ᶕỰ㍒ᇎֻⲴᯩ⌅ҏᴤ࣐༽ᵲǄCBR ᐢ㻛ᓄ⭘Ҿ䀓ߣᖸཊ䰞仈ˈ∄ྲˈ
ṩᦞᮠᦞᓃѝᆈۘⲴԕࡽⲴ䇮䇑മ㓨ˈᶕ䘋㹼ᵪỠ䇮༷Ⲵᙫփ䇮䇑˄Sycara et al. 1992˅
(solidworks, proe ㅹй㔤䖟ԦⲴ䇮䇑৏⨶ԕ৺ᓄ⭘ੇ˛˛˛)˗ṩᦞԕࡽⲴ㻱ߣᶕሩᯠⲴ⌅ᖻ
ṸԦ䘋㹼᧘⨶˄Ashley 1990˅˗䙊䗷ሩԕࡽⲴ⴨լ䰞仈Ⲵ䀓ߣᯩṸⲴ༽⭘ᡆਸᒦˈᶕ䀓ߣ㿴
ࡂ઼䈳ᓖ䰞仈˄Veloso 1992˅Ǆ
֌Ѫԕਾ䇘䇪Ⲵส⹰ˈ䇙ᡁԜ㘳㲁สҾṸֻⲴ᧘⨶㌫㔏ⲴањֻᆀǄCADET㌫㔏
˄Sycara et al. 1992˅䟷⭘สҾṸֻⲴ᧘⨶ᶕ䖵ࣙㆰঅᵪỠ䇮༷˄ֻྲ≤嗉ཤ˅Ⲵᙫփ䇮䇑Ǆ
ᆳ֯⭘ањᮠᦞᓃˈަѝवਜ਼བྷ㓖 75 њԕࡽⲴ䇮䇑ᡆ䇮䇑⡷ᯝˈᶕ᧘㦀ㅖਸᯠⲴ䇮䇑㿴Ṭ
Ⲵᙫփ䇮䇑Ǆ޵ᆈѝ⇿ањᇎֻᱟ䙊䗷ᆳⲴ㔃ᶴ઼ᇊᙗⲴ࣏㜭ᶕ㺘⽪ⲴǄ⴨ᓄⲴˈᯠⲴ䇮䇑
䰞仈ᱟ䙊䗷ᡰ㾱≲Ⲵ࣏㜭઼㔃ᶴᶕ㺘⽪ⲴǄമ 8-3 ⭫ࠪҶ䘉њ䰞仈ǄമⲴкॺ䜘࠶ᱮ⽪Ҷа
њިරⲴᆈۘṸֻˈ㻛〠ѪTර᧕ཤ㇑ǄᆳⲴ࣏㜭㻛㺘⽪Ѫ䗃‫઼ޕ‬䗃ࠪ⛩Ⲵ⍱䟿઼⑙ᓖ䰤Ⲵ
ᇊᙗ‫ޣ‬㌫Ǆ൘ਣ‫࣏Ⲵח‬㜭᧿䘠ѝˈḷᴹĀ+āⲴ㇝ཤ㺘᰾㇝ཤཤ䜘Ⲵਈ䟿䲿⵰㇝ཤቮ䜘Ⲵਈ
䟿кॷǄֻྲˈ䗃ࠪ⍱䟿Q3䲿⵰䗃‫ޕ‬⍱䟿Q1໎䮯Ǆ㊫լൠˈ Ā-āḷ䇠㺘᰾㇝ཤཤ䜘Ⲵਈ䟿䲿
⵰㇝ཤቮ䜘Ⲵਈ䟿л䱽Ǆ䘉ᑵമⲴлॺ䜘࠶⭫ࠪҶањᯠⲴ䇮䇑䰞仈ˈᆳ䙊䗷ᯠ䇮䇑ѝᡰ㾱
≲Ⲵ࣏㜭ᶕ᧿䘠Ǆ䘉њ࣏㜭᧿㔈Ҷа⿽≤嗉ཤⲴ㹼Ѫ⢩ᖱǄ䘉䟼Qcᤷ䘋‫ޕ‬嗉ཤⲴߧ≤⍱䟿ˈ
Qhᤷ✝≤Ⲵ䗃‫ޕ‬⍱䟿ˈQmᤷ⍱ࠪ嗉ཤⲴঅа␧ਸ⍱䟿Ǆ㊫լൠˈTcǃTh઼Tm࠶࡛ᤷ✝≤ǃߧ
≤઼␧ਸ≤⍱Ⲵ⑙ᓖǄਈ䟿Ct㺘⽪䗃‫ࡠޕ‬嗉ཤⲴ⑙ᓖ᧗ࡦؑਧˈCf㺘⽪ሩ≤⍱Ⲵ᧗ࡦؑਧǄ
⌘᜿ˈᡰ㾱≲Ⲵ࣏㜭᧿䘠ѝᤷࠪˈ䘉Ӌ᧗ࡦؑਧCt઼Cf⭘ᶕᖡ૽≤⍱Qc઼QhˈӾ㘼䰤᧕ᖡ૽
嗉ཤⲴ䗃ࠪ⍱䟿Qm઼⑙ᓖTmǄ
A stored case: T-junction tube- ањᆈۘⲴṸֻ˖T ර᧕ཤ㇑
Structure-㔃ᶴ
Function-࣏㜭
A problem specification: Water faucet- ањ䰞仈Ⲵ㿴Ṭ䈤᰾˖≤嗉ཤ
Structure-㔃ᶴ
Function-࣏㜭
മ 8-3 ањᆈۘⲴṸ઼ֻањᯠ䰞仈
кॺ䜘࠶᧿㔈Ҷ CADET ṸֻᓃѝањިරⲴ䇮䇑⡷ᯝǄᆳⲴ࣏㜭ᱟ䙊䗷 T ර᧕ཤਈ䟿䰤Ⲵᇊᙗ‫׍‬

䎆‫ޣ‬㌫മ㺘⽪Ⲵ˄൘↓᮷ѝާփ᧿䘠˅Ǆлॺ䜘࠶ᱮ⽪ҶањިරⲴ䇮䇑䰞仈Ǆ
㔉ᇊᯠ䇮䇑䰞仈Ⲵ࣏㜭䈤᰾ˈCADET ӾᆳⲴṸֻᓃѝᩌ㍒ᆈۘⲴֻˈ֯ᆳⲴ࣏㜭᧿䘠
઼ᯠ䇮䇑䰞仈⴨३䝽Ǆྲ᷌ਁ⧠Ҷањ㋮⺞Ⲵ३䝽ˈ㺘᰾ḀњᆈۘṸֻ㋮⺞ᇎ⧠Ҷᡰ㾱≲Ⲵ
࣏㜭ˈ䛓Ѹਟԕ䘄എ䘉њṸֻ֌Ѫᯠ䇮䇑䰞仈Ⲵᔪ䇞ᯩṸǄྲ᷌⋑ᴹਁ⧠㋮⺞Ⲵ३䝽ˈ
CADET ਟ㜭᢮ࡠ३䝽ᡰ䴰࣏㜭Ⲵн਼ᆀമⲴṸֻǄֻྲˈ൘മ 8-3 ѝ T ර᧕ཤⲴ࣏㜭३䝽
Ҷ ≤ 嗉 ཤ ࣏ 㜭 മ Ⲵ а њ ᆀ മ Ǆ ᴤ а 㡜 ൠ 䇢 ˈ CADET ൘ є њ ࣏ 㜭 മ 䰤 ᩌ ㍒ ਼ ᶴ ᆀ മ
˄isomorphisms subgraph˅
ˈԕਁ⧠ањṸֻⲴḀ䜘࠶ˈ֯ᆳ३䝽䇮䇑㿴Ṭ䈤᰾Ⲵ⴨ᓄ䜘࠶Ǆ
↔ཆˈ㌫㔏ਟԕ࣐ᐕ৏࿻Ⲵ࣏㜭䈤᰾മˈӗ⭏ㅹԧⲴᆀമԕ३䝽ᴤཊⲴṸֻǄᆳ֯⭘‫ޣ‬Ҿ⢙
⨶ᝏᓄⲴа㡜⸕䇶ᶕࡋᔪ䘉ṧⲴ࣐ᐕ䗷Ⲵ࣏㜭മǄֻྲˈ࡙⭘а⿽䟽߉㿴ࡉਟԕᢺ䘉њᝏᓄ˖

A
o B
䟽߉Ѫ

A
o x
o B
䘉њ䟽߉㿴ࡉਟԕ㻛䀓䟺Ѫ˖ྲ᷌ B 䲿 A кॷˈ䛓Ѹаᇊਟԕ᢮ࡠḀњ xˈ┑䏣 B 䲿 x

кॷ㘼ф x 䲿 A кॷǄ䘉䟼 x ᱟањ‫〠ޘ‬䟿ॆⲴਈ䟿ˈᆳ൘࣏㜭മо䘉њṸֻᓃ३䝽ᰦ㓖ᶏ
ࡠ⺞ᇊ٬Ǆһᇎкˈമ 8-3 ѝⲴ≤嗉ཤⲴ࣏㜭മቡᱟᓄ⭘䘉њ䟽߉㿴ࡉӾ৏ᶕⲴ࣏㜭䈤᰾ѝ
࣐ᐕᗇࡠⲴǄ
䙊䗷Ự㍒३䝽н਼ᆀമⲴཊњṸֻˈᴹᰦਟԕ᤬᧕ᗇࡠᮤњ䇮䇑Ǆа㡜ᶕ䈤ˈӾཊњỰ
㍒ࡠⲴṸֻӗ⭏ᴰ㓸ᯩṸⲴ䗷〻ਟԕᖸ༽ᵲǄѪҶਸᒦᆈۘṸֻѝⲴỰ㍒ࡠⲴ䜘࠶ˈਟ㜭䴰
㾱Ӿཤ䇮䇑㌫㔏Ⲵ਴њ䜘࠶Ǆҏਟ㜭䴰㾱എⓟԕࡽⲴ䇮䇑ᆀⴞḷˈӾ㘼ђᔳࡽ䶒Ự㍒ࡠⲴṸ
ֻǄCADET ਸᒦ઼㠚䘲ᓄᐢỰ㍒ࡠṸֻᒦᖒᡀᴰ㓸䇮䇑Ⲵ㜭࣋ᖸᴹ䲀ˈᆳѫ㾱‫׍‬䎆⭘ᡧᶕ
‫ڊ‬㠚䘲ᓄ䱦⇥Ⲵ༴⨶Ǆ↓ྲ Sycara et al.˄1992˅ᡰ᧿䘠ⲴˈCADET ᱟањ⹄ウ⭘Ⲵ৏ර㌫
㔏ˈ⭘ᶕ᧒㍒สҾṸֻⲴ᧘⨶൘ᙫփ䇮䇑ѝⲴ▌൘֌⭘Ǆᆳнާ༷⭘ᶕᢺ䘉Ӌᣭ䊑Ⲵᙫփ䇮
䇑ᨀ⛬ᡀᴰ㓸䇮䇑Ⲵ࠶᷀㇇⌅Ǆ
ሩҾCADETⲴ䰞仈Ṷ઼ᷦสҾᇎֻⲴᯩ⌅˄ֻྲk-䘁䛫㇇⌅˅Ⲵа㡜Ṷᷦˈ࠶᷀ᆳԜ
ѻ䰤Ⲵሩᓄѻ༴ᱟᴹ⳺ⲴǄ൘CADETѝ⇿њᆈۘⲴ䇝㓳ṧֻ᧿㔈Ҷањ࣏㜭മԕ৺ᇎ⧠䈕
࣏㜭Ⲵ㔃ᶴǄᯠⲴḕ䈒ሩᓄᯠⲴ࣏㜭മǄഐ↔ˈᡁԜਟԕᢺCADETⲴ䰞仈᱐ሴࡠḷ߶Ⲵᆖ
Ґ䰞仈ᇊѹѝǄަѝᇎֻオ䰤XᇊѹѪᡰᴹ࣏㜭മⲴオ䰤Ǆⴞḷ࠭ᮠf᱐ሴࡠᇎ⧠䘉Ӌ࣏㜭Ⲵ
㔃ᶴǄ⇿њᆈۘ䇝㓳ṧֻ<x, f(x)>ᱟањᒿ‫᧿ˈڦ‬䘠Ḁњ࣏㜭മx઼ᇎ⧠xⲴ㔃ᶴf(x)Ǆ㌫㔏ᗵ
享ᆖҐ䇝㓳Ṹֻˈԕ䗃ࠪ┑䏣࣏㜭മḕ䈒䗃‫ޕ‬xqⲴ㔃ᶴf(xq)Ǆ
к䶒‫ޣ‬Ҿ CADET ㌫㔏ㆰ㾱᧿䘠ˈ䈤᰾ҶสҾṸֻⲴ᧘⨶㌫㔏४࡛Ҿ k-䘁䛫䘉ṧⲴᯩ⌅

Ⲵ㤕ᒢа㡜⢩ᖱ˖
x ᇎֻᡆṸֻਟԕ⭘ѠᇼⲴㅖਧ᧿䘠㺘⽪ˈቡ‫ ۿ‬CADET ѝ֯⭘Ⲵ࣏㜭മǄ䘉ਟ㜭
䴰㾱н਼Ҿ⅗∿䐍⿫Ⲵ⴨լᙗᓖ䟿ˈ∄ྲєњ࣏㜭മⲴᴰབྷਟ‫ޡ‬ӛᆀമⲴབྷሿǄ
x Ự㍒ࡠⲴཊњṸֻਟԕਸᒦᖒᡀᯠ䰞仈Ⲵ䀓ߣᯩṸǄ䘉о k-䘁䛫ᯩ⌅⴨լüü
ཊњ⴨լⲴṸֻ⭘ᶕᶴᡀሩᯠḕ䈒ⲴഎㆄǄ❦㘼ˈਸᒦཊњỰ㍒ࡠⲴṸֻⲴ䗷
〻о k-䘁䛫ᴹᖸབྷн਼ˈᆳ‫׍‬䎆Ҿ⸕䇶᧘⨶㘼нᱟ㔏䇑ᯩ⌅Ǆ
x ṸֻỰ㍒ǃสҾ⸕䇶Ⲵ᧘⨶઼䰞仈≲䀓䰤ᱟ㍗ᇶ㙖ਸ൘а䎧ⲴǄֻྲ CADET
㌫㔏൘ቍ䈅᢮ࡠ३䝽ⲴṸֻ䗷〻ѝˈᆳ֯⭘ᴹ‫ޣ‬⢙⨶ᝏᓄⲴа㡜⸕䇶䟽߉Ҷ࣏
㜭മǄӪԜᐢ㓿ᔰਁࠪᖸཊަԆⲴ㌫㔏ˈ䘉Ӌ㌫㔏ᴤ࣐ᆼᮤൠᢺสҾṸֻⲴ᧘
⨶䳶ᡀࡠสҾᩌ㍒Ⲵ䰞仈≲䀓㌫㔏ѝǄANAPRON˄Golding & Rosenbloom 1991˅
઼ Prodigy/Analogy˄Veloso 1992˅ᱟєњֻᆀǄ
ᾲᤜൠ䇢ˈสҾṸֻⲴ᧘⨶ᱟа⿽สҾᇎֻⲴᆖҐᯩ⌅ˈ൘䘉⿽ᯩ⌅ѝˈᇎֻ˄Ṹֻ˅
ਟԕᱟѠᇼⲴ‫ޣ‬㌫᧿䘠˗㘼ф൘䈕ᯩ⌅ѝˈѪҶ䀓ߣᖃࡽḕ䈒ˈṸֻỰ㍒઼ਸᒦ䗷〻ਟ㜭‫׍‬
䎆Ҿ⸕䇶᧘⨶઼ᩌ㍒ᇶ䳶Ⲵ䰞仈≲䀓ᯩ⌅Ǆⴞࡽ‫ޣ‬ҾสҾṸֻⲴ᧘⨶⹄ウⲴањ䈮仈ᱟˈ᭩
䘋㍒ᕅṸֻⲴᯩ⌅Ǆ䘉䟼Ⲵѝᗳ䰞仈ᱟਕ⌅⴨լᓖ䟿˄ֻྲˈ࣏㜭മѻ䰤Ⲵᆀമ਼ᶴ˅ӵ㜭
䘁լൠᤷࠪ⢩ᇊṸֻо⢩ᇊ䰞仈Ⲵ⴨‫ޣ‬ᓖǄᖃ CBR ㌫㔏䈅മ༽⭘Ự㍒ࡠⲴṸֻᰦˈᆳਟ㜭
䙷ࡠਕ⌅⴨լᓖ䟿ѝ⋑ᴹᦅ᥹ࡠⲴ䳮⛩Ǆֻྲˈ൘ CADET ѝˈỰ㍒ࡠⲴཊњ䇮䇑⡷ᯝਟ㜭
ᖬ↔нެᇩˈ֯ᗇᆳԜᰐ⌅㻛ਸᒦࡠањ㔏аⲴᴰ㓸䇮䇑ѝǄа㡜ᖃ䘉⿽ᛵߥਁ⭏ᰦˈCBR
㌫㔏ਟഎⓟᩌ㍒ਖཆⲴṸֻԕ䘲ᓄ⧠ᴹⲴṸֻˈᡆ㘵≲ࣙҾަԆⲴ䰞仈≲䀓ᯩ⌅Ǆ䟽㾱Ⲵᱟˈ
ᖃỰ⍻ࡠ䘉ṧⲴ䳮⛩ᰦˈᆳԜҏᨀ‫׋‬Ҷ⭘ᶕ᭩䘋⴨լᙗᓖ䟿˄ᡆㅹԧⲴˈṸֻᓃ㍒ᕅ㔃ᶴ˅
Ⲵ䇝㓳ᮠᦞǄ⺞࠷ൠ䇢ˈྲ᷌ṩᦞ⴨լᙗᓖ䟿Ự㍒ࡠҶањṸֻˈն൘䘋а↕Ⲵ࠶᷀ѝਁ⧠
䘉њṸֻоᖃࡽⲴ䇮䇑ᱟᰐ‫ˈⲴޣ‬䛓Ѹ䘉њ⴨լᙗᓖ䟿Պ㻛᭩䘋ˈԕ‫ׯ‬ሩҾԕਾⲴ㊫լḕ䈒
ᤂ㔍䘉њṸֻǄ
8.6 ሯ⎾ᶷᆜҖૂ〥ᶷᆜҖⲺ䇺䇰
൘䘉аㄐѝᡁԜ㘳㲁Ҷй⿽⎸ᶱᆖҐ˄lazy learning˅ᯩ⌅˖k-䘁䛫㇇⌅ǃተ䜘࣐ᵳഎᖂ
઼สҾṸֻⲴ᧘⨶Ǆѻᡰԕ〠䘉Ӌᯩ⌅ᱟ⎸ᶱⲴˈᱟഐѪᆳԜᔦ䘏ҶྲօӾ䇝㓳ᮠᦞѝ⌋ॆ
Ⲵߣㆆˈⴤࡠ䙷ࡠањᯠⲴḕ䈒Ǆᵜㄐ䇘䇪Ҷа⿽〟ᶱᆖҐᯩ⌅˖ᆖҐᖴੁส࠭ᮠ㖁㔌Ⲵᯩ
⌅Ǆѻᡰԕ〠䘉⿽ᯩ⌅ᱟ〟ᶱⲴˈᱟഐѪᆳ൘㿱ࡠᯠⲴḕ䈒ѻࡽቡ‫ྭڊ‬Ҷ⌋ॆⲴᐕ֌üü൘
䇝㓳ᰦᨀӔҶᇊѹަⴞḷ࠭ᮠ䙬䘁Ⲵ㖁㔌㔃ᶴ઼ᵳ٬Ǆṩᦞ਼ṧⲴ⨶䀓ˈᵜҖަԆㄐ㢲䇘䇪
ⲴᡰᴹަԆ㇇⌅䜭ᱟ〟ᶱᆖҐ㇇⌅˄ֻྲˈ৽ੁՐ᫝㇇⌅ǃC4.5˅Ǆ
൘㇇⌅㜭࣋ᯩ䶒ˈ⎸ᶱᯩ⌅઼〟ᶱᯩ⌅ᴹ᰾ᱮᐞᔲੇ˛ᡁԜ‫ݸ‬४࠶є⿽ᐞᔲ˖䇑㇇ᰦ䰤
Ⲵᐞᔲˈ઼ሩᯠḕ䈒Ⲵ࠶㊫ᐞᔲǄ൘䇑㇇ᰦ䰤ᯩ䶒⎸ᶱᯩ⌅઼〟ᶱᯩ⌅ᱮ❦ᴹᐞᔲǄֻྲˈ
⎸ᶱᯩ⌅൘䇝㓳ᰦа㡜䴰㾱䖳ቁⲴ䇑㇇ˈն൘亴⍻ᯠḕ䈒Ⲵⴞḷ٬ᰦ䴰㾱ᴤཊⲴ䇑㇇ᰦ䰤Ǆ
ᴤสᵜⲴ䰞仈ᱟˈ൘ᖂ㓣‫ٿ‬㖞ᯩ䶒⎸ᶱ઼〟ᶱᯩ⌅ᱟ੖ᴹᇎ䍘ᙗⲴᐞᔲ઒˛൘䘉ᯩ䶒є
⿽ᯩ⌅Ⲵ‫ޣ‬䭞ᐞᔲᱟ˖
x ⎸ᶱᯩ⌅൘ߣᇊྲօӾ䇝㓳ᮠᦞDѝ⌋ॆᰦ㘳㲁ḕ䈒ᇎֻxqǄ
x 〟ᶱᯩ⌅н㜭‫ࡠڊ‬䘉а⛩ˈഐѪ൘㿱ࡠḕ䈒ᇎֻxqࡽˈᆳԜᐢ㓿䘹ਆҶሩⴞ
ḷ࠭ᮠⲴ˄‫ޘ‬ተ˅䙬䘁Ǆ
䘉њ४࡛Պᖡ૽ᆖҐಘⲴ⌋ॆ㋮ᓖੇ˛ྲ᷌㾱≲⎸ᶱⲴ઼〟ᶱⲴᆖҐಘ䟷⭘਼ањ‫ٷ‬
䇮オ䰤Hˈ䛓ѸㆄṸᱟ㛟ᇊⲴǄѪҶ䈤᰾䘉а⛩ˈ㘳㲁⭡㓯ᙗ࠭ᮠ㓴ᡀⲴ‫ٷ‬䇮オ䰤Ǆࡽ䶒䇘
䇪Ⲵተ䜘࣐ᵳഎᖂ㇇⌅ᱟสҾ䘉ṧⲴ‫ٷ‬䇮オ䰤Ⲵ⎸ᶱᆖҐᯩ⌅ǄሩҾ⇿њᯠḕ䈒xqˈᆳṩᦞ
xq䱴䘁Ⲵ䇝㓳ṧֻ䘹ᤙањᯠⲴ‫ٷ‬䇮Ӿ䇝㓳ᮠᦞѝ⌋ॆǄ⴨৽ˈањ֯⭘਼ṧⲴ㓯ᙗ࠭ᮠ‫ٷ‬
䇮オ䰤Ⲵ〟ᶱᆖҐಘᗵ享൘㿱ࡠḕ䈒ѻࡽ䘹ᤙሩⴞḷ࠭ᮠⲴ䙬䘁Ǆᡰԕ〟ᶱᆖҐಘᗵ享ᨀӔ
অњⲴ㓯ᙗ࠭ᮠ‫ٷ‬䇮ˈԕ㾶ⴆᮤњᇎֻオ䰤઼ᡰᴹᵚᶕⲴḕ䈒Ǆ⎸ᶱᆖҐᯩ⌅ᴹ᭸ൠ֯⭘Ҷ
ᴤѠᇼⲴ‫ٷ‬䇮オ䰤ˈഐѪᆳ֯⭘ᖸཊн਼Ⲵተ䜘㓯ᙗ࠭ᮠᶕᖒᡀሩⴞḷ࠭ᮠⲴ䳀ਜ਼Ⲵ‫ޘ‬ተ䙬
䘁Ǆ⌘᜿ަԆⲴаӋᆖҐಘ઼‫ٷ‬䇮オ䰤ҏㅖਸ਼ṧⲴᛵߥǄֻྲ৽ੁՐ᫝㇇⌅Ⲵ⎸ᶱ⡸ᵜˈ
ਟԕሩ⇿њ⤜・Ⲵḕ䈒⛩ᆖҐн਼Ⲵ⾎㓿㖁㔌Ǆ䘉оㅜ 4 ㄐ䇘䇪Ⲵ৽ੁՐ᫝㇇⌅Ⲵ〟ᶱ⡸ᵜ
ᖒᡀሩ➗Ǆ
к䶒а⇥ⲴṨᗳ㿲⛩ᱟˈ⎸ᶱⲴᆖҐಘਟԕ䙊䗷ᖸཊተ䜘䙬䘁Ⲵ㓴ਸ˄䳀ਜ਼ൠ˅㺘⽪ⴞ
ḷ࠭ᮠˈ❦㘼〟ᶱⲴᆖҐಘᗵ享൘䇝㓳ᰦᨀӔঅњⲴ‫ޘ‬ተ䙬䘁Ǆഐ↔〟ᶱᆖҐⲴ઼⎸ᶱᆖҐ
ѻ䰤Ⲵᐞᔲ᜿ણ⵰ሩⴞḷ࠭ᮠⲴ‫ޘ‬ተ䙬䘁઼ተ䜘䙬䘁ⲴᐞᔲǄ
֯⭘ཊњተ䜘䙬䘁Ⲵ〟ᶱᯩ⌅ˈਟԕӗ⭏о⎸ᶱᯩ⌅Ⲵተ䜘䙬䘁਼ṧⲴ᭸᷌ੇ˛ᖴੁส
࠭ᮠ㖁㔌ਟԕ㻛ⴻ֌ੁ䘉њⴞḷⲴቍ䈅ǄRBF ᆖҐᯩ⌅ᱟ൘䇝㓳ᰦᨀӔⴞḷ࠭ᮠ‫ޘ‬ተ䙬䘁
Ⲵ〟ᶱᯩ⌅Ǆ❦㘼ˈањ RBF 㖁㔌ᢺ䘉њ‫ޘ‬ተ࠭ᮠ㺘⽪Ѫཊњተ䜘Ṩ࠭ᮠⲴ㓯ᙗ㓴ਸǄн
䗷ˈഐѪ RBF ᆖҐᯩ⌅ᗵ享൘⸕䚃ḕ䈒⛩ѻࡽᨀӔ‫ٷ‬䇮ˈᡰԕᆳԜࡋᔪⲴተ䜘䙬䘁н㜭䗮
ࡠ‫⎸ۿ‬ᶱᆖҐᯩ⌅ѝ䛓ṧ⢩࡛䪸ሩḕ䈒⛩ǄԓᴯൠˈRBF 㖁㔌ᱟӾԕ䇝㓳ṧֻѪѝᗳⲴተ
䜘䙬䘁ѝ㻛Ā〟ᶱāᔪ・Ⲵˈᡆ㘵䈤ᱟԕ䇝㓳ṧֻⲴ㚊㊫Ѫѝᗳˈнᱟԕᵚ⸕Ⲵᵚᶕḕ䈒⛩
ѪѝᗳǄ
ᾲ㘼䀰ѻˈ⎸ᶱᆖҐᯩ⌅ਟԕሩҾ⇿ањḕ䈒ᇎֻ䘹ᤙн਼Ⲵ‫ٷ‬䇮˄ᡆⴞḷ࠭ᮠⲴተ䜘
䙬䘁˅Ǆ֯⭘਼ṧ‫ٷ‬䇮オ䰤Ⲵ〟ᶱᯩ⌅ᱟᴤ࣐ਇ䲀ࡦⲴˈഐѪᆳԜᗵ享ᨀӔањ㾶ⴆᮤњᇎ
ֻオ䰤Ⲵঅа‫ٷ‬䇮Ǆᖃ❦ˈ〟ᶱⲴᯩ⌅ਟԕ֯⭘ਸᒦҶཊњተ䜘䙬䘁Ⲵ‫ٷ‬䇮オ䰤ˈቡ䊑 RBF
㖁㔌аṧǄ❦㘼ˈণ֯ᱟ䘉ӋਸᒦⲴተ䜘䙬䘁ˈҏн㜭֯〟ᶱᯩ⌅ᆼ‫ާޘ‬ᴹ⎸ᶱᯩ⌅䛓⿽䪸
ሩᵚ⸕ḕ䈒֌ࠪ‫ٷ‬䇮Ⲵ㜭࣋Ǆ
x สҾᇎֻⲴᆖҐᯩ⌅н਼ҾަԆⲴ࠭ᮠ䙬䘁ᯩ⌅ˈഐѪᆳԜ᧘䘏༴⨶䇝㓳ṧֻˈ
ⴤࡠᗵ享࠶㊫ањᯠḕ䈒ᇎֻᰦǄഐ↔ˈᆳԜнᗵᖒᡀањ᰾⺞Ⲵ‫ٷ‬䇮ᶕᇊѹ
ᮤњᇎֻオ䰤кⲴᆼᮤⴞḷ࠭ᮠǄ⴨৽ˈᆳԜਟԕሩ⇿њḕ䈒ᇎֻᖒᡀањн
਼Ⲵⴞḷ࠭ᮠተ䜘䙬䘁Ǆ
x สҾᇎֻⲴᯩ⌅ⲴՈ⛩वᤜ˖䙊䗷а㌫ࡇнཚ༽ᵲⲴተ䜘䙬䘁ᶕ⁑ᤏ༽ᵲⴞḷ
࠭ᮠⲴ㜭࣋˗нՊᦏཡ䇝㓳ṧֻѝ㮤ਜ਼Ⲵԫօؑ᚟˄ഐѪһֻᵜ䓛㻛ⴤ᧕ൠᆈ
ۘ䎧ᶕ˅Ǆѫ㾱Ⲵᇎ䐥䰞仈वᤜ˖࠶㊫ᯠᇎֻⲴ᭸⦷˄ᡰᴹⲴ༴⨶䜭൘ḕ䈒ᵏ
䘋㹼㘼нᱟһ‫˗˅ྭ༷߶ݸ‬䳮ԕ䘹ᤙ⭘ᶕỰ㍒⴨‫ޣ‬ᇎֻⲴਸ䘲Ⲵ䐍⿫ᓖ䟿˄⢩
࡛ᱟᖃᇎֻᱟ⭘༽ᵲⲴㅖਧ㺘⽪᧿䘠Ⲵᰦ‫˗˅ى‬ᰐ‫ޣ‬⢩ᖱሩ䐍⿫ᓖ䟿Ⲵ䍏֌⭘Ǆ
x k-䘁䛫ᱟ⭘ᶕ䙬䘁ᇎᮠ٬ᡆ⿫ᮓ٬ⴞḷ࠭ᮠⲴสҾᇎֻ㇇⌅ˈᆳ‫ٷ‬ᇊᇎֻሩᓄҾ
n 㔤⅗∿オ䰤ѝⲴ⛩Ǆањᯠḕ䈒Ⲵⴞḷ࠭ᮠ٬ᱟṩᦞ k њоަᴰ䘁Ⲵ䇝㓳ṧֻ
Ⲵ٬ՠ䇑ᗇࡠⲴǄ
x ተ䜘࣐ᵳഎᖂ⌅ᱟ k-䘁䛫ᯩ⌅Ⲵ᧘ᒯˈ൘䘉⿽ᯩ⌅ѝˈѪ⇿њḕ䈒ᇎֻᔪ・а
њ᰾⺞Ⲵⴞḷ࠭ᮠⲴተ䜘䙬䘁Ǆⴞḷ࠭ᮠⲴተ䜘䙬䘁ਟԕสҾ‫ۿ‬ᑨᮠǃ㓯ᙗ࠭
ᮠᡆҼ⅑࠭ᮠ䘉ṧⲴབྷ䟿Ⲵ࠭ᮠᖒᔿˈҏਟԕสҾオ䰤ተ䜘ॆⲴṨ࠭ᮠǄ
x ᖴੁส࠭ᮠ˄RBF˅㖁㔌ᱟа㊫⭡オ䰤ተ䜘ॆṨ࠭ᮠᶴᡀⲴӪᐕ⾎㓿㖁㔌Ǆᆳਟ
㻛ⴻ֌ᱟสҾᇎֻⲴᯩ⌅˄⇿њṨ࠭ᮠⲴᖡ૽ᱟ㻛ተ䜘ॆⲴ˅઼⾎㓿㖁㔌ᯩ⌅
˄൘䇝㓳ᵏᖒᡀҶሩⴞḷ࠭ᮠⲴ‫ޘ‬ተ䙬䘁ˈ㘼нᱟ൘ḕ䈒ᵏᖒᡀተ䜘䙬䘁˅Ⲵ
␧ਸǄᖴੁส࠭ᮠ㖁㔌ᐢ㻛ᡀ࣏ൠᓄ⭘ࡠᖸཊ䈮仈ˈ∄ྲ㿶㿹൪Ჟ࠶᷀
˄interpreting visual scenes˅ˈަѝ‫ٷ‬ᇊオ䰤ተ䜘Ⲵᖡ૽ᱟᖸਸ⨶ⲴǄ
x สҾṸֻⲴ᧘⨶ҏᱟа⿽สҾᇎֻⲴᆖҐᯩ⌅ˈն䘉⿽ᯩ⌅֯⭘༽ᵲⲴ䙫䗁᧿
䘠㘼нᱟ⅗∿オ䰤ѝⲴ⛩ᶕ㺘⽪ᇎֻǄ㔉ᇊᇎֻⲴㅖਧ᧿䘠ˈӪԜᐢ㓿ᨀࠪҶ
བྷ䟿Ⲵᯩ⌅⭘Ҿᢺ䇝㓳ṧֻ᱐ሴᡀᯠᇎֻⲴⴞḷ࠭ᮠ٬ǄสҾṸֻⲴ᧘⨶ᯩ⌅
ᐢ㓿ᓄ⭘ࡠᖸཊᇎ䱵䰞仈ѝˈ∄ྲ⁑ᤏ⌅ᖻ᧘⨶ˈԕ৺൘༽ᵲⲴ⭏ӗ઼䘀䗃㿴
ࡂ䰞仈ѝᕅሬᩌ㍒Ǆ
k-䘁䛫㇇⌅ᱟᵪಘᆖҐѝ㻛࠶᷀ᗇᴰ䘿ᖫⲴ㇇⌅ѻаˈ৏ഐа䜘࠶ᱟ⭡Ҿᆳࠪ⧠Ⲵ䖳
ᰙˈਖཆҏ⭡ҾᆳⲴㆰ᰾ᙗǄCover & Hart˄1967˅ᨀࠪҶᰙᵏⲴ⨶䇪㔃᷌ˈDuda & Hart
˄1973˅ᨀ‫׋‬ҶањᖸྭⲴᾲ㿲ǄBishop˄1995˅䇘䇪Ҷ k-䘁䛫㇇⌅ԕ৺ᆳоᾲ⦷ᇶᓖՠ䇑
Ⲵ‫ޣ‬㌫ǄAtkeson et al.˄1997˅ሩተ䜘࣐ᵳഎᖂᯩ⌅㔉ࠪҶањ䶎ᑨྭⲴ㓥㿸ǄAtkeson et al.
˄1997b˅䈳ḕҶ䘉Ӌᯩ⌅൘ᵪಘӪ᧗ࡦᯩ䶒Ⲵᓄ⭘Ǆ
Bishop˄1995˅ᨀ‫׋‬Ҷањሩᖴੁส࠭ᮠⲴ‫ޘ‬䶒䇘䇪ǄަԆ䇪䘠⭡ Powell˄1987˅઼
Poggio & Girosi˄1990˅㔉ࠪǄᵜҖⲴ 6.12 ሿ㢲䇘䇪Ҷ EM ㇇⌅઼ᆳ൘䘹ᤙ␧ਸ儈ᯟ൷٬ᯩ
䶒Ⲵᓄ⭘Ǆ
Kolodner˄1993˅ᨀ‫׋‬ҶሩสҾṸֻⲴ᧘⨶Ⲵа㡜ӻ㓽Ǆԕл᮷⥞㔉ࠪҶަԆⲴаӋ‫ޣ‬
Ҿ䘁ᶕⲴ⹄ウᡀ᷌Ⲵ㓥㿸઼≷䳶˖Aamodt et al.˄1994˅ˈAha et al.˄1991˅ˈHaton et al.˄1995˅ˈ
Riesbeck & Schank˄1989˅
ˈSchank et al.˄1994˅ˈVeloso and Aamodt˄1995˅ˈWatson˄1995˅ˈ
Wess et al.˄1994˅Ǆ
Ґ仈
8.1 ሩҾ‫ޜ‬ᔿ˄8.7˅ѝⲴⴞḷ࠭ᮠⲴањ䐍⿫࣐ᵳተ䜘㓯ᙗ䙬䘁ˈ᧘ሬởᓖл䱽⌅ࡉǄ
8.2 ᙍ㘳ԕлѪ䀓ߣተ䜘࣐ᵳഎᖂѝⲴ䐍⿫ᓖ䟿Ⲵਖа⿽ᯩ⌅Ǆྲлᔪ・ањ㲊ᤏⲴ䇝
㓳ṧֻ䳶ਸDc˖ሩҾ৏࿻䇝㓳ᮠᦞ䳶ਸDѝⲴ⇿ањ䇝㓳ṧֻ<x, f(x)>ˈ൘Dcѝࡋᔪࠪаᇊ
ᮠ䟿˄ਟ㜭ᱟ࠶ᮠ˅Ⲵ<x, f(x)>Ⲵᤧ䍍ˈަѝᤧ䍍Ⲵᮠ䟿ᱟK(d(xq, x))Ǆ⧠൘䇝㓳ањ㓯ᙗ䙬
䘁ᶕᴰሿॆԕл䈟ᐞ߶ࡉ˖
1
E4 { ¦ ( f ( x) fˆ ( x)) 2
2 xDc
䘉䟼Ⲵᜣ⌅ᱟሩ䶐䘁ḕ䈒ᇎֻⲴ䇝㓳ṧֻӗ⭏䖳ཊⲴᤧ䍍ˈ䐍⿫䘌Ⲵᤧ䍍䖳ቁǄ᧘ሬࠪ
䘉њ䈟ᐞ߶ࡉⲴởᓖл䱽⌅ࡉǄᢺ䘉њ⌅ࡉ㺘⽪ᡀ൘ D ⲴᡀઈкⲴ≲઼ˈ㘼нᱟ൘ DcⲴᡀ
ઈк≲઼ˈᒦᢺᆳ઼‫ޜ‬ᔿ˄8.6˅઼˄8.7˅ѝⲴ⌅ࡉ䘋㹼∄䖳Ǆ
8.3 ߣㆆṁᆖҐ㇇⌅ ID3˄㿱ㅜ 3 ㄐ˅ᱟ〟ᶱⲴᆖҐᯩ⌅ˈᨀࠪ䘉⿽㇇⌅Ⲵањ⎸ᶱ⡸

ᵜǄоᵜᶕⲴ〟ᶱ㇇⌅⴨∄ˈ֐Ⲵ⎸ᶱ㇇⌅ᴹӰѸՈ⛩઼㕪⛩˛
৲㘳᮷⥞
ㅢ9ㄖ䚍Ֆ㇍⌋
遗传算法提供了一种大致基于模拟进化的学习方法。其中的假设常被描述为
二进制位串，位串的含义依赖于具体的应用。然而，假设也可以被描述为符号表
达式或者甚至是计算机程序。对合适假设的搜索是从若干初始假设的群体
（population）或汇集（collection）开始的。当前群体的成员通过模仿生物进化
的方式来产生下一代群体，比如说随机变异（mutation）和交叉（crossover）。在
每一步，根据给定的适应度（fitness）度量评估当前群体中的假设，而后使用概
率方法选出适应度最高的假设作为产生下一代的种子。遗传算法已被成功地应用
到多种学习任务和最优化问题中。例如，遗传算法已被用于学习机器人控制的规
则集，以及优化人工神经网络的拓扑结构和学习参数。这一章既覆盖了用位串描
述假设的遗传算法（genetic algorithms），也覆盖了用计算机程序描述假设的遗传
编程（genetic programming）。
9.1 ࣞᵰ
䚇Ր㇇⌅˄GA˅ᨀ‫׋‬Ҷа⿽ਇ⭏⢙䘋ॆ੟ਁⲴᆖҐᯩ⌅Ǆᆳн޽ᱟӾа㡜ࡠ⢩↺ᡆӾ
ㆰঅࡠ༽ᵲൠᩌ㍒‫ٷ‬䇮ˈ㘼ᱟ䙊䗷ਈᔲ઼䟽㓴ᖃࡽᐢ⸕Ⲵᴰྭ‫ٷ‬䇮ᶕ⭏ᡀਾ㔝Ⲵ‫ٷ‬䇮Ǆ൘⇿
а↕ˈ㻛〠Ѫᖃࡽ㗔փ˄population˅Ⲵа㓴‫ٷ‬䇮㻛ᴤᯠˈᯩ⌅ᱟ䙊䗷֯⭘ⴞࡽ䘲ᓄᓖᴰ儈
Ⲵ‫ٷ‬䇮Ⲵਾԓᴯԓ㗔փⲴḀњ䜘࠶Ǆ䘉њ䗷〻ᖒᡀҶሩ‫ٷ‬䇮Ⲵ⭏ᡀᒦ⍻䈅˄generate-and-test˅
ḡ⣦ᩌ㍒˄beam-search˅ˈަѝ㤕ᒢњᴰ֣ᖃࡽ‫ٷ‬䇮Ⲵਈփᴰᴹਟ㜭൘ла↕㻛㘳㲁ǄGA
ⲴᲞ৺઼ਁኅᗇ⳺Ҿԕлഐ㍐˖
x ൘⭏⢙㌫㔏ѝ䘋ॆ㻛䇔Ѫᱟа⿽ᡀ࣏Ⲵ㠚䘲ᓄᯩ⌅ˈᒦфާᴹᖸྭⲴ励ἂ
ᙗǄ
x GA ᩌ㍒Ⲵ‫ٷ‬䇮オ䰤ѝˈ‫ٷ‬䇮Ⲵ਴њ䜘࠶⴨ӂ֌⭘ˈ⇿а䜘࠶ሩᙫⲴ‫ٷ‬䇮䘲
ᓄᓖⲴᖡ૽䳮ԕᔪ⁑Ǆ
x 䚇Ր㇇⌅᱃Ҿᒦ㹼ॆˈфਟ䱽վ⭡Ҿ֯⭘䎵ᕪ䇑㇇ᵪ⺜ԦⲴᑖᶕⲴᰲ䍥䍩
⭘Ǆ
䘉аㄐ᧿䘠Ҷ䚇Ր㇇⌅ˈѮֻ╄⽪ҶᆳⲴ⭘⌅ˈᒦ࠶᷀Ҷᆳᩌ㍒Ⲵ‫ٷ‬䇮オ䰤Ⲵ⢩ᙗǄᡁ
Ԝҏ᧿䘠ҶᆳⲴањਈփˈ〠Ѫ䚇Ր㕆〻ˈ൘䘉⿽ᯩ⌅ѝˈᮤњ䇑㇇ᵪ〻ᒿੁ⵰Ḁњ䘲ᓄᓖ
߶ࡉ䘋ॆǄ䚇Ր㇇⌅઼䚇Ր㕆〻ᱟ䘋ॆ䇑㇇˄evolutionary computation˅亶ฏⲴѝⲴє⿽⍱
㹼ᯩ⌅Ǆ൘ᵜㄐⲴᴰਾа㢲ᡁԜሶ᧕䀖аӋ⹄ウ⭏⢙䘋ॆⲴ䈮仈ˈवᤜ劽ᗧ⑙᭸ᓄ˄Baldwin
effect˅ˈᆳ᧿䘠ҶњփⲴᆖҐ㜭࣋оᮤњ㗔փ䘋ॆ䙏ᓖѻ䰤ᴹ䏓Ⲵ⴨ӂ֌⭘Ǆ
9.2 䚍Ֆ㇍⌋
GA ⹄ウⲴ䰞仈ᱟᩌ㍒ањ‫ى‬䘹‫ٷ‬䇮Ⲵオ䰤ˈԕ⺞ᇊᴰ֣Ⲵ‫ٷ‬䇮Ǆ൘ GA ѝˈĀᴰ֣‫ٷ‬
䇮ā㻛ᇊѹѪᱟ֯Ā䘲ᓄᓖ˄fitness˅āᴰՈⲴ‫ٷ‬䇮ˈ䘲ᓄᓖᱟѪᖃࡽ䰞仈亴‫ݸ‬ᇊѹⲴᮠᆇ
ᓖ䟿Ǆֻྲˈྲ᷌ᆖҐԫ࣑ᱟ൘㔉ᇊањᵚ⸕࠭ᮠⲴ䗃‫ޕ‬䗃ࠪ䇝㓳ṧֻਾ䙬䘁䘉њ࠭ᮠˈ䛓
Ѹ䘲ᓄᓖਟ㻛ᇊѹѪ‫ٷ‬䇮൘䇝㓳ᮠᦞкⲴ㋮ᓖǄྲ᷌ԫ࣑ᱟᆖҐлഭ䱵䊑ỻⲴㆆ⮕ˈ䛓Ѹ䘲
ᓄᓖਟ㻛ᇊѹѪ䈕њփ൘ᖃࡽ㗔փѝоަԆњփሩᔸⲴ㜌⦷Ǆ
ቭ㇑䚇Ր㇇⌅Ⲵн਼ᇎ⧠൘㓶㢲кᴹᡰн਼ˈնᆳԜ䜭ާᴹԕлⲴ‫਼ޡ‬㔃ᶴ˖㇇⌅䘝ԓ
ᴤᯠањ‫ٷ‬䇮⊐ˈ䘉њ‫ٷ‬䇮⊐〠Ѫ㗔փǄ൘⇿а⅑䘝ԓѝˈṩᦞ䘲ᓄᓖ࠭ᮠ䇴ՠ㗔փѝⲴᡰ
ᴹᡀઈǄ❦ਾӾᖃࡽ㗔փѝ⭘ᾲ⦷ᯩ⌅䘹ਆ䘲ᓄᓖᴰ儈Ⲵњփӗ⭏ᯠⲴаԓǄ൘䘉Ӌ㻛䘹ѝ
Ⲵњփѝˈа䜘࠶‫؍‬ᤱ৏ṧൠ䘋‫ޕ‬лаԓ㗔փˈަԆⲴ㻛⭘֌ӗ⭏ਾԓњփⲴส⹰ˈަѝᓄ
⭘䊑Ӕ৹઼ਈᔲ䘉ṧⲴ䚇Րᯩ⌅Ǆ
㺘 9-1 ᧿䘠Ҷањ䚇Ր㇇⌅৏රǄ㇇⌅Ⲵ䗃‫ޕ‬वᤜ˖⭘ᶕᧂᒿ‫ى‬䘹‫ٷ‬䇮Ⲵ䘲ᓄᓖ࠭ᮠ˗
ᇊѹ㇇⌅㓸→ᰦ䘲ᓄᓖⲴ䰸٬˗㾱㔤ᤱⲴ㗔փབྷሿ˗઼ߣᇊྲօӗ⭏ਾ㔗㗔փⲴ৲ᮠ˖⇿а
ԓ㗔փѝ㻛␈⊠Ⲵ∄઼ֻਈᔲ⦷Ǆ
㺘 9-1 䚇Ր㇇⌅৏ර
㇇⌅ѝ㔤ᤱањवਜ਼pњ‫ٷ‬䇮Ⲵ㗔փǄ൘⇿а⅑䘝ԓѝˈਾ㔗㗔փPSⲴᖒᡀ䙊䗷є⿽䙄ᖴ˖ṩ
ᦞ‫ٷ‬䇮Ⲵ䘲ᓄᓖ⭘ᾲ⦷ᯩ⌅䘹ᤙњփˈԕ৺࣐‫ޕ‬ᯠ‫ٷ‬䇮Ǆᯠ‫ٷ‬䇮䙊䗷є⿽ᯩ⌅ᗇࡠ˖ሩᴰ儈䘲ᓄᓖ
‫ٷ‬䇮ሩᓄ⭘Ӕ৹㇇ᆀ˗ሩ䙊䗷䘹ᤙ઼Ӕ৹ӗ⭏Ⲵᯠаԓ㗔փѝⲴ䜘࠶‫ٷ‬䇮䘋㹼অ⛩ਈᔲǄ䟽༽䘉њ
䘝ԓ䗷〻ˈⴤࡠਁ⧠䘲ᓄᓖ䏣ཏྭⲴ‫ٷ‬䇮ǄިරⲴӔ৹઼ਈᔲ㇇ᆀᇊѹ൘ਾ䶒Ⲵ㺘ṬѝǄ
GA(Fitness, Fitness_threshold, p, r, m)
Fitness˖䘲ᓄᓖ䇴࠶࠭ᮠˈѪ㔉ᇊ‫ٷ‬䇮䍻Ҹањ䇴ՠᗇ࠶Ǆ
Fitness_threshold˖ᤷᇊ㓸→ࡔᦞⲴ䰸٬Ǆ
p˖㗔փѝवਜ਼Ⲵ‫ٷ‬䇮ᮠ䟿Ǆ
r˖⇿а↕ѝ䙊䗷Ӕ৹ਆԓ㗔փᡀઈⲴ∄ֻǄ
m˖ਈᔲ⦷Ǆ
x ࡍ࿻ॆ㗔փ˖Pm䲿ᵪӗ⭏Ⲵ p њ‫ٷ‬䇮
x 䇴ՠ˖ሩҾ P ѝⲴ⇿ањ hˈ䇑㇇ Fitness(h)
x ᖃ[ max Fitness(h)]<Fitness_thresholdˈ‫˖ڊ‬
h
ӗ⭏ᯠⲴаԓPS˖
1ˊ 䘹ᤙ˖⭘ᾲ⦷ᯩ⌅䘹ᤙPⲴ(1-r)pњᡀઈ࣐‫ޕ‬PS ǄӾPѝ䘹ᤙ‫ٷ‬䇮hi Ⲵᾲ⦷
Pr(hi)䙊䗷л䶒‫ޜ‬ᔿ䇑㇇˖
Fitness(hi )
Pr(hi )
¦
p
j 1
Fitness(h j )
2ˊ Ӕ৹˖ṩᦞк䶒㔉ࠪⲴPr(hi)ˈӾPѝ᤹ᾲ⦷䘹ᤙrp/2 ሩ‫ٷ‬䇮ǄሩҾ⇿аሩ
‫ٷ‬䇮<h1, h2>ᓄ⭘Ӕ৹㇇ᆀӗ⭏єњਾԓǄᢺᡰᴹⲴਾԓ࣐‫ޕ‬PSǄ
3ˊ ਈᔲ˖֯⭘൷रⲴᾲ⦷ӾPSѝ䘹ᤙmⲮ࠶∄ⲴᡀઈǄሩҾ䘹ࠪⲴ⇿њᡀઈˈ
൘ᆳⲴ㺘⽪ѝ䲿ᵪ䘹ᤙањսਆ৽Ǆ
4ˊ ᴤᯠ˖PmPSǄ
5ˊ 䇴ՠ˖ሩҾ P ѝⲴ⇿ањ h 䇑㇇ Fitness(h)
x Ӿ P ѝ䘄എ䘲ᓄᓖᴰ儈Ⲵ‫ٷ‬䇮Ǆ
൘䘉њ㇇⌅Ⲵ⇿а⅑䘝ԓѝˈสҾᖃࡽⲴ㗔փӗ⭏ᯠаԓⲴ‫ٷ‬䇮Ǆ俆‫ˈݸ‬ӾᖃࡽⲴ㗔փ
ѝ䘹ᤙаᇊᮠ䟿Ⲵ‫ٷ‬䇮वਜ਼൘лаԓѝǄ䘉Ӌ‫ٷ‬䇮ᱟ⭘ᾲ⦷ᯩ⌅䘹ᤙⲴˈަѝ䘹ᤙ‫ٷ‬䇮hiⲴ
ᾲ⦷ᱟ䙊䗷лᔿ䇑㇇Ⲵ˖
Fitness(hi )
Pr(hi ) ˄9.1˅
¦
p
j 1
Fitness(h j )
ഐ↔ˈањ‫ٷ‬䇮㻛䘹ᤙⲴᾲ⦷оᆳ㠚ᐡⲴ䘲ᓄᓖᡀ↓∄ˈᒦфоᖃࡽ㗔փѝަԆㄎҹ‫ٷ‬
䇮Ⲵ䘲ᓄᓖᡀ৽∄Ǆ
൘ᖃࡽԓⲴ䘉Ӌᡀઈᐢ㻛䘹‫ޕ‬лаԓ㗔փਾˈ޽֯⭘а⿽Ӕ৹᫽֌ӗ⭏ަԆⲴᡀઈǄӔ
৹᫽֌ሶ൘ла㢲㻛ާփᇊѹˈᆳӾᖃࡽԓѝਆєњৼӢ‫ٷ‬䇮ˈᒦ䙊䗷䟽ᯠ㓴ਸৼӢⲴ਴䜘
࠶ӗ⭏єњਾԓ‫ٷ‬䇮ǄৼӢ‫ٷ‬䇮ᱟӾᖃࡽ㗔փѝ᤹ᾲ⦷䘹ࠪⲴˈҏ֯⭘‫ޜ‬ᔿ˄9.1˅Ⲵᾲ⦷
࠭ᮠǄ൘䙊䗷䘉⿽Ӕ৹᫽֌ӗ⭏ᯠⲴᡀઈਾˈᯠаԓ㗔փᐢ㓿वਜ਼Ҷᡰ䴰ᮠ䟿ⲴᡀઈǄ᧕л
ᶕˈӾ䘉Ӌᡀઈѝ䲿ᵪ䘹ࠪаᇊ∄ֻ˄m˅ˈᒦ䘋㹼䲿ᵪਈᔲǄ
ഐ↔ˈ䘉⿽ GA ㇇⌅ᢗ㹼а⿽䲿ᵪⲴǃᒦ㹼ḡ⣦‫ٷ‬䇮ᩌ㍒ˈṩᦞ䘲ᓄᓖ࠭ᮠਁ⧠䖳ྭⲴ
‫ٷ‬䇮Ǆ൘л䶒Ⲵሿ㢲ѝᡁԜሶᴤ䈖ቭൠ᧿䘠䘉њ㇇⌅ѝ֯⭘Ⲵ‫ٷ‬䇮㺘⽪઼䚇Ր㇇ᆀǄ
9.2.1 㺞⽰‫ٽ‬䇴
GA ѝ‫ٷ‬䇮㓿ᑨ㻛㺘⽪ѪҼ䘋ࡦսѢˈ䘉ṧਟԕᖸᯩ‫ׯ‬ൠ⭘ਈᔲ઼Ӕ৹䚇Ր㇇ᆀᶕ᫽
֌Ǆ֯⭘䘉ṧⲴսѢ㺘⽪Ⲵ‫ٷ‬䇮ਟ㜭䶎ᑨ༽ᵲǄֻྲˈif-then 㿴ࡉቡਟԕᖸᇩ᱃ൠ⭘䘉⿽ᯩ
ᔿ㺘⽪ˈ‫⌅ڊ‬ᱟ䘹ᤙ㿴ࡉⲴа⿽㕆⸱ˈަѝѪ⇿њ㿴ࡉⲴࡽԦ઼ਾԦ࠶䝽⢩ᇊⲴᆀѢǄ
Holland˄1986˅˗Grefenstette˄1988˅
˗DeJong et al.˄1993˅ѝ᧿䘠Ҷ GA ㌫㔏ѝ䘉⿽㿴ࡉ
㺘⽪ⲴֻᆀǄ
ѪҶ䈤᰾ྲօᢺ if-then 㿴ࡉ㕆⸱ᡀսѢˈ俆‫ݸ‬㘳㲁ᘾṧ֯⭘սѢ᧿䘠অњ኎ᙗⲴ٬㓖

ᶏǄֻྲ㘳㲁኎ᙗ OutlookˈᆳⲴ٬ਟԕਆԕл 3 њ٬ѝⲴԫањ˖SunnyˈOvercast ᡆ RainǄ
㺘⽪ Outlook 㓖ᶏⲴањ᰾ᱮⲴᯩ⌅ᱟˈ֯⭘ањ䮯ᓖѪ 3 ⲴսѢˈ⇿սሩᓄањਟ㜭٬Ǆ
㤕ḀսѪ 1 㺘⽪䘉њ኎ᙗਟԕਆሩᓄⲴ٬ǄֻྲˈѢ 010 㺘⽪ Outlook ᗵ享ਆㅜҼњ٬Ⲵ㓖
ᶏˈᡆ㘵䈤 Outlook=OvercastǄ㊫լⲴˈѢ 011 㺘⽪ᴤа㡜Ⲵ㓖ᶏˈOutlook ਟԕਆєњਟ㜭
٬ˈᡆ㘵䈤˄Outlook=OvercastRain˅
Ǆ⌘᜿ 111 㺘⽪ᴰа㡜Ⲵ㓖ᶏˈ㺘᰾ᡁԜн‫ޣ‬ᗳ䘉њ
኎ᙗਆଚњ٬Ǆ
ᴹҶ㺘⽪অњ኎ᙗ㓖ᶏⲴᯩ⌅ˈ䛓Ѹሩཊњ኎ᙗ㓖ᶏⲴਸਆਟԕᖸᇩ᱃ൠ㺘⽪Ѫሩᓄս
ѢⲴ䘎᧕Ǆֻྲˈ㘳㲁ㅜҼњ኎ᙗ Windˈᆳਟԕਆєњ٬ Strong ᡆ WeakǄ䛓Ѹ‫ۿ‬л䶒Ⲵ㿴
ࡉࡽԦ˖
(Outlook=OvercastRain)(Wind=Strong)
ਟ㻛㺘⽪Ѫ䮯ᓖѪ 5 ⲴսѢ˖
Outlook Wind
011 10
㿴ࡉⲴਾԦ˄ֻྲ PlayTennis=yes˅ਟԕ⭘⴨լⲴᯩᔿ㺘⽪ǄҾᱟˈᮤњ㿴ࡉ㺘⽪ਟԕ
䙊䗷ᢺ᧿䘠㿴ࡉࡽԦ઼ਾԦⲴսѢ䘎᧕䎧ᶕǄֻྲˈл䶒Ⲵ㿴ࡉ
IF Wind=Strong THEN PlayTennis=yes
ሶ㻛㺘⽪ѪԕлⲴսѢ˖
Outlook Wind PlayTennis
111 10 10
ަѝˈࡽйս᧿䘠Ҷሩ Outlook ⲴĀн‫ޣ‬ᗳ˄don’t care˅ā㓖ᶏˈ᧕лᶕєս᧿䘠Ҷሩ

Wind Ⲵ㓖ᶏˈᴰਾєս᧿䘠Ҷ㿴ࡉⲴਾԦ˄䘉䟼‫ٷ‬ᇊ PlayTennis ਟԕਆєњ٬ Yes ᡆ No˅Ǆ
⌘᜿ˈ㺘⽪㿴ࡉⲴսѢሩ‫ٷ‬䇮オ䰤ѝⲴ⇿њ኎ᙗᴹањᆀѢˈণ֯䈕኎ᙗн㻛㿴ࡉⲴࡽԦᡰ
㓖ᶏǄ䘉ṧᗇࡠҶањപᇊ䮯ᓖⲴ㿴ࡉսѢ㺘⽪ˈަѝ൘⢩ᇊս㖞ⲴᆀѢ᧿䘠ሩ⢩ᇊ኎ᙗⲴ
㓖ᶏǄᴹҶঅњ㿴ࡉⲴ㺘⽪ᯩ⌅ˈᡁԜਟԕㆰঅൠᢺঅњ㿴ࡉⲴսѢ㺘⽪䘎᧕䎧ᶕˈӾ㘼㺘
⽪㿴ࡉ䳶Ǆ
൘ѪḀњ‫ٷ‬䇮オ䰤䇮䇑սѢ㕆⸱ᰦˈᴹᗵ㾱䇙⇿њਕ⌅ਸ⌅ⲴսѢ㺘⽪ањᴹ᜿ѹⲴ‫ٷ‬
䇮Ǆ∄ྲˈ㤕֯⭘ка⇥Ⲵ㿴ࡉ㕆⸱ᯩᔿˈ䛓ѸսѢ 111 10 11 㺘⽪Ҷањ㿴ࡉˈᆳⲴਾԦ
н㓖ᶏⴞḷ኎ᙗ PlayTennisǄྲ᷌㾱䚯‫ݽ‬㘳㲁䘉њ‫ٷ‬䇮ˈਟԕ䟷⭘н਼Ⲵ㕆⸱ᯩᔿ˄ֻྲˈ
ӵ࠶䝽ањս㔉ਾԦ PlayTennisˈ㺘⽪ᆳⲴ٬ᱟ Yes ᡆ No˅˗ᡆ᭩ਈ䚇Ր㇇ᆀԕ᰾⺞䚯‫ݽ‬ᔪ
・䘉ṧⲴսѢ˗ᡆᒢ㜶ᢺᖸվⲴ䘲ᓄᓖ䍻㔉䘉ṧⲴѢǄ
൘аӋ GA ѝˈ‫ٷ‬䇮ᱟ⭘ㅖਧ᧿䘠ᶕ㺘⽪Ⲵˈ㘼нᱟ⭘սѢǄֻྲˈ൘ 9.5 㢲ѝˈᡁԜ

䇘䇪Ҷањᢺ‫ٷ‬䇮㕆⸱Ѫ䇑㇇ᵪ〻ᒿⲴ䚇Ր㇇⌅Ǆ
9.2.2 䚍Ֆ㇍ᆆ
൘ GA ѝ䙊䗷а㌫ࡇ㇇ᆀ˄operators˅ᶕߣᇊਾԓˈ㇇ᆀሩᖃࡽ㗔փѝ䘹ᇊⲴᡀઈ䘋㹼
䟽㓴઼ਈᔲǄ㺘 9-1 ѝࡇࠪҶ⭘ᶕ᫽֌սѢⲴިර GA ㇇ᆀǄ䘉Ӌ㇇ᆀᱟ⭏⢙䘋ॆѝⲴ䚇Ր
䗷〻Ⲵ⨶ᜣॆᖒᔿǄᴰᑨ㿱Ⲵєњ㇇ᆀᱟӔ৹˄crossover˅઼ਈᔲ˄mutation˅Ǆ
Ӕ৹㇇ᆀӾєњৼӢѢѝ䙊䗷༽ࡦ䘹ᇊսӗ⭏єњᯠⲴਾԓǄ⇿њਾԓⲴㅜ i սᱟӾᆳ
ⲴḀњৼӢⲴㅜ i ս༽ࡦᶕⲴǄ㠣ҾৼӢѝⲴଚањ൘ㅜ i ս䎧֌⭘ˈ䘉ᱟ⭡ਖཆањ〠Ѫ
Ӕ৹᧙⸱˄crossover mask˅ⲴսѢߣᇊⲴǄл䶒╄⽪ал䘉њ䗷〻ˈ㘳㲁㺘 9-2 ѝᴰк䗩Ⲵ
অ⛩˄single-point˅Ӕ৹㇇ᆀǄ‫ݸ‬㘳㲁ަѝк䶒ањਾԓǄ䘉њਾԓӾㅜањৼӢѝਆࡽ 5
սˈަ։Ⲵ 6 սᶕ㠚ㅜҼњৼӢˈഐѪӔ৹᧙⸱ 11111000000 Ѫ⇿њսᤷᇊ䘉Ӌ䘹ᤙǄㅜҼ
њਾԓ֯⭘਼ṧⲴӔ৹᧙⸱ˈնӔᦒҶৼӢⲴ䀂㢢Ǆᡰԕˈᆳवਜ਼Ҷㅜањਾԓ⋑ᴹ⭘䗷Ⲵ
սǄ൘অ⛩Ӕ৹ѝˈӔ৹᧙⸱ᙫᱟ䘉ṧ㓴ᡀⲴˈᆳԕ䘎㔝Ⲵ n њ 1 ᔰ࿻ˈਾ䶒䐏䲿ᗵ㾱њᮠ
Ⲵ 0 ⴤ㠣㔃ᶏǄ䘉ṧⲴ㔃᷌ᱟਾԓѝࡽ n սᶕ㠚ㅜањৼӢˈ։лⲴսᶕ㠚ㅜҼњৼӢǄ⇿
⅑ᓄ⭘অ⛩Ӕ৹㇇ᆀᰦˈӔ৹⛩ n ᱟ䲿ᵪ䘹ਆⲴˈ❦ਾ޽ӗ⭏Ӕ৹᧙⸱ᒦᓄ⭘Ǆ
㺘 9-2 䚇Ր㇇⌅ᑨ㿱㇇ᆀ
䘉Ӌ㇇ᆀᖒᡀ⭘սѢ㺘⽪Ⲵਾԓ‫ٷ‬䇮ǄӔ৹㇇ᆀӾєњৼӢѝӗ⭏єњਾԓˈ֯⭘Ӕ৹᧙⸱ᶕ
ߣᇊଚањৼӢ֌⭘Ҿ⴨ᓄⲴսǄਈᔲӾঅаⲴৼӢѝӗ⭏অаⲴਾԓˈ䙊䗷䲿ᵪ䘹ਆаսᒦਆ৽Ǆ
initial strings –ࡍ࿻Ѣ
Crossover Mask-Ӕ৹᧙⸱
Offspring-ਾԓ
Single-point crossover:-অ⛩Ӕ৹
Two-point crossover:-є⛩Ӕ৹
Uniform crossover-൷रӔ৹
Point mutation:-⛩ਈᔲ
൘є⛩Ӕ৹˄two-point crossover˅ѝˈਾԓⲴӗ⭏䙊䗷ᢺањৼӢѢⲴѝ䰤⡷⇥ᴯᦒㅜ
ҼњৼӢѢⲴѝ䰤⡷⇥Ǆᦒਕ䈍ᶕ䇢ˈӔ৹᧙⸱ԕn0њ 0 ᔰ࿻ˈਾ䶒䐏䲿n1њ 1ˈ޽䐏䲿ᗵ
㾱ᮠ䟿Ⲵ 0 㔃ᶏǄ⇿⅑ᓄ⭘є⛩Ӕ৹㇇ᆀᰦˈ䙊䗷䲿ᵪ䘹ਆєњᮤᮠn0઼n1ᶕӗ⭏᧙⸱Ǆֻ
ྲˈ൘㺘 9-2 ᱮ⽪Ⲵֻᆀѝˈᱟ֯⭘n0=2 ઼n1=5 Ⲵ᧙⸱ᶕӗ⭏ਾԓⲴǄ઼к䶒аṧˈ䙊䗷
䖜ᦒєњৼӢⲴ䀂㢢ᶕӗ⭏䘉єњਾԓǄ
൷रӔ৹˄uniform crossover˅ਸᒦҶӾєњৼӢԕ൷रᾲ⦷ᣭਆⲴսˈྲ㺘 9-2 ᡰ⽪Ǆ

൘䘉⿽ᛵߥлˈӗ⭏ањ䲿ᵪⲴսѢ֌ѪӔ৹᧙⸱ˈ⇿аսⲴ䘹ਆ䜭ᱟ䲿ᵪⲴᒦф⤜・Ҿަ
ԆսǄ
䲔Ҷ䙊䗷㓴ਸৼӢⲴ਴䜘࠶ӗ⭏ਾԓⲴ䟽㓴㇇ᆀˈਖа⿽㊫රⲴ㇇ᆀӾঅаⲴৼӢӗ⭏
ਾԓǄ⺞࠷ൠ䇢ˈਈᔲ˄mutation˅㇇ᆀ⭘ҾሩսѢӗ⭏䲿ᵪⲴሿਈॆˈᯩ⌅ᱟ䘹ਆањսˈ
❦ਾਆ৽Ǆਈᔲ㓿ᑨᱟ൘ᓄ⭘ҶӔ৹ѻਾ䘋㹼Ⲵˈ‫ۿ‬㺘 9-1 ѝⲴ৏ර㇇⌅䛓ṧǄ
аӋ GA ㌫㔏ᓄ⭘ҶަԆⲴ㇇ᆀˈ⢩࡛ᱟаӋу䰘䪸ሩ㌫㔏ѝ⢩ᇊ‫ٷ‬䇮㺘⽪Ⲵ㇇ᆀǄֻ
ྲˈGrefenstette et al.˄1991˅᧿䘠ҶањᆖҐᵪಘӪ᧗ࡦ㿴ࡉ䳶Ⲵ㌫㔏Ǆᆳ䲔Ҷ֯⭘ਈᔲ
઼Ӕ৹㇇ᆀˈ䘈֯⭘Ҷањ㇇ᆀԕ֯㿴ࡉ⢩ॆǄJanikow˄1993˅᧿䘠ҶањᆖҐ㿴ࡉ䳶Ⲵ
㌫㔏ˈަѝ֯⭘Ҷཊ⿽ⴤ᧕⌋ॆ઼⢩ॆ㿴ࡉⲴ㇇ᆀ˄ֻྲⴤ᧕ᢺањ኎ᙗᶑԦᴯᦒѪĀн‫ޣ‬
ᗳā˄don’t care˅˅Ǆ
9.2.3 䘸ᓊᓜ࠳ᮦૂ‫ٽ‬䇴䘿᤟
䘲ᓄᓖ࠭ᮠᇊѹҶ‫ى‬䘹‫ٷ‬䇮Ⲵᧂ਽߶ࡉˈᒦфᱟԕᾲ⦷ᯩ⌅䘹ᤙлаԓ㗔փⲴ߶ࡉǄྲ
᷌ԫ࣑ᱟᆖҐ࠶㊫Ⲵ㿴ࡉˈ䛓Ѹ䘲ᓄᓖ࠭ᮠѝՊᴹа亩⭘ᶕ䇴ԧ⇿њ㿴ࡉሩ䇝㓳ṧֻ䳶ਸⲴ
࠶㊫㋮ᓖǄ䘲ᓄᓖ࠭ᮠѝҏਟ㜭वਜ਼ަԆⲴ߶ࡉˈֻྲ㿴ࡉⲴ༽ᵲᓖ઼а㡜ᙗ˄generality˅Ǆ
ᴤа㡜ൠ䇢ˈᖃսѢ㻛䀓䟺Ѫ༽ᵲⲴ䗷〻ᰦ˄ֻྲˈᖃսѢ㺘⽪а㌫ࡇ㿴ࡉˈ䘉Ӌ㿴ࡉ㾱㻛
䬮᧕൘а䎧᧗ࡦањᵪಘӪ䇮༷˅ˈ䘲ᓄᓖ࠭ᮠਟԕ⍻䟿⭏ᡀⲴ䗷〻ᙫփᙗ㜭㘼нᱟঅњ㿴
ࡉⲴᙗ㜭Ǆ
൘㺘 9-1 ѝᱮ⽪ⲴGA৏රѝˈ䘹ᤙḀ‫ٷ‬䇮Ⲵᾲ⦷ᱟ䙊䗷ᆳⲴ䘲ᓄᓖоᖃࡽ㗔փѝަԆ
ᡀઈⲴ䘲ᓄᓖⲴ∄٬ᗇࡠⲴˈྲ‫ޜ‬ᔿ˄9.1˅ᡰ⽪Ǆ䘉⿽ᯩ⌅ᴹᰦ㻛〠Ѫ䘲ᓄᓖ∄ֻ䘹ᤙ˄fitness
proportionate selection˅ˈᡆ〠Ѫ䖞ⴈ䍼ķ䘹ᤙ˄roulette wheel selection˅ǄӪԜҏᨀࠪҶަԆ
֯⭘䘲ᓄᓖᶕ䘹ᤙ‫ٷ‬䇮Ⲵᯩ⌅Ǆֻྲ䭖ḷ䎋䘹ᤙ˄tournament selection˅ ˈᆳ‫ݸ‬Ӿᖃࡽ㗔փѝ
䲿ᵪ䘹ਆєњ‫ٷ‬䇮ˈ޽᤹➗һ‫ݸ‬ᇊѹⲴᾲ⦷p䘹ᤙ䘲ᓄᓖ䖳儈Ⲵ‫ٷ‬䇮ˈ᤹➗ᾲ⦷ 1-p䘹ᤙ䘲
ᓄᓖ䖳վⲴ‫ٷ‬䇮Ǆ䭖ḷ䎋䘹ᤙᑨᑨ∄䘲ᓄᓖ∄ֻ⌅ᗇࡠᴤ࣐ཊṧॆⲴ㗔փ˄Goldberg and Deb
1991˅Ǆ൘ਖа⿽㻛〠Ѫᧂ਽䘹ᤙ˄rank selection˅Ⲵᯩ⌅ѝˈᖃࡽ㗔փѝⲴ‫ٷ‬䇮‫᤹ݸ‬䘲ᓄ
ᓖᧂᒿǄ❦ਾˈ䘹ᤙḀ‫ٷ‬䇮Ⲵᾲ⦷оᆳ൘䘉њᧂᒿࡇ㺘ѝⲴս㖞ᡀ∄ֻˈ㘼нᱟоᆳⲴ䘲ᓄ
ᓖᡀ∄ֻǄ
9.3 ⽰ׁ
䚇Ր㇇⌅ਟԕ㻛ⴻ֌䙊⭘ⲴᴰՈॆᯩ⌅ˈᆳᩌ㍒ањᐘབྷⲴ‫ى‬䘹ሩ䊑オ䰤ˈṩᦞ䘲ᓄᓖ
࠭ᮠḕ᢮㺘⧠ᴰྭⲴሩ䊑Ǆቭ㇑н‫؍‬䇱ਁ⧠ᴰՈⲴሩ䊑ˈն GA 㓿ᑨᡀ࣏ൠਁ⧠ާᴹ䖳儈䘲
ᓄᓖⲴሩ䊑ǄGA ᐢ㓿㻛ᓄ⭘ࡠᵪಘᆖҐԕཆⲴབྷ䟿ᴰՈॆ䰞仈ˈवᤜ‫⭥ۿ‬䐟ᐳ㓯઼ԫ࣑䈳
ᓖ䘉ṧⲴ䰞仈Ǆ൘ᵪಘᆖҐ亶ฏˈGA нӵ㻛ᓄ⭘ࡠ࠭ᮠ䙬䘁䰞仈ˈ䘈ᓄ⭘ࡠ‫ۿ‬䘹ਆӪᐕ⾎
㓿㖁㔌Ⲵᤃᢁ㔃ᶴ䘉ṧⲴԫ࣑Ǆ
ѪҶ䈤᰾ GA ൘ᾲᘥᆖҐᯩ䶒Ⲵᓄ⭘ˈᡁԜㆰ㾱ᾲ䘠ал DeJong et al.˄1993˅Ⲵ GABIL

㌫㔏ǄGABIL ֯⭘ GA ᶕᆖҐԕભ仈㿴ࡉⲴ᷀ਆ䳶ਸ㺘⽪ⲴᐳቄᾲᘥǄ൘ሩࠐњᾲᘥᆖҐ
䰞仈Ⲵᇎ傼ѝˈਁ⧠൘⌋ॆ㋮ᓖᯩ䶒 GABIL оަԆⲴᆖҐ㇇⌅བྷփ⴨ᖃˈ䘉䟼ⲴަԆ㇇⌅
वᤜߣㆆṁᆖҐ㇇⌅ C4.5 ઼㿴ࡉᆖҐ㌫㔏 AQ14Ǆ䘉њ⹄ウѝⲴᆖҐԫ࣑ᰒᴹӪѪ䇮䇑Ⲵ
⭘ᶕ⹄ウ㌫㔏⌋ॆ㋮ᓖⲴᆖҐԫ࣑ˈ৸ᴹң㞪ⱼ䇺ᯝ䘉ṧⲴ⧠ᇎ䰞仈Ǆ
GABIL ֯⭘Ⲵ㇇⌅ቡᱟ㺘 9-1 ѝ᧿䘠Ⲵ㇇⌅Ǆ൘ DeJong et al.˄1993˅ᣕ੺Ⲵᇎ傼ѝˈ

ߣᇊ䙊䗷Ӕ৹ᴯᦒ⡦ԓ∄ֻⲴ৲ᮠ r 㻛䇮㖞Ѫ 0.06Ǆߣᇊਈᔲ⦷Ⲵ৲ᮠ m 㻛䇮㖞Ѫ 0.001Ǆ
䘉ᱟ䘉Ӌ৲ᮠⲴިර䇮㖞Ǆ㗔փབྷሿ p Ӿ 100 ࡠ 1000 нㅹˈ㿶⢩ᇊᆖҐԫ࣑㘼ᇊǄ
GA ൘ GABIL ѝⲴާփᓄ⭘ਟԕ㻛ᾲᤜѪԕлࠐ⛩˖
㺘⽪ ൘GABILѝ⇿њ‫ٷ‬䇮ሩᓄҾањભ仈㿴ࡉⲴ᷀ਆ䳶ˈᒦ᤹➗ 9.2.1 㢲᧿䘠Ⲵᯩ

⌅㕆⸱Ǆ⺞࠷ൠ䇢ˈ㿴ࡉࡽԦⲴ‫ٷ‬䇮オ䰤⭡ሩањപᇊⲴ኎ᙗ䳶Ⲵ㓖ᶏⲴਸਆ㓴ᡀˈቡ‫ࡽۿ‬
䶒᧿䘠Ⲵ䛓ṧǄѪҶ㺘⽪㿴ࡉ䳶ˈঅњ㿴ࡉⲴսѢ㺘⽪㻛䘎᧕䎧ᶕǄֻྲˈ㘳㲁䘉ṧањ‫ٷ‬
䇮オ䰤ˈަѝ㿴ࡉⲴࡽԦᱟሩєњᐳቄ኎ᙗa1઼a2Ⲵ㓖ᶏⲴਸਆǄ㿴ࡉⲴਾԦᱟ⭘অњⲴս
᧿䘠Ⲵˈ㺘⽪ⴞḷ኎ᙗcⲴ亴⍻٬ǄҾᱟˈ⭡єњ㿴ࡉ㓴ᡀⲴ‫ٷ‬䇮˖
IF a1=T a2=F THEN c=T˗IF a2=T THEN c=F
ሶ㻛㺘⽪ѪѢ˖
ķ
䈁⌘˖ 䖞ⴈ䍼ᱟᤷа⿽䍼ঊ㘵ᢃ䍼䖜ⴈк᯻䖜Ⲵሿ⨳ሶ‫→ڌ‬Ҿⴈкଚањ‭޵Ⲵ⑨ᠿˈ䘉䟼Ⲵਜ਼ѹ
ᱟᾲ⦷བྷⲴ‫ٷ‬䇮ঐᦞⴈк䖳བྷⲴᡷ४ˈഐ㘼㻛䘹ѝⲴᵪՊ䖳བྷǄ
a1 a2 c a1 a2 c
10 01 1 11 10 0
⌘᜿սѢⲴ䮯ᓖ䲿⵰‫ٷ‬䇮ѝ㿴ࡉⲴᮠ䟿໎䮯Ǆ⭡ҾսѢ䮯ᓖⲴਟਈᙗˈ䴰㾱ሩӔ৹㇇ᆀ
֌ቁ䇨‫؞‬᭩ˈ䘉ሶ൘л䶒᧿䘠Ǆ
䚇Ր㇇ᆀ GABIL֯⭘㺘 9-2 ѝⲴḷ߶ਈᔲ㇇ᆀˈ䲿ᵪ䘹ਆањսˈᒦ⭘ᆳⲴ৽⸱ਆ

ԓ䘉аսǄGABIL֯⭘ⲴӔ৹㇇ᆀᱟ㺘 9-2 ᧿䘠Ⲵє⛩Ӕ৹㇇ᆀⲴањ⴨ᖃḷ߶ⲴᢙኅǄ⺞
࠷ൠ䇢ˈѪҶ䘲ᓄ㕆⸱㿴ࡉ䳶ⲴսѢⲴ䮯ᓖਟਈᙗˈᒦф䲀ࡦ㌫㔏ԕ֯Ӕ৹ӵਁ⭏൘սѢⲴ
⴨լ⡷⇥䰤ˈ䟷ਆҶл䶒Ⲵ࣎⌅Ǆ俆‫ݸ‬൘ㅜањৼӢѢк䲿ᵪ䘹ਆєњӔ৹⛩ˈᆳԜѻ䰤ࡂ
࠶ࠪҶањսѢ⡷⇥Ǆ⭡ҾսѢ㺘⽪Ⲵᱟањ㿴ࡉ䳶ˈᡁԜਟԕḷ䇠ࠪަѝ⇿њ㿴ࡉⲴ䗩⭼Ǆ
䘉њսѢ⡷⇥ਟ㜭䐘䎺㤕ᒢњ㿴ࡉ䗩⭼Ǆ❦ਾԔd1㺘⽪⡷⇥Ⲵᴰᐖаսࡠᆳᐖ‫ח‬ㅜањ㿴ࡉ
䗩⭼Ⲵ䐍⿫Ǆd2㺘⽪⡷⇥Ⲵᴰਣаսࡠᆳᐖ‫ח‬ㅜањ㿴ࡉ䗩⭼Ⲵ䐍⿫Ǆ᧕лᶕˈ൘ㅜҼњৼ
Ӣк䲿ᵪ䘹ਆӔ৹⛩ˈਚ㾱䘹ᤙⲴӔ৹⛩ާᴹ਼ṧⲴd1઼d2٬Ǆֻྲˈྲ᷌єњৼӢѢᱟ
a1 a2 c a1 a2 c
h1˖ 10 01 1 11 10 0
a1 a2 c a1 a2 c
h2˖ 01 11 0 10 01 0
ᒦфѪㅜањৼӢ䘹ਆӔ৹⛩ս㖞ᱟㅜ 1 ઼ㅜ 8 սˈྲлᡰ⽪˖
a1 a2 c a1 a2 c
h1˖ 1[0 01 1 11 1]0 0
ަѝĀ[ā઼Ā]ā㺘⽪Ӕ৹⛩ˈ䛓Ѹd1=1 ᒦфd2=3Ǆᡰԕˈ‫ݱ‬䇨䘹ਆⲴㅜҼњৼӢӔ৹⛩Ⲵ
ս㖞ᴹ<1, 3>ˈ<1, 8>઼<6, 8>Ǆྲ᷌ᚠᐗ䘹ਆҶ<1ˈ3>ˈ
a1 a2 c a1 a2 c
h2˖ 0[1 1]1 0 10 01 0
䛓Ѹ㔃᷌⭏ᡀⲴєњਾԓᱟ
a1 a2 c
h3˖ 11 10 0
a1 a2 c a1 a2 c a1 a2 c
h4˖ 00 01 1 11 11 0 10 01 0
ྲ↔ֻᡰ⽪ˈ䘉⿽Ӕ৹ᯩ⌅ѝਾԓਟԕवਜ਼оৼӢн਼ᮠ䟿Ⲵ㿴ࡉˈ਼ᰦ‫؍‬䇱Ҷ᤹䘉⿽ᯩᔿ
ӗ⭏ⲴսѢ㺘⽪㢟ᇊѹⲴ˄well-defined˅㿴ࡉ䳶Ǆ
䘲ᓄᓖ࠭ᮠ ⇿њ㿴ࡉ䳶Ⲵ䘲ᓄᓖᱟṩᦞᆳ൘䇝㓳ᮠᦞкⲴ࠶㊫㋮ᓖ䇑㇇ⲴǄ⺞࠷ൠ䇢ˈ
ᓖ䟿䘲ᓄᓖⲴ࠭ᮠᱟ˖
Fitness(h)=(correct(h))2
ަѝˈcorrect(h)ᱟ‫ٷ‬䇮 h ࠶㊫ᡰᴹ䇝㓳ṧֻⲴ↓⺞⦷Ǆ
൘∄䖳 GABIL ઼‫ ۿ‬C4.5 ઼ ID5R 䘉ṧⲴߣㆆṁᆖҐ㇇⌅ԕ৺㿴ࡉᆖҐ㇇⌅ AQ14 Ⲵᇎ
傼ѝˈṩᦞሩн਼ᆖҐԫ࣑Ⲵ⍻䈅ˈDeJong et al.˄1993˅ᣕ੺Ҷ䘉Ӌ㌫㔏ާᴹབྷփ⴨ᖃⲴ
ᙗ㜭ǄֻྲˈሩӪѪ䇮䇑Ⲵ 12 њ䰞仈ˈGABIL 䗮ࡠҶ 92.1%Ⲵᒣ൷⌋ॆ㋮ᓖˈ㘼ަԆ㌫㔏
ⲴⲴᙗ㜭ᱟ൘ 91.2%ࡠ 96.6%ѻ䰤Ǆ
9.3.1 ᢟኋ
DeJong et al.˄1993˅ѝҏ᧒㍒Ҷሩ GABIL สᵜ䇮䇑Ⲵєњᴹ䏓ⲴᢙኅǄ൘а㓴ᇎ傼ѝˈ

ԆԜ⹄ウҶਖཆєњᯠⲴ䚇Ր㇇ᆀˈ䘉єњ㇇ᆀਇࡠҶᖸཊㅖਧᆖҐᯩ⌅ѝᑨ㿱Ⲵ⌋ॆ㇇ᆀ
Ⲵ੟ਁǄㅜањ㇇ᆀѪ AddAlternativeˈᆳ⌋ॆሩḀњ⢩ᇊ኎ᙗⲴ㓖ᶏˈᯩ⌅ᱟᢺ䘉њ኎ᙗ
ሩᓄⲴᆀѢѝⲴањ 0 ᭩Ѫ 1Ǆֻྲˈྲ᷌ањ኎ᙗⲴ㓖ᶏ֯⭘Ѣ 10010 㺘⽪ˈ䛓Ѹ䘉њ㇇
ᆀਟ㜭ᢺᆳ᭩Ѫ 10110Ǆ䘉њ㇇ᆀ൘⇿аԓ㗔փѝሩ䘹ᇊⲴᡀઈ᤹➗ 0.01 Ⲵᾲ⦷ᓄ⭘ǄㅜҼ
њ㇇ᆀѪ DropConditionˈᆳ䟷⭘а⿽ᴤ࣐ᶱㄟⲴ⌋ॆ᧚ᯭˈᢺањ⢩ᇊ኎ᙗⲴᡰᴹս䜭ᴯ
ᦒѪ 1Ǆ䘉њ㇇ᆀ⴨ᖃҾ䙊䗷ᆼ‫ޘ‬᫔䬰኎ᙗ㓖ᶏᶕ⌋ॆ㿴ࡉˈᆳ᤹➗ᾲ⦷ 0.60 ൘⇿аԓѝᓄ
⭘ǄDeJong et al.˄1993˅ѝᣕ੺Ҷ䘉њ᭩䘋Ⲵ㌫㔏ሩҾк䶒ᡰ䈤ⲴӪѪ䇮䇑ԫ࣑䗮ࡠҶ 95.2%
Ⲵᒣ൷⌋ॆ㋮ᓖˈ⴨∄䖳สᵜⲴ GA Ѫ 92.1%Ǆ
൘к䶒Ⲵᇎ傼ѝˈєњ㇇ᆀሩ⇿аԓ㗔փѝⲴ⇿њ‫ٷ‬䇮ᱟԕ਼ṧⲴᾲ⦷ᓄ⭘ⲴǄ൘ਖа
њᇎ傼ѝˈሩ‫ٷ‬䇮ⲴսѢ㺘⽪䘋㹼Ҷᢙኅˈ֯ަवਜ਼ਖཆєսԕߣᇊᱟ੖ਟԕሩ䈕‫ٷ‬䇮ᓄ⭘
䘉єњ㇇ᆀǄ൘䘉њᢙኅⲴ㺘⽪ѝˈањިරⲴ㿴ࡉ䳶‫ٷ‬䇮ⲴսѢѪ
a1 a2 c a1 a2 c AA DC
01 11 0 10 01 0 1 0
ަѝᴰਾⲴєњս㺘⽪൘䘉⿽ᛵߥлਟԕሩ䈕Ѣᓄ⭘ AddAlternative ㇇ᆀˈ㘼нਟԕᓄ⭘
DropCondition ㇇ᆀǄ䘉єњᯠⲴսᇊѹҶ䜘࠶Ⲵ GA ᩌ㍒ㆆ⮕ˈ㘼фᆳԜᵜ䓛ҏ઼ѢѝⲴ
ަԆսа䎧㻛਼ṧⲴӔ৹઼ਈᔲ㇇ᆀ‫؞‬᭩઼䘋ॆǄDeJong et al.˄1993˅ᣕ੺Ҷ䘉⿽ᯩ⌅Ⲵ
㔃᷌Ո࣓৲ॺ˄ҏቡᱟሩḀӋ䰞仈ᨀ儈Ҷᙗ㜭ˈሩަԆ䰞仈䱽վҶᙗ㜭˅ˈᆳֻ⽪Ҷ GA ൘
৏ࡉкᱟྲօ֯ަ‫ٷ‬䇮Ⲵᩌ㍒ᯩ⌅䘋ॆⲴǄ
9.4 ‫ٽ‬䇴グ䰪ᩒ㍘
ྲкᡰ⽪ˈGA 䟷⭘а⿽䲿ᵪॆⲴḡ⣦ᩌ㍒ᶕራ᢮ᴹᴰབྷ䘲ᓄᓖⲴ‫ٷ‬䇮Ǆ䘉⿽ᩌ㍒оᵜ
Җѝᐢ㘳㲁ⲴަԆᆖҐᯩ⌅Ⲵᩌ㍒ᆼ‫ޘ‬н਼Ǆֻྲˈ∄䖳 GA ֯⭘Ⲵᩌ㍒オ䰤઼⾎㓿㖁㔌৽
ੁՐ᫝㇇⌅֯⭘Ⲵᩌ㍒オ䰤˖൘৽ੁՐ᫝㇇⌅ѝˈởᓖл䱽ᩌ㍒Ӿањ‫ٷ‬䇮ᒣ━〫ࣘࡠањ
䶎ᑨ⴨լⲴᯠ‫ٷ‬䇮Ǆо↔н਼ˈGA ᩌ㍒Ⲵ〫ࣘਟ㜭䶎ᑨケ❦ˈ֯⭘઼ৼӢṩᵜн਼Ⲵਾԓ
ᴯᦒৼӢ‫ٷ‬䇮Ǆ⌘᜿ GA ᩌ㍒ഐ↔нཚਟ㜭‫ۿ‬ởᓖл䱽ᯩ⌅䛓ṧާᴹ䲧‫ޕ‬ተ䜘ᴰሿ٬Ⲵ䰞
仈Ǆ
൘аӋ GA ᓄ⭘ѝˈањᇎ䐥кⲴ䳮仈ᱟᤕᥔ˄crowding˅䰞仈Ǆᤕᥔᱟ䘉ṧа⿽⧠䊑ˈ
㗔փѝḀањփ䘲ᓄᓖབྷབྷ儈ҾަԆњփˈഐ↔ᆳ䗵䙏㑱⇆ˈԕ㠣Ҿ䘉њњփ઼оᆳ⴨լⲴ
њփঐᦞҶ㗔փⲴ㔍བྷ䜘࠶ǄᤕᥔⲴн㢟ᖡ૽ᱟ䱽վҶ㗔փⲴཊṧᙗ˄diversity˅ˈӾ㘼߿ធ
Ҷ GA Ⲵ䘋а↕䘋ॆǄӪԜᐢ㓿᧒㍒Ҷ㤕ᒢ䱽վᤕᥔⲴㆆ⮕Ǆа⿽ᯩ⌅ᱟ‫؞‬᭩䘹ᤙ࠭ᮠˈ֯
⭘‫ۿ‬䭖ḷ䎋䘹ᤙᡆᧂ਽䘹ᤙ䘉ṧⲴ߶ࡉਆԓ䘲ᓄᓖ∄ֻ䖞ⴈ䍼䘹ᤙǄањ⴨‫Ⲵޣ‬ㆆ⮕ᱟĀ䘲
ᓄᓖ‫ޡ‬ӛ˄fitness sharing˅
āˈަѝṩᦞ㗔փѝоḀњփ⴨լⲴњփᮠ䟿ˈ߿ሿ䈕њփⲴ䘲ᓄ
ᓖǄㅜй⿽ᯩ⌅ᱟሩਟ䟽㓴⭏ᡀਾԓⲴњփ⿽㊫䘋㹼䲀ࡦǄֻྲˈ䙊䗷ਚ‫ݱ‬䇨ᴰ⴨լⲴњփ
䟽㓴ˈਟԕ൘㗔փѝ‫׳‬ᡀ⴨լⲴњփ㚊㊫ˈᡆཊњӊ⿽˄subspecies˅Ǆа⿽⴨‫Ⲵޣ‬ᯩ⌅ᱟ᤹
オ䰤࠶ᐳњփˈᒦфӵ‫ݱ‬䇨⴨䛫Ⲵњփ䟽㓴Ǆ䘉Ӌᢰᵟᖸཊ䜭ᱟਇࡠҶ⭏⢙䘋ॆⲴ੟⽪Ǆ
9.4.1 㗚։䘑ौૂ⁗ᕅ⨼䇰
н࿘ᙍ㘳ањᴹ䏓Ⲵ䰞仈˖ᱟ੖㜭⭘ᮠᆖⲴᯩ⌅࡫⭫ GA ѝ㗔փ䲿ᰦ䰤䘋ॆⲴ䗷〻˛
Holland˄1975˅Ⲵ⁑ᔿ৏⨶˄schema theorem˅ᨀ‫׋‬Ҷа⿽࡫⭫ᯩ⌅ǄᆳสҾ᧿䘠սѢ䳶ਸ
Ⲵ⁑ᔿ˄schemaˈᡆ pattern˅Ǆ㋮⺞ൠ䇢ˈањ⁑ᔿᱟ⭡㤕ᒢ 0ǃ1 ઼*㓴ᡀⲴԫ᜿ѢǄĀ*ā
㺘⽪ањн‫ޣ‬ᗳⲴսǄֻྲ⁑ᔿ 0*10 㺘⽪ⲴսѢ䳶ਸѝਚवਜ਼ 0010 ઼ 0110Ǆ
অњսѢਟԕ㻛ⴻ֌оᆳ३䝽Ⲵ⇿њ⁑ᔿⲴԓ㺘ǄֻྲˈսѢ 0010 ਟԕ㻛䇔Ѫ 24њ⴨

ᔲ⁑ᔿⲴԓ㺘ˈֻྲ 00**ˈ0*10ˈ****ㅹǄ㊫լൠˈањսѢⲴ㗔փਟԕ㻛ⴻ֌˖սѢᡰ
ԓ㺘Ⲵ⁑ᔿⲴ䳶ਸˈԕ৺о⇿њ⁑ᔿ‫ޣ‬㚄Ⲵњփᮠ䟿Ǆ
⁑ᔿ⨶䇪ṩᦞ⇿њ⁑ᔿⲴᇎֻᮠ䟿ᶕ࡫⭫ GA ѝ㗔փⲴ䘋ॆǄԔ m(s, t)㺘⽪㗔փѝⲴ⁑

ᔿ s ൘ᰦ䰤 t˄ҏቡᱟ൘ㅜ t ԓᵏ䰤˅Ⲵᇎֻᮠ䟿Ǆ⁑ᔿ⨶䇪ṩᦞ m(s, t)઼⁑ᔿǃ㗔փ৺ GA
৲ᮠⲴަԆ኎ᙗˈᶕ᧿䘠 m(s, t+1)Ⲵᵏᵋ٬Ǆ
GAѝ㗔փⲴ䘋ॆ‫׍‬䎆Ҿࠐњ↕僔ˈণ䘹ᤙ↕ǃ䟽㓴↕઼ਈᔲ↕Ǆ‫ݸ‬Ӿਚ㘳㲁䘹ᤙ↕Ⲵ
ᖡ૽ᔰ࿻Ǆ֯⭘f(h)㺘⽪սѢњփhⲴ䘲ᓄᓖˈᒦ⭘ f (t)㺘⽪൘ᰦ䰤t㗔փѝᡰᴹњփⲴᒣ൷
䘲ᓄᓖǄ䇮nѪ㗔փѝњփⲴᙫᮠ䟿Ǆ֯⭘hspt㺘⽪њփhᰒᱟ⁑ᔿsⲴањԓ㺘ˈ৸ᱟᰦ
䰤t㗔փⲴањᡀઈǄᴰਾˈԔ uˆ (s, t)㺘⽪൘ᰦ䰤t㗔փѝ⁑ᔿsⲴᇎֻⲴᒣ൷䘲ᓄᓖǄ
ᡁԜᝏ‫ޤ‬䏓Ⲵᱟ m(s, t+1)Ⲵᵏᵋ٬ˈ⭘ E[m(s, t+1)]ᶕ㺘⽪Ǆਟԕ֯⭘‫ޜ‬ᔿ˄9.1˅ѝ㔉

ࠪⲴᾲ⦷࠶ᐳᶕ䇑㇇ E[m(s, t+1)]ˈᒦ֯⭘ⴞࡽⲴㅖਧᢺᆳ䟽ᯠ㺘⽪ᡀྲлᖒᔿ˖
f ( h)
Pr(h)
¦
n
i 1
f (hi )
f (h)
nf (t )
⧠൘ྲ᷌ṩᦞ䘉њᾲ⦷࠶ᐳ䘹ᤙᯠ㗔փⲴањᡀઈˈ䛓Ѹ䘹ࡠ⁑ᔿ s Ⲵањԓ㺘Ⲵᾲ⦷
ᱟ˖
f ( h)
Pr( h s ) ¦
hs pt nf (t )
˄9.2˅
uˆ ( s, t )
m( s, t )
nf (t )
к䶒ⲴㅜҼ↕ṩᦞԕлⲴᇊѹ㘼ᗇ˖
uˆ ( s, t )
¦ hs pt
f ( h)
m( s , t )
‫ޜ‬ᔿ˄9.2˅㔉ࠪҶ GA 䘹ᤙⲴањ‫ٷ‬䇮ᱟ⁑ᔿ s ⲴᇎֻⲴᾲ⦷ǄᡰԕˈሩҾӗ⭏ᮤњ

ᯠаԓⲴ n ⅑⤜・䘹ᤙ↕ˈᗇࡠⲴ s ⲴᇎֻⲴᵏᵋᮠ䟿ቡᱟ䘉њᾲ⦷Ⲵ n ‫ؽ‬Ǆ
uˆ ( s, t )
E[m(s, t+1)]= m(s, t) ˄9.3˅
f (t )
‫ޜ‬ᔿ˄9.3˅㺘᰾ˈ൘ t+1 ԓѝ⁑ᔿ s Ⲵᇎֻᵏᵋᮠ䟿ˈо൘ᰦ䰤 t 䘉њ⁑ᔿⲴᇎֻⲴᒣ
൷䘲ᓄᓖ û (s, t)ᡀ↓∄ˈᒦоᰦ䰤 t ѝ㗔փⲴᡰᴹᡀઈⲴᒣ൷䘲ᓄᓖ f (t)ᡀ৽∄Ǆഐ↔ˈᡁ
Ԝਟԕᵏᵋˈ൘ਾ㔗Ⲵ਴ԓѝ儈Ҿᒣ൷䘲ᓄᓖⲴ⁑ᔿࠪ⧠仁⦷Պॷ儈Ǆྲ᷌ᡁԜᢺ GA ⴻ֌ˈ
൘ሩњփオ䰤䘋㹼ᱮᔿᩌ㍒Ⲵ਼ᰦˈሩਟ㜭⁑ᔿオ䰤䘋㹼⵰㲊ᤏⲴᒦ㹼ᩌ㍒ˈ䛓Ѹ‫ޜ‬ᔿ˄9.3˅
ᤷࠪ䘲ᓄᓖ儈Ⲵ⁑ᔿⲴᖡ૽࣋Պ䲿⵰ᰦ䰤໎࣐Ǆ
❦㘼к䶒Ⲵ࠶᷀ӵ㘳㲁 GA ѝ䘹ᤙ↕Ⲵᖡ૽ˈᡰԕҏᓄ䈕㘳㲁Ӕ৹઼ਈᔲ↕Ⲵᖡ૽Ǆ⁑
ᔿ⨶䇪ӵ㘳㲁䘉Ӌ㇇ᆀਟ㜭䙐ᡀⲴ䍏䶒ᖡ૽˄ֻྲˈ䲿ᵪਈᔲਟ㜭䱽վ s Ⲵԓ㺘ᮠ䟿ˈ⤜・
Ҿ uˆ (s,t)˅ˈᒦфӵ㘳㲁অ⛩Ӕ৹ⲴᛵߥǄᡰԕᆼᮤⲴ⁑ᔿ⨶䇪㔉ࠪҶ⁑ᔿ s Ⲵᵏᵋ仁⦷Ⲵл
⭼ˈྲлᡰ⽪˖
uˆ ( s, t ) § d (s) ·
E[m( s, t 1)] t m( s, t )¨1 pc ¸(1 p m )
o( s )
˄9.4˅
f (t ) © l 1 ¹
䘉䟼ˈpcᱟሩԫ᜿њփᓄ⭘অ⛩Ӕ৹㇇ᆀⲴᾲ⦷ˈpmᱟሩԫ᜿њփⲴԫ᜿ս֯⭘ਈᔲ㇇
ᆀ䘋㹼ਈᔲⲴᾲ⦷Ǆo(s)ᱟ⁑ᔿsѝ⺞ᇊս˄defined bits˅Ⲵњᮠˈ0 ઼ 1 ᱟ⺞ᇊⲴսˈ*нᱟǄ
d(s)ᱟ⁑ᔿsѝᴰᐖ䗩Ⲵ⺞ᇊս઼ᴰਣ䗩Ⲵ⺞ᇊս䰤Ⲵ䐍⿫Ǆᴰਾˈlᱟ൘㗔փѝњփսѢ䮯
ᓖǄ⌘᜿ˈ‫ޜ‬ᔿ˄9.4˅ѝⲴᴰᐖа亩о‫ޜ‬ᔿ˄9.3˅ᱟаṧⲴˈ䘉а亩᧿䘠Ҷ䘹ᤙ↕僔Ⲵᖡ
૽Ǆѝ䰤а亩᧿䘠Ҷঅ⛩Ӕ৹㇇ᆀⲴᖡ૽ˈ⢩࡛٬ᗇ⌘᜿Ⲵᱟˈ䘉а亩᧿䘠Ҷԓ㺘sⲴԫ᜿
њփ൘ᓄ⭘ҶӔ৹㇇ᆀਾ䘈㺘⽪sⲴᾲ⦷Ǆᴰਣа亩᧿䘠Ҷԓ㺘⁑ᔿsⲴԫ᜿њփ൘ᓄ⭘Ҷਈ
ᔲ㇇ᆀਾ䘈㺘⽪sⲴᾲ⦷Ǆঅ⛩Ӕ৹઼ਈᔲⲴᖡ૽䲿⵰⁑ᔿѝ⺞ᇊսⲴᮠ䟿o(s)઼⺞ᇊս䰤Ⲵ
䐍⿫d(s)໎䮯Ǆഐ↔ˈਟԕ⁑ᔿ⨶䇪ਟԕ㻛㋇⮕ൠ䀓䟺Ѫ˖ᴤ儈䘲ᓄᓖⲴ⁑ᔿⲴᖡ૽࣋䎻ੁ
໎བྷˈቔަᱟवਜ਼䖳ቁᮠ䟿Ⲵ⺞ᇊս˄ҏቡᱟवਜ਼བྷ䟿Ⲵ*˅Ⲵ⁑ᔿ઼䘉Ӌ⺞ᇊս൘սѢѝ
ᖬ↔䶐䘁Ⲵ⁑ᔿǄ
ሩ GA ѝⲴ㗔փ䘋ॆ䗷〻ˈ⁑ᔿ⨶䇪ਟ㜭ᱟ㻛ᕅ⭘ᗇᴰᒯ⌋Ⲵ࡫⭫ᯩᔿǄᆳнᆼ༷Ⲵа
䶒ᱟᰐ⌅㘳㲁Ӕ৹઼ਈᔲⲴ˄བྷᾲⲴ˅↓䶒ᖡ૽Ǆᴰ䘁ӪԜᐢ㓿ᨀࠪҶᖸཊᯠⲴ⨶䇪࠶᷀ˈ
वᤜสҾ傜ቄਟཛ䬮⁑ර˄Markov chain model˅઼㔏䇑࣋ᆖ⁑ර˄statistical mechanics
models˅Ⲵ࠶᷀Ǆֻྲਟԕ৲㿱 Whitley & Vose˄1995˅઼ Mitchell˄1996˅Ǆ
9.5 䚍Ֆ㕌ぁ
䚇Ր㕆〻˄Genetic ProgrammingˈGP˅ᱟ䘋ॆ䇑㇇Ⲵа⿽ᖒᔿˈަѝ䘋ॆ㗔փѝⲴњ
փᱟ䇑㇇ᵪ〻ᒿ㘼нᱟսѢǄKoza˄1992˅᧿䘠ҶสᵜⲴ䚇Ր㕆〻ᯩ⌅ᒦф㔉ࠪҶᖸཊㆰ
অⲴਟԕ㻛 GP ᡀ࣏ᆖҐⲴ〻ᒿǄ
9.5.1 ぁᓅ㺞⽰
GP ᫽֌Ⲵ〻ᒿа㡜㻛㺘⽪Ѫ〻ᒿⲴ䀓᷀˄parse˅ṁǄ⇿њ࠭ᮠ䈳⭘㻛㺘⽪ѪṁⲴањ
㢲⛩ˈ࠭ᮠⲴ৲ᮠ䙊䗷ᆳⲴᆀ㔃⛩㔉ࠪǄֻྲˈമ 9-1 ⭫ࠪҶ࠭ᮠ sin(x)+ x y Ⲵṁ㺘⽪Ǆ

2
ѪҶᓄ⭘䚇Ր㕆〻ࡠḀњ⢩ᇊⲴ亶ฏˈ⭘ᡧᗵ享ᇊѹᖵ㘳㲁Ⲵ৏ᆀ࠭ᮠ˄primitive functions˅
˄ֻྲˈsinˈcosˈᔰᯩˈ+ˈ-ˈᤷᮠ˅ˈԕ৺ㄟ⛩˄terminals˅
˄ֻྲ xǃy ԕ৺ᑨᮠ˅Ǆ᧕л
ᶕˈ䚇Ր㕆〻㇇⌅֯⭘䘋ॆᩌ㍒ˈᶕ᧒㍒֯⭘䘉Ӌ৏ᆀ᧿䘠Ⲵ〻ᒿⲴᐘབྷオ䰤Ǆ
о൘䚇Ր㇇⌅ѝаṧˈ৏රⲴ䚇Ր㕆〻㇇⌅㔤ᣔ⭡ཊњњփ˄൘䘉䟼ᱟ〻ᒿṁ˅㓴ᡀⲴ
㗔փǄ൘⇿а↕䘝ԓѝˈᆳ֯⭘䘹ᤙǃӔ৹઼ਈᔲӗ⭏ᯠаԓњփǄ㗔փѝḀњփ〻ᒿⲴ䘲
ᓄᓖа㡜䙊䗷൘䇝㓳ᮠᦞкᢗ㹼䘉њ〻ᒿᶕߣᇊǄӔ৹᫽֌ᱟ䘉ṧ䘋㹼Ⲵ˖൘ањৼӢ〻ᒿ
ѝ䲿ᵪ䘹ᤙањᆀṁˈ❦ਾ⭘ਖањৼӢⲴᆀṁᴯԓ䘉њᆀṁǄമ 9-2 ╄⽪ҶањިරⲴӔ
৹᫽֌Ǆ
മ 9-1 䚇Ր㕆〻ѝⲴ〻ᒿṁ㺘⽪
ԫ᜿〻ᒿਟԕ㺘⽪ѪᆳԜⲴ䀓᷀ṁǄ
മ 9-2 ሩєњৼӢ〻ᒿṁ䘋㹼Ӕ৹᫽֌
ৼӢ〻ᒿṁᱮ⽪൘кᯩˈᆙᆀṁ൘лᯩǄӔ৹⛩˄к䗩࣐㋇ᱮ⽪Ⲵ㢲⛩˅ᱟ䲿ᵪ䘹ਆⲴǄ❦ਾԕ䘉
ӋӔ৹⛩ѪṩⲴᆀṁӂᦒԕӗ⭏ᆙᆀṁǄ
Koza˄1992˅᧿䘠Ҷᓄ⭘ GP ࡠཊњԫ࣑Ⲵᇎ傼Ǆ൘ԆⲴᇎ傼ѝˈṩᦞ䘲ᓄᓖᾲ⦷䘹ᤙ
ᖃࡽ㗔փⲴ 10%н࣐᭩ਈⲴ‫ࡠ⮉؍‬лаԓǄ޽ṩᦞ䘲ᓄᓖᾲ⦷Ӿᖃࡽ㗔փѝ䘹ᤙ〻ᒿሩˈ
ᓄ⭘Ӕ৹᫽֌ӗ⭏ᯠаԓⲴަ։䜘࠶Ǆ൘䘉њᇎ傼㌫ࡇѝ⋑ᴹ֯⭘ਈᔲ㇇ᆀǄ
9.5.2 ⽰ׁ
Koza˄1992˅㔉ࠪⲴањ⽪ֻᱟᆖҐањ㇇⌅ᶕึ⸼മ 9-3 ᡰ⽪ⲴᆇඇǄ䘉њԫ࣑ᱟᔰ

ਁањ䙊⭘Ⲵ㇇⌅ᶕᢺᆇඇึਐᡀঅњḸ˄stack˅ˈ᤬ࠪঅ䇽Āuniversalāˈᰐ䇪䘉Ӌᆇඇࡍ
࿻Ⲵ㔃ᶴྲօǄਟᢗ㹼Ⲵࣘ֌ᱟ⇿⅑ਚ‫ݱ‬䇨〫ࣘањᆇඇǄ⺞࠷ൠ䇢ˈ൘Ḹѝᴰк䶒Ⲵᆇඇ
ਟԕ㻛〫ࡠṼ䶒кˈᡆ㘵Ṽ䶒кⲴᆇඇਟԕ㻛〫ࣘࡠḸ亦Ǆ
മ 9-3 ањᆇඇึਐ䰞仈
䘉њ䚇Ր㕆〻Ⲵԫ࣑ᱟਁ⧠ањ〻ᒿˈਟԕᢺᴹԫ᜿ࡍ࿻㔃ᶴⲴᆇඇਈᦒᡀањḸ᤬ࠪঅ䇽
ĀuniversalāǄᒦᨀ‫׋‬Ҷ 166 ⿽ࡍ࿻㔃ᶴᶕ䇴ՠ‫ى‬䘹〻ᒿⲴ䘲ᓄᓖ˄᪈㠚 Koza 1992˅Ǆ
൘བྷཊᮠ GP ᓄ⭘ѝˈ䰞仈㺘⽪ᯩ⌅Ⲵ䘹ᤙሩҾ亪࡙ൠ䀓ߣ䰞仈䎧⵰䶎ᑨ䟽㾱Ⲵ֌⭘Ǆ
൘ Koza Ⲵ䇮䇑ѝˈ⭘ԕ㓴ᡀ〻ᒿⲴ৏ᆀ࠭ᮠवਜ਼л䶒Ⲵйњㄟ⛩৲ᮠ˖
x CS˄current stack˅˖ᖃࡽḸˈᤷḸ亦ᆇඇⲴ਽ᆇˈᡆ⋑ᴹᖃࡽḸᰦѪ FǄ
x TB˄top correct bloack˅˖ᴰк↓⺞ᆇඇˈᤷ䈕ᆇඇ઼ᆳԕлᆇඇ൷Ѫ↓⺞
亪ᒿⲴᆇඇǄ
x NN˄next necessary˅˖лањᡰ䴰ᆇඇˈᤷѪҶ᤬ᡀঅ䇽ĀuniversalāˈḸ
޵㍗䛫 TB ѻкⲴᡰ䴰ᆇඇⲴ਽ᆇˈᡆ㘵ᖃн޽䴰㾱ᆇඇᰦѪ FǄ
ਟԕⴻࠪˈ䘹ᤙ䘉ṧⲴㄟ⛩৲ᮠሩҾ᧿䘠↔ԫ࣑Ⲵᆇඇ᫽㓥〻ᒿᨀ‫׋‬Ҷа⿽㠚❦Ⲵ㺘
⽪Ǆ⴨৽ˈ䇮ᜣྲ᷌ᢺ⇿њᆇඇⲴ xǃy ඀ḷᇊѹѪㄟ⛩৲ᮠˈ䛓Ѹ㾱ᇎ⧠䘉њԫ࣑⴨ሩ㾱
ഠ䳮ᗇཊǄ
䲔Ҷ䘉Ӌㄟ⛩৲ᮠˈ䘉њᓄ⭘ѝⲴ〻ᒿ䈝䀰䘈वᤜл䶒Ⲵ৏ᆀ࠭ᮠ˖
x ˄MS x˅˄move to stack˅˖〫ࣘࡠḸǄྲ᷌ᆀඇ x ൘Ṽ䶒кˈ䘉њ᫽֌ᢺ
x 〫ࣘࡠḸ亦ᒦф䘄എ TǄ੖ࡉˈᆳӰѸҏн‫ڊ‬ᒦф䘄എ FǄ
x ˄MT x˅˄move to table˅˖〫ࣘࡠṼ䶒Ǆྲ᷌ᆇඇ x ᱟ൘ḸѝḀњս㖞ˈ
䘉њ᫽֌ᢺḸ亦Ⲵᆇඇ〫ࣘࡠṼ䶒ᒦф䘄എ TǄ੖ࡉ䘄എ FǄ
x ˄EQ x y˅˄equal˅˖⴨ㅹˈྲ᷌ x ㅹҾ y 䘄എ Tˈ੖ࡉ䘄എ FǄ
x ˄NOT x˅˖ྲ᷌ x=F 䘄എ Tˈྲ᷌ x=T 䘄എ FǄ
x ˄DU x y˅˄do until˅˖৽༽ᢗ㹼㺘䗮ᔿ x ⴤࡠ㺘䗮ᔿ y 䘄എ TǄ
ѪҶ䇴ՠԫ᜿㔉ᇊ〻ᒿⲴ䘲ᓄᓖˈKoza ᨀ‫׋‬Ҷ 166 њ䇝㓳䰞仈ˈ㺘⽪Ҷᖸཊ⿽н਼Ⲵ

ࡍ࿻ᆇඇ㔃ᶴˈ䰞仈Ⲵ䳮ᓖ਴ᔲǄԫ᜿㔉ᇊ〻ᒿⲴ䘲ᓄᓖቡᱟᆳ䀓ߣҶⲴ䇝㓳䰞仈Ⲵᮠ䟿Ǆ
㗔փ㻛ࡍ࿻ॆѪ 300 њ䲿ᵪ〻ᒿⲴ䳶ਸǄ㓿䗷Ҷ 10 ԓਾˈ㌫㔏ਁ⧠Ҷл䶒Ⲵ〻ᒿ䀓ߣҶᡰ
ᴹ 166 њ䰞仈Ǆ
˄EQ˄DU˄MT CS˅˄NOT CS˅˅˄DU˄MS NN˅˄NOT NN˅˅˅
⌘᜿䘉њ〻ᒿवਜ਼Ҷєњ DU˄ҏቡᱟĀDo Untilā˅䈝ਕⲴᒿࡇǄㅜањ DU 䈝ਕᗚ⧟

ൠᢺᖃࡽⲴḸ亦ᆇඇ〫ࣘࡠṼ䶒ⴤࡠᢺḸ〫オǄ❦ਾˈㅜҼњĀDo Untilā䈝ਕᗚ⧟ൠᢺл
ањᡰ䴰ᆇඇӾṼ䶒〫ࣘࡠḸ亦Ǆ䘉䟼ᴰཆቲⲴ EQ 㺘䗮ᔿ䎧ࡠⲴ֌⭘ᱟᨀ‫׋‬ањਸ⌅Ⲵਕ
⌅ᶕᧂࡇ䘉єњĀDo Untilāᗚ⧟Ǆ
ཊቁᴹӋԔӪ᛺ཷˈӵӵ㓿䗷Ҷࠐԓˈ䘉њ GP ቡਁ⧠Ҷ㜭䀓ߣᡰᴹ 166 њ䇝㓳䰞仈Ⲵ

〻ᒿǄ㌫㔏Ⲵ䘉њ㜭࣋ᖸབྷ〻ᓖк‫׍‬䎆Ҿᨀ‫Ⲵ׋‬สᵜ৲ᮠ઼৏ᆀ࠭ᮠˈԕ৺⭘ᶕ䇴ՠ䘲ᓄᓖ
Ⲵ䇝㓳ṧֻ䳶ਸǄ
9.5.3 䚍Ֆ㕌ぁ䈪᱄
↓ྲк䶒Ⲵֻᆀᡰ╄⽪Ⲵˈ䚇Ր㕆〻ᢺ䚇Ր㇇⌅ᢙኅࡠሩᆼᮤⲴ䇑㇇ᵪ〻ᒿⲴ䘋ॆǄቭ
㇑ᆳᗵ享㾱ᩌ㍒ᐘབྷⲴ‫ٷ‬䇮オ䰤ˈնᐢ㓿䇱ᇎ൘⴨ᖃᮠ䟿Ⲵᓄ⭘ѝ䚇Ր㕆〻ӗ⭏ҶԔӪ⵰䘧
Ⲵ㔃᷌ǄO’Reilly and Oppacher˄1994˅∄䖳Ҷ GP ㇇⌅઼ަԆᩌ㍒䇑㇇ᵪ〻ᒿオ䰤Ⲵ㇇⌅ˈ
ֻྲ⡜ኡ⌅˄hill climbing˅઼⁑ᤏ䘰⚛⌅˄simulated annealing˅Ǆ
ᖃ❦к䶒Ⲵ GP ㇇⌅ֻᆀᱟ⴨ᖃㆰঅⲴˈKoza et al.˄1996˅ᾲᤜҶ GP ㇇⌅൘аӋᴤ༽

ᵲⲴԫ࣑ѝⲴᓄ⭘ˈֻྲ䇮䇑⭥ᆀ└⌒⭥䐟઼࠶㊫㳻ⲭ䍘࠶ᆀ⡷⇥Ǆ└⌒⭥䐟䇮䇑䰞仈ᨀ‫׋‬
Ҷањ⴨ᖃ༽ᵲⲴ䰞仈Ǆ䘉䟼ˈ〻ᒿⲴ䘋ॆᱟӾㆰঅⲴപᇊ⿽ᆀ⭥䐟䖜ਈѪᴰ㓸Ⲵ⭥䐟䇮䇑Ǆ
GP ㇇⌅ѝ㓴ᔪ〻ᒿⲴ৏ᆀ࠭ᮠ䙊䗷ᨂ‫ޕ‬ᡆࡐ䲔⭥䐟䴦Ԧ઼ሬ㓯䘎᧕ᶕ㕆䗁䘉њ⿽ᆀ⭥䐟Ǆ
⇿њ〻ᒿⲴ䘲ᓄᓖᱟ䘉ṧ䇑㇇Ⲵ˖‫⁑ݸ‬ᤏ䘉њ⭥䐟Ⲵ䗃ࠪ˄֯⭘ SPICE ⭥䐟ԯⵏಘ˅ˈ❦ਾ
ⴻ䘉њ⭥䐟оᵏᵋⲴ䇮䇑Ⲵ䘲ᓄᓖⲴᐞ䐍Ǆ㋮⺞ൠ䇢ˈ䘲ᓄᓖ࠶٬ᱟሩҾ 101 њн਼Ⲵ䗃‫ޕ‬
仁⦷ˈ䇑㇇ᇎ䱵⭥䐟䗃઼ࠪᵏᵋ⭥䐟䗃ࠪ䰤䈟ᐞ䟿Ⲵ઼Ǆ൘䘉њֻᆀѝˈ㔤ᣔⲴ㗔փབྷሿᱟ
640,000ˈ䘹ᤙӗ⭏ 10%Ⲵਾԓ㗔փˈӔ৹ӗ⭏ 89%ˈਈᔲӗ⭏ 1%Ǆ㌫㔏ᱟ൘аਠ 64 㢲⛩
Ⲵᒦ㹼༴⨶ᵪкᢗ㹼ⲴǄ൘ᴰࡍⲴ䲿ᵪӗ⭏Ⲵ㗔փѝˈ⭥䐟ᱟྲ↔Ⲵнਸ⨶ԕ㠣Ҿ 98%Ⲵ
⭥䐟㹼Ѫᰐ⌅㻛 SPICE ԯⵏಘԯⵏǄ൘ㅜаԓѻਾᰐ⌅ԯⵏⲴ⭥䐟ⲴⲮ࠶∄л䱽ࡠ 84.9%ˈ
ㅜҼԓਾл䱽ࡠ 75.0%ˈ޽л䱽ࡠਾᶕ਴ԓѝᒣ൷ 9.6%Ǆ൘ࡍ࿻㗔փѝᴰྭ⭥䐟Ⲵ䘲ᓄᓖ
࠶٬ᱟ 159ˈо↔⴨∄ˈ20 ԓਾ࠶٬ᱟ 39ˈ137 ԓਾ࠶٬ᱟ 0.8Ǆ137 ԓਾⲴᴰ֣⭥䐟䗮ࡠ
Ⲵᙗ㜭о㾱≲Ⲵ䶎ᑨ⴨䘁Ǆ
൘བྷཊᮠᛵߥлˈ㺘⽪ᯩ⌅Ⲵ䘹ᤙ઼䘲ᓄᓖ࠭ᮠⲴ䘹ᤙሩ䚇Ր㕆〻Ⲵᙗ㜭ᱟ㠣‫ޣ‬䟽㾱
ⲴǄ⭡Ҿ䘉њ৏ഐˈⴞࡽ⹄ウⲴањ⍫䏳亶ฏᱟ㠚ࣘਁ⧠઼ਸᒦᆀ〻ᒿˈ᭩ழᴰࡍⲴ৏ᆀ࠭
ᮠ䳶ਸˈӾ㘼‫ݱ‬䇨㌫㔏ࣘᘱൠ᭩ਈ⭘ԕᶴᔪњփⲴ৏ᆀǄֻྲਟԕ৲㿱 Koza˄1994˅Ǆ
9.6 䘑ौૂᆜҖ⁗ශ
൘ᖸཊ㠚❦㌫㔏ѝˈ⭏⢙њփ൘ᆳԜа⭏ᖃѝ䜭൘ᆖҐྲօᴤྭൠ䘲ᓄ⧟ຳǄ਼ᰦˈ⭏
⢙઼⽮Պ䗷〻‫ݱ‬䇨ᆳԜⲴ⢙⿽൘ањवਜ਼ᖸཊԓⲴᰦᵏ޵䘲ᓄ⧟ຳǄ‫ޣ‬Ҿ䘋ॆ㌫㔏Ⲵањᴹ
䏓䰞仈ᱟ˖Āঅањփ⭏ભᵏ䰤ⲴᆖҐˈоᮤњ⢙⿽䖳䮯ᰦᵏ޵⭡䘋ॆ‫׳‬ᡀⲴᆖҐˈᆳԜⲴ
‫ޣ‬㌫ᱟӰѸ˛ā
9.6.1 ᣿傢‫ށ‬䘑ौ
、ᆖᇦ᣹傜‫˄ݻ‬Lamarck˅൘ॱҍц㓚ᵛᨀࠪˈཊԓⲴ䘋ॆⴤ᧕ਇࡠҶњ࡛⭏⢙փ൘ᆳ
Ԝ⭏ભᵏ䰤Ⲵ㓿傼Ⲵᖡ૽Ǆ⺞࠷ൠ䇢ˈԆᨀࠪњ࡛⭏⢙փⲴ㓿傼ⴤ᧕ᖡ૽ަਾԓⲴ䚇Ր㔃ᶴ˖
ྲ᷌ањњփ൘⭏ભᵏ޵ᆖՊҶ䚯ᔰḀ⿽ᴹ∂伏⢙ˈᆳ‫ׯ‬㜭ᢺ䘉⿽⢩ᖱ䚇Ր㔉ᆳⲴਾԓǄ䘉
ᱟањᖸ੨ᕅӪⲴ⥌ᜣˈഐѪ∄䎧ᘭ⮕њփ㓿傼ⲴĀ⭏ᡀᒦ⍻䈅˄generate-and-test˅ā䗷〻
˄ྲGA઼GPѝ䛓ṧ˅ˈᆳਟ㜭㧧ᗇᴤ儈᭸Ⲵ䘋ॆ䗷〻Ǆቭ㇑䘉њ⨶䇪ᖸᴹ੨ᕅ࣋ˈնⴞࡽ
Ⲵ、ᆖ䇱ᦞо᣹傜‫⁑ݻ‬රᖫᓅߢケǄⴞࡽ㻛᧕ਇⲴ㿲⛩ᱟˈњփⲴ䚇Ր㔃ᶴһᇎкнਇᆳⲴ
ৼӢⲴ⭏ᆈ㓿傼Ⲵᖡ૽Ǆቭ㇑䘉ᱟ᰾ᱮⲴ⭏⢙ᆖкⲴһᇎˈն䘁ᶕⲴ䇑㇇ᵪ⹄ウᐢ㓿㺘᰾ˈ
᣹傜‫ݻ‬䗷〻ᴹᰦਟԕᨀ儈䇑㇇ᵪ䚇Ր㇇⌅Ⲵ᭸⦷˄৲㿱Grefenstette 1991˗Ackley & Littman
Ǆ
1994˗Hart & Belew 1995˅
9.6.2 勃ᗭ⑟᭾ᓊ
ቭ㇑᣹傜‫ݻ‬䘋ॆ⁑ර⋑ᴹ㻛⭏⢙䘋ॆ䗷〻ᡰ᧕ਇˈӪԜᐢ㓿ᨀࠪҶަԆⲴᵪࡦˈ䙊䗷䘉
ӋᵪࡦњփᆖҐਟԕ᭩ਈ䘋ॆ䘋〻Ǆަѝа⿽㻛〠Ѫ劽ᗧ⑙᭸ᓄ˄Baldwin effect˅ ˈᱟṩᦞ
俆‫ݸ‬ᨀࠪ䘉⿽ᙍᜣⲴ J. M. Baldwin˄1896˅Ⲵ֌㘵਽ᆇભ਽ⲴǄ劽ᗧ⑙᭸ᓄสҾԕл⧠䊑˖
x ྲ᷌ањ⢙⿽൘ањਈॆⲴ⧟ຳѝ䘋ॆˈ䛓Ѹ䘋ॆⲴ঻࣋Պ᭟ᤱᴹᆖҐ㜭࣋
ⲴњփǄֻྲˈྲ᷌൘䘋ॆ⧟ຳѝࠪ⧠ҶањᯠⲴᦅ伏㘵ˈ䛓Ѹ㜭ᆖՊ䚯ᔰ
ᦅ伏㘵ⲴњփˈՊ∄н㜭ᆖՊ↔㜭࣋Ⲵњփᴤᡀ࣏Ǆ൘᭸᷌кˈ䘉⿽ᆖҐⲴ
㜭࣋ਟԕ֯њփ൘ަ⭏ભᵏ䰤ᢗ㹼а⿽ሿⲴተ䜘ᩌ㍒ˈԕᴰབྷॆᆳⲴ䘲ᓄ
ᓖǄ⴨৽ˈнᆖҐⲴњփⲴ䘲ᓄᓖᆼ‫ޘ‬ਆߣҾᆳⲴ䚇Ր㔃ᶴˈՊ༴Ҿ⴨ሩⲴ
࣓࣯Ǆ
x 䛓Ӌ㜭ཏᆖҐᖸཊ⢩ᙗ˄trait˅ⲴњփˈՊ䖳ቁൠ‫׍‬䎆Ҿ䚇Րԓ⸱ᶕ⺜ᙗൠ
㿴ᇊަ⢩ᙗǄ㔃᷌ˈ䘉Ӌњփਟԕ‫׍‬䎆њփᆖҐ‫ݻ‬ᴽ䚇Րԓ⸱ѝⲴĀђཡⲴā
ᡆĀᒦ䶎ᴰՈⲴā⢩ᙗˈӾ㘼᭟ᤱᴤ࣐ཊṧॆⲴสഐ⊐˄gene pool˅Ǆ᧕л
ᶕˈ䘉њᴤ࣐ཊṧॆⲴสഐ⊐ਟԕ‫׳‬䘋䘲ᓄᙗᴤᘛ䙏ൠ䘋ॆǄഐ↔ˈњփⲴ
ᆖҐ㜭࣋ާᴹ䰤᧕࣐䙏ᮤњ㗔փ䘋ॆ䘲ᓄⲴ֌⭘Ǆ
ֻྲˈ䇮ᜣḀњ⢙⿽Ⲵ⧟ຳѝਁ⭏ҶḀњᯠⲴਈॆˈ∄ྲࠪ⧠ањᯠⲴᦅ伏㘵Ǆ䘉ṧⲴ
ਈॆՊᴹ࡙Ҿ㜭ᆖՊ䚯ᔰᦅ伏㘵ⲴњփǄ䲿⵰㗔փѝ㠚ᡁᨀ儈ⲴњփⲴ∄ֻⲴ໎䮯ˈ㗔փՊ
᭟ᤱᴤ࣐ཊṧॆⲴสഐ⊐ˈ‫ݱ‬䇨䘋ॆ䗷〻˄ণ֯ᱟ䶎᣹傜‫Ⲵݻ‬Ā⭏ᡀᒦ⍻䈅ā䗷〻˅䘲ᓄᗇ
ᴤᘛǄ᧕лᶕˈ䘉⿽࣐䙏Ⲵ䘲ᓄਟԕ֯ḷ߶Ⲵ䘋ॆ䗷〻ᴤᘛൠ䘋ॆࠪа⿽䚇Ր⢩ᖱ˄䶎ᆖࡠ
Ⲵ⢩ᖱ˅ᶕ䚯ᔰᦅ伏㘵˄ֻྲˈа⿽ሩᦅ伏㘵Ⲵᵜ㜭᜗ᙅ˅Ǆഐ↔劽ᗧ⑙᭸ᓄᨀ‫׋‬Ҷа⿽䰤
᧕Ⲵᵪࡦˈ֯њփⲴᆖҐਟԕ↓䶒ᖡ૽䘋ॆ䙏ᓖǄ䙊䗷ᨀ儈⢙⿽Ⲵ⭏ᆈ઼࣋䚇Րཊṧᙗˈњ
փᆖҐՊ࣐ᘛ䘋ॆ䘋〻ˈӾ㘼໎࣐䘉њ⢙⿽䘋ॆࠪᴤྭൠ䘲ᓄᯠ⧟ຳⲴ䚇Ր⢩ᙗⲴᵪՊǄ
ӪԜаⴤࣚ࣋ᔰਁ⹄ウ劽ᗧ⑙᭸ᓄⲴ䇑㇇⁑රǄֻྲˈHinton & Nowlan˄1987˅ሩа

њㆰঅ⾎㓿㖁㔌Ⲵ㗔փ䘋㹼Ҷ䈅傼ˈ൘ањ㖁㔌њփⲴĀ⭏ભᵏā 䰤ˈᆳⲴаӋᵳ٬ᱟപ
ᇊⲴˈ㘼ަԆⲴᵳᱟਟԕ㻛䇝㓳ⲴǄ䘉њњփⲴ䚇Ր㔃ᶴߣᇊҶଚӋᵳ٬ᱟਟԕ㻛䇝㓳Ⲵˈ
䛓ӋᱟപᇊⲴǄ൘ᇎ傼ѝˈᖃн‫ݱ‬䇨њփᆖҐᰦˈ㗔փн㜭䲿⵰ᇎ䐥ᨀ儈ᆳⲴ䘲ᓄᓖǄ❦㘼ˈ
ᖃ‫ݱ‬䇨њփᆖҐᰦˈ㗔փ䗵䙏ൠᨀ儈ᆳⲴ䘲ᓄᓖǄ൘㗔փ䘋ॆࡍᵏⲴ਴ԓѝˈާᴹᖸཊਟ䇝
㓳ᵳ٬Ⲵњփঐᦞ䖳བྷⲴ∄ֻǄն䲿⵰䘋ॆⲴ䘋㹼ˈ㗔փੁ⵰䚇Ր㔉ᇊᵳ٬઼䖳ቁ‫׍‬䎆њփ
ᆖҐᵳ٬Ⲵᯩੁ䘋ॆˈ↓⺞Ⲵപᇊᵳ٬Ⲵᮠ䟿䎻Ҿ໎䮯ǄBelew˄1990˅ ˈHarvey˄1993˅઼
French & Messinger˄1994˅ᣕ੺Ҷሩ劽ᗧ⑙᭸ᓄⲴަԆ䇑㇇ᙗ⹄ウǄMitchell˄1996˅ѝᴹ
ањ‫ޣ‬Ҿ䘉њѫ仈Ⲵ㋮ᖙ㔬䘠Ǆǉ䚇Ր䇑㇇Ǌ˄Evolutionary Computation˅ᵲᘇⲴаᵏ⢩࠺
˄Turney et al. 1997˅वਜ਼Ҷࠐㇷᴹ‫ޣ‬劽ᗧ⑙᭸ᓄⲴ᮷ㄐǄ
9.7 ᒬ㺂䚍Ֆ㇍⌋
GA ᖸ㠚❦ൠ䘲ਸᒦ㹼ᇎ⧠ˈ㘼фᐢ㓿᧒㍒ࠪҶᖸཊᒦ㹼ॆⲴᯩ⌅Ǆ㋇㋂ᓖ˄coarse grain˅
ᒦ㹼ᯩ⌅ᢺ㗔փ㓶࠶ᡀ⴨ሩ⤜・Ⲵњփ㗔ˈ〠Ѫ㊫኎˄deme˅Ǆ❦ਾѪ⇿њ㊫኎࠶䝽㔉ањ
н਼Ⲵ䇑㇇㢲⛩ˈ൘⇿њ㢲⛩䘋㹼ḷ߶Ⲵ GA ᩌ㍒Ǆ㊫኎ѻ䰤Ⲵ䙊઼ؑӔ৹ਁ⭏Ⲵ仁⦷о㊫
኎޵⴨∄䖳վǄ㊫኎ѻ䰤ⲴӔᦒ䙊䗷䗱〫˄migration˅ᶕ䘋㹼ˈҏቡᱟḀӋњփӾањ㊫኎
༽ࡦᡆӔᦒࡠަԆⲴ㊫኎Ǆ䘉њ䗷〻⁑ᤏҶԕлⲴ⭏⢙䘋ॆᯩᔿˈণ㠚❦⭼ѝᔲփਇ㋮ਟ㜭
ਁ⭏൘࠶⿫Ⲵ⢙⿽ᆀ㗔փѻ䰤Ǆ䘉⿽ᯩ⌅Ⲵањྭ༴ᱟᆳ߿ቁҶ䶎ᒦ㹼 GA 㓿ᑨ⻠ࡠⲴᤕᥔ
䰞仈ˈ൘䶎ᒦ㹼㇇⌅ѝˈ⭡Ҿ䗷ᰙࠪ⧠᭟䝽ᮤњ㗔փⲴสഐරˈ֯㌫㔏䲧‫ޕ‬ተ䜘ᴰՈǄTanese
˄1989˅઼ Cohoon et al.˄1987˅᧿䘠Ҷ㋇㋂ᓖᒦ㹼 GA ㇇⌅ⲴֻᆀǄ
⴨ሩҾ㋇㋂ᓖᒦ㹼ᇎ⧠ˈ㓶㋂ᓖ˄fine-grained˅ᇎ⧠а㡜㔉㗔փѝⲴ⇿њњփ࠶䝽ањ
༴⨶ಘǄ❦ਾ⴨䛫Ⲵњփ䰤ਁ⭏䟽㓴ǄӪԜᐢ㓿ᨀࠪҶࠐњ⴨䛫⁑රˈӾᒣ䶒㖁Ṭࡠ䎵⧟㔃
ᶴǄSpiessens & Manderick˄1991˅᧿䘠Ҷ䘉ṧⲴ㌫㔏ⲴᇎֻǄStender˄1993˅ѝਟԕᗇࡠ
‫ޣ‬Ҿᒦ㹼 GA ㇇⌅Ⲵ䇪᮷䳶Ǆ
x 䚇Ր㇇⌅˄GA˅䘋㹼а⿽䲿ᵪॆⲴᒦ㹼⡜ኡᩌ㍒ˈᶕਁ⧠֯亴‫ݸ‬ᇊѹⲴ䘲
ᓄᓖ࠭ᮠᴰՈⲴ‫ٷ‬䇮Ǆ
x GA ᡰ䟷ਆⲴᩌ㍒ᱟสҾሩ⭏⢙䘋ॆⲴ⁑ᤏǄGA 㔤ᣔањ⭡ㄎҹ‫ٷ‬䇮㓴ᡀ
Ⲵཊṧॆ㗔փǄ൘⇿а⅑䘝ԓѝˈ䘹ࠪ㗔փѝ䘲ᓄᓖᴰ儈Ⲵᡀઈᶕӗ⭏ਾԓˈ
ᴯԓ㗔փѝ䘲ᓄᓖᴰᐞⲴᡀઈǄ‫ٷ‬䇮ᑨ㻛㕆⸱ᡀսѢˈਟԕ䙊䗷Ӕ৹㇇ᆀ㓴
ਸˈսѢкҏਟ㜭ਁ⭏䲿ᵪⲴਈᔲǄ
x GA 䱀᰾ҶྲօᢺᆖҐ䗷〻ⴻᡀᴰՈॆ䗷〻Ⲵањ⢩ֻǄާփᶕ䈤ˈᆖҐԫ
࣑ቡᱟṩᦞ亴‫ݸ‬ᇊѹⲴ䘲ᓄᓖ࠭ᮠਁ⧠ᴰՈⲴ‫ٷ‬䇮Ǆ䘉㺘᰾ަԆⲴᴰՈॆᢰ
ᵟˈֻྲ⁑ᤏ䘰⚛⌅ˈҏਟԕᓄ⭘ࡠᵪಘᆖҐ䰞仈Ǆ
x GA ᐢ㓿㻛Პ䙽ᓄ⭘ࡠᵪಘᆖҐཆⲴᴰՈॆ䰞仈ѝˈֻྲ䇮䇑Ոॆ䰞仈Ǆᖃ
ᢺ GA ᓄ⭘ࡠᆖҐԫ࣑ᰦˈᆳ⢩࡛䘲ਸ‫ٷ‬䇮ᖸ༽ᵲⲴԫ࣑˄ֻྲˈ‫ٷ‬䇮ᱟᵪ
ಘӪ᧗ࡦⲴ㿴ࡉ䳶ᡆ䇑㇇ᵪ〻ᒿ˅઼ᴰՈॆⲴⴞḷᱟ‫ٷ‬䇮Ⲵ䰤᧕࠭ᮠⲴԫ࣑
˄ֻྲˈ㾱≲ᗇࡠⲴ㿴ࡉ䳶ਟԕᡀ࣏ൠ᧗ࡦањᵪಘӪ˅Ǆ
x 䚇Ր㕆〻ᱟ䚇Ր㇇⌅Ⲵањਈփˈ൘䚇Ր㕆〻ѝˈ㻛᫽֌Ⲵ‫ٷ‬䇮ᱟ䇑㇇ᵪ〻
ᒿ㘼нᱟսѢǄӔ৹઼ਈᔲ᫽֌㻛᧘ᒯԕᓄ⭘Ҿ〻ᒿ㘼нᱟսѢǄӪԜᐢ㓿
╄⽪Ҷ䚇Ր㕆〻ᆖҐ䪸ሩḀӋԫ࣑Ⲵ〻ᒿˈ∄ྲ⁑ᤏᵪಘӪ᧗ࡦ˄Koza
1992˅઼䇶࡛㿶㿹൪Ჟ˄visual scenes˅ѝⲴ⢙փ˄Teller and Veloso 1994˅Ǆ
൘䇑㇇ᵪ、ᆖⲴᰙᵏˈӪԜቡᔰ࿻᧒㍒สҾ䘋ॆⲴ䇑㇇ᯩ⌅˄ֻྲ Box 1957 ઼ Bledsoe

1961˅Ǆ60 ᒤԓᨀࠪҶࠐњн਼Ⲵ䘋ॆᯩ⌅ˈਾᶕ৸㻛䘋а↕⹄ウǄRechenberg˄1965ˈ1973˅
ᔰਁⲴ䘋ॆㆆ⮕⭘ᶕՈॆᐕ〻䇮䇑ѝⲴᮠᆇ৲ᮠˈSchwefel˄1975ˈ1977ˈ1995˅઼ަԆа
ӋӪ㔗㔝⹄ウҶ䘉⿽ㆆ⮕ǄFolgel & Owens & Walsh˄1966˅ᔰਁҶ䘋ॆ㕆〻ˈ֌Ѫ䘋ॆᴹ
䲀⣦ᘱᵪⲴа⿽ᯩ⌅ˈབྷ䟿Ⲵ⹄ウ㘵㔗㔝᧒㍒Ҷ䘉⿽ᯩ⌅˄ֻྲ Fogel & Atmar 1993˅Ǆ
Holland˄1962ˈ1975˅ᨀࠪⲴ䚇Ր㇇⌅वਜ਼Ҷ㔤ᣔњփⲴ㓴ᡀབྷ㗔փⲴᾲᘥˈᒦфᕪ䈳൘
䘉ṧⲴ㌫㔏ѝӔ৹ᱟањ‫ޣ‬䭞Ⲵ᫽֌ǄKoza˄1992˅ӻ㓽Ҷ䚇Ր㕆〻ˈᢺ䚇Ր㇇⌅Ⲵᩌ㍒
ㆆ⮕ᓄ⭘ࡠ⭡䇑㇇ᵪ〻ᒿ㓴ᡀⲴ‫ٷ‬䇮ѝǄ䲿⵰䇑㇇ᵪ⺜Ԧнᯝൠਈᗇᴤᘛ઼ᴤ‫ׯ‬ᇌˈሩ䘋ॆ
ᯩ⌅Ⲵ‫ޤ‬䏓ҏнᯝ໎䮯Ǆ
֯⭘ GA ᆖҐ㿴ࡉ䳶Ⲵа⿽ᯩ⌅ᱟ⭡ K. DeJong ઼ԆⲴᆖ⭏൘ Pittsburgh བྷᆖᔰਁⲴ˄৲

㿱 Smith 1980˅Ǆ൘䘉⿽ᯩ⌅ѝˈ⇿њ㿴ࡉ䳶ᱟㄎҹ‫ٷ‬䇮㓴ᡀⲴ㗔փⲴањᡀઈˈቡ‫ۿ‬ᵜㄐ
䇘䇪Ⲵ GABIL ㌫㔏ѝⲴаṧǄHolland ઼ԆⲴᆖ⭏˄Holland 1986˅൘ Michigan བྷᆖᔰਁҶ
а⿽н਼Ⲵᯩ⌅ˈަѝ⇿њ㿴ࡉᱟ㗔փⲴањᡀઈˈ㘼㗔փᵜ䓛ᱟањ㿴ࡉ䳶ǄWright˄1977˅
Ӿ⭏⢙ᆖ䀂ᓖ࠶᷀Ҷਈᔲǃ㑱⇆ǃӔ৹㑱⇆઼䘋ॆ䘹ᤙⲴ֌⭘Ǆ
Mitchell˄1996˅઼ Goldberg˄1989˅ᱟ䇘䇪䚇Ր㇇⌅䘉аѫ仈ⲴєᵜᮉᶀǄForrest˄1993˅
ᾲᤜҶ GA ѝⲴᢰᵟ䰞仈ˈGoldberg˄1994˅ᾲᤜҶᴰ䘁Ⲵࠐњᓄ⭘ǄKoza˄1992˅‫ޣ‬Ҿ䚇
Ր㕆〻Ⲵу㪇ᱟሩ䚇Ր㇇⌅ᢙኅࡠ᫽֌䇑㇇ᵪ〻ᒿⲴḷ߶৲㘳Ǆਁ㺘ᯠᡀ᷌Ⲵѫ㾱Պ䇞ᱟ䚇
Ր㇇⌅ഭ䱵Պ䇞˄ICGA˅ǄަԆ⴨‫Ⲵޣ‬Պ䇞वᤜ㠚䘲ᓄ㹼ѪԯⵏՊ䇞˄CSAB˅ˈӪᐕ⾎㓿
㖁㔌઼䚇Ր㇇⌅ഭ䱵Պ䇞˄ICANNGA˅ˈԕ৺ IEEE 䘋ॆ䇑㇇ഭ䱵Պ䇞˄ICEC˅Ǆⴞࡽҏᴹ
䚇Ր㕆〻ᯩ䶒ⲴᒤՊ˄Koza et al. 1996b˅Ǆǉ䘋ॆ䇑㇇ᵲᘇǊ˄Evolutionary Computation
Journal˅ᱟ䘉њ亶ฏᴰᯠ⹄ウᡀ᷌ⲴањᶕⓀǄ ǉᵪಘᆖҐǊ
˄Machine Learning˅ᵲᘇⲴа
Ӌ⢩࠺ҏᱟ䪸ሩ GA ⲴǄ
Ґ仈
9.1 Ѫㅜ 3 ㄐѝ᧿䘠Ⲵ PlayTennis 䰞仈䇮䇑ањ䚇Ր㇇⌅ˈᆖҐਸਆⲴ࠶㊫㿴ࡉǄ㋮⺞

ൠ᧿䘠ࠪަѝሩ‫ٷ‬䇮ⲴսѢ㕆⸱઼а㓴Ӕ৹㇇ᆀǄ
9.2 ᇎ⧠㓳Ґ 9.1 ѝⲴㆰঅ GAǄ⭘н਼Ⲵ㗔փབྷሿ pǃ⇿аԓѝ㻛␈⊠Ⲵ∄ֻ r ઼ਈᔲ

⦷ m 䘋㹼䈅傼Ǆ
9.3 ᢺ GP ਁ⧠Ⲵ〻ᒿ˄൘ㅜ 9.5.2 㢲ѝ᧿䘠˅䟽ᯠ㺘⽪ѪṁǄ❦ਾሶṁⲴєњᤧ䍍֌Ѫ

єњৼӢˈ൘ަкᓄ⭘ GP ⲴӔ৹㇇ᆀǄ䈤᰾ަѝӔ৹㇇ᆀⲴ᫽֌䗷〻Ǆ
9.4 㘳㲁ᢺ GA ᓄ⭘ࡠሩӪᐕ⾎㓿㖁㔌˄⢩࡛ᱟо৽ੁՐ᫝㇇⌅䇝㓳Ⲵ㖁㔌а㠤Ⲵࡽ侸

㖁㔌ˈ㿱ㅜ 4 ㄐ˅ራ᢮а㓴ਸ䘲Ⲵᵳ٬Ǆ㘳㲁ањ 3u2u1 Ⲵ࠶ቲࡽ侸㖁㔌Ǆ᧿䘠а⿽ᢺ㖁
㔌ᵳ٬㕆⸱ᡀսѢⲴᯩ⌅ˈᒦ᧿䘠а྇䘲ᖃⲴӔ৹㇇ᆀǄᨀ⽪˖н㾱൘սѢк‫ݱ‬䇨ᡰᴹਟ㜭
ⲴӔ৹᫽֌Ǆᤷࠪ൘䇝㓳㖁㔌ᵳ٬ᯩ䶒ˈ֯⭘ GA о৽ੁՐ᫝㇇⌅⴨∄ⲴањՈ⛩઼ањ㕪
⛩Ǆ
৲㘳᮷⥞
ㅢ10ㄖ ᆜҖ㿺ࡏ䳼ਾ
ᇍᄺдࠄⱘ‫؛‬䆒ˈ᳔‫݋‬᳝㸼ᕕ࡯ⱘ੠᳔㛑ЎҎ㉏᠔⧚㾷ⱘ㸼⼎ᮍ⊩ПϔЎ if-then 㾘߭
ⱘ䲚ড়Ǆᴀゴ᥶㋶њ㢹ᑆ㛑ᄺд䖭ḋⱘ㾘߭䲚ড়ⱘㅫ⊩Ǆ݊Ё᳔䞡㽕ⱘϔ⾡ᰃᄺдࣙ৿ব䞣
ⱘ㾘߭䲚ড়ˈ៪⿄Ўϔ䰊 Horn ᄤহ䲚ড়Ǆ⬅Ѣϔ䰊 Horn ᄤহ䲚ড়ৃ㹿㾷䞞Ў䘏䕥㓪⿟䇁
㿔 Prolog Ёⱘ⿟ᑣˈᄺдⱘ䖛⿟㒣ᐌ㹿⿄Ўᔦ㒇䘏䕥㓪⿟˄ILP˅Ǆᴀゴ㗗ᆳњ໮⾡ᄺд㾘
߭䲚ড়ⱘ䗨ᕘˈ݊Ёϔ⾡䗨ᕘ෎Ѣᴎ఼ᅮ⧚䆕ᯢ఼Ёⓨ㒢᪡԰ⱘ䗚䕀Ǆ
10.1 ԁ㔃
೼䆌໮ᚙ‫މ‬ϟˈ᳝ᖙ㽕ᄺдϔϾⳂᷛߑ᭄ˈᅗ㸼⼎Ў݅ৠᅮН䆹ߑ᭄ⱘ㢹ᑆ if-then 㾘
߭ⱘ䲚ড়Ǆབ㄀ 3 ゴ᠔⼎ˈᄺд㾘߭䲚ড়ⱘϔ⾡ࡲ⊩ᰃ佪‫ܜ‬ᄺдࠄ‫އ‬ㄪᷥˈ✊ৢᇚℸᷥ䕀ᤶ
Ўϔㄝᬜⱘ㾘߭䲚ড়Ǆ঺ϔ⾡ᮍ⊩ᰃ೼㄀ 9 ゴҟ㒡ⱘ䘫Ӵㅫ⊩ˈᅗ⫼ԡІ㓪ⷕ↣Ͼ㾘߭䲚ড়ˈ
✊ৢ⫼䘫Ӵ᧰㋶ㅫᄤᴹ᥶㋶ᭈϾ‫؛‬䆒ぎ䯈Ǆᴀゴ៥Ӏ䅼䆎ϔᮣϡৠⱘㅫ⊩ˈᅗⳈ᥹ᄺд㾘߭
䲚ড়ˈ䖭ᮣㅫ⊩Ϣࠡ䴶ⱘㅫ⊩Ⳍ↨᳝ϸ⚍݇䬂ⱘϡৠǄ佪‫ˈܜ‬ᅗӀৃᄺдࣙ৿ব䞣ⱘϔ䰊㾘
߭䲚ড়ˈ䖭ϔ⚍ᕜ䞡㽕ˈ಴Ўϔ䰊ᄤহⱘ㸼⼎㛑࡯↨ੑ乬㾘߭㽕ᔎᕫ໮Ǆ㄀Ѡˈ䖭䞠䅼䆎ⱘ
ㅫ⊩Փ⫼ᑣ߫㽚Ⲫㅫ⊩ˈϔ⃵ᄺдϔϾ㾘߭ˈҹ䗦๲ഄᮍᓣᔶ៤᳔㒜ⱘ㾘߭䲚ড়Ǆ
԰Ўϔ䰊㾘߭䲚ড়ⱘ՟ᄤˈ㗗㰥ҹϟϸϾ㾘߭ˈᅗӀ݅ৠᦣ䗄њⳂᷛὖᗉ AncestorǄ䖭
䞠៥ӀՓ⫼䇧䆡 Parent(x,y)ᴹ㸼⼎ y ᰃ x ⱘ⠊҆៪↡҆ˈ㗠䇧䆡 Ancestor(x,y)㸼⼎ y ᰃ x ⱘ
ӏᛣҷⱘ⼪‫ܜ‬Ǆ
IF Parent(x,y) THEN Ancestor(x,y)
IF Parent(x,z)ġAncestor(z,y) THEN Ancestor(x,y)
⊼ᛣҹϞϸϾ㾘߭ᕜ㋻‫ޥ‬ഄᦣ䗄њϔϾ䗦ᔦߑ᭄ˈᅗᕜ䲒⫼‫އ‬ㄪᷥ៪݊Ҫⱘੑ乬ⱘᮍ⊩
ᴹ㸼⼎ǄЎ䇈ᯢϔ䰊㾘߭ⱘᔎ໻ⱘ㸼⼎㛑࡯ˈৃҹ㗗㰥䗮⫼ⱘ㓪⿟䇁㿔 PrologǄ೼ Prolog
Ёˈ⿟ᑣᰃϔ䰊㾘߭ⱘ䲚ড়ˈབϞ᠔⼎ⱘ䙷ḋ˄䖭⾡ᔶᓣⱘ㾘߭г㹿⿄Ў Horn ᄤহ˅Ǆᅲ
䰙Ϟˈབᵰ⿡⿡ׂᬍϞ䴶ϸϾ㾘߭ⱘ䇁⊩ˈህৃҹᕫࠄϔϾড়⊩ⱘ Prolog ⿟ᑣˈᅗ⫼ᴹ䅵
ㅫ Ancestor ݇㋏Ǆ಴ℸˈϔϾৃҹᄺд䖭⾡㾘߭䲚ড়ⱘ䗮⫼ㅫ⊩ˈৃ㹿ⳟ԰ᰃҢḋ՟Ё㞾
ࡼ᥼ᇐߎ Prolog ⿟ᑣⱘㅫ⊩Ǆᴀゴҟ㒡њ೼㒭ᅮ䗖ᔧⱘ䆁㒗՟䲚ড়ᯊˈ㛑໳ᄺд䖭⾡㾘߭
ⱘᄺдㅫ⊩Ǆ
ᅲ䏉Ё෎Ѣϔ䰊㸼⼎ⱘᄺд㋏㒳Ꮖ៤ࡳഄᑨ⫼Ѣ৘⾡䯂乬ˈབ೼䋼䈅ҾЁᄺдાϔϾ࣪
ᄺ㥃ક㛑㉬ড়⹢⠛BuchananLindsayˈᄺдાϔϾ࣪ᄺѮ㒧ᵘӮѻ⫳䇅ᇐ᳝
ᴎԧさবⱘᬒᇘᗻ⠽䋼˄ϔϾ݇Ѣ㟈Ⱐ⠽䋼ⱘሲᗻ˅˄Srinivasanetal˅ˈҹঞᄺд᳝
䰤ऩ‫ܗ‬㔥ҹߚᵤ⠽⧚㒧ᵘЁⱘᑨ࡯˄Dolsak Muggleton˅Ǆ೼↣Ͼᑨ⫼Ёˈ‫؛‬䆒ⱘ㸼
⼎ᖙ乏ࣙ৿݇㋏ᮁ㿔ˈᅗৃ⬅ϔ䰊㸼⼎ᴹㅔऩഄ㸼䖒ˈैᕜ䲒⫼ੑ乬㸼⼎ᴹᦣ䗄Ǆ
೼ᴀゴЁˈ៥Ӏ‫ܜ‬ҟ㒡㛑໳ᄺдੑ乬㾘߭䲚ⱘㅫ⊩ˈेϡ৿ব䞣ⱘ㾘߭Ǆ೼䖭⾡ḚᶊЁˈ
᧰ᇏ‫؛‬䆒ぎ䯈ҹᄺдࠄᵤপ㾘߭䲚ড়ⱘㅫ⊩↨䕗ᯧѢ⧚㾷Ǆ✊ৢˈ៥Ӏ㗗㰥њᇚ䖭ѯㅫ⊩ᠽ
ሩࠄϔ䰊㾘߭Ǆ᥹ϟᴹ䅼䆎њᔦ㒇䘏䕥ⱘϸ⾡䗮⫼䗨ᕘˈҹঞᔦ㒇੠ⓨ㒢᥼⧚ⱘ෎ᴀ݇㋏Ǆ
10.2 ᓅࡍ㾼ⴌ㇍⌋
䖭䞠㗗㰥ⱘϔᮣㅫ⊩ˈ݊ᄺд㾘߭䲚ⱘㄪ⬹Ў˖ᄺдϔϾ㾘߭ˈ⿏এᅗ㽚Ⲫⱘ᭄᥂ˈ‫ݡ‬
䞡໡䖭ϔ䖛⿟Ǆ䖭ḋⱘㅫ⊩㹿⿄Ўᑣ߫㽚Ⲫ˄sequentialcovering˅ㅫ⊩Ǆᛇ䈵៥ӀᏆ᳝њ
ϔϾᄤ⿟ᑣ learnoneruleˈᅗⱘ䕧ܹЎϔ㒘ℷ՟੠ড՟ˈ✊ৢ䕧ߎऩϾ㾘߭ˈᅗ㛑໳㽚Ⲫ
䆌໮ℷ՟ˈᑊϨ㽚Ⲫᕜᇥⱘড՟Ǆ៥Ӏ㽕∖䖭ϔ䕧ߎⱘ㾘᳝߭䕗催ⱘ㊒⹂ᑺˈԚϡᖙ᳝䕗催
ⱘ㽚ⲪᑺǄ䕗催ⱘ㊒⹂ᑺ䇈ᯢᅗ᠔‫ⱘߎخ‬乘⌟ᑨЎℷ⹂ⱘǄৃ᥹ফ䕗Ԣⱘ㽚Ⲫᑺˈ㸼⼎ᅗϡ
ᖙᇍ↣Ͼ䆁㒗ḋ՟䛑԰ߎ乘⌟Ǆ
᳝њ䖭ḋϔϾᄺдऩϾ㾘߭ⱘ learnonerule ᄤ⿟ᑣˈ㽕ᄺд㾘߭䲚ˈϔϾᯢᰒⱘᮍ⊩

ᰃ೼᠔᳝ৃ⫼䆁㒗ḋ՟Ϟᠻ㸠 learnoneruleˈ‫⿏ݡ‬এ⬅݊ᄺࠄⱘ㾘߭㽚Ⲫⱘℷ՟ˈ‫ݡ‬೼࠽
ԭⱘ䆁㒗ḋ՟Ϟᠻ㸠ᅗҹᄺд㄀ѠϾ㾘߭Ǆ䆹䖛⿟ৃ䞡໡㢹ᑆ⃵ˈⳈࠄ᳔ৢᄺдࠄᵤপ㾘߭
䲚ˈᅗӀ݅ৠ㽚Ⲫℷ՟ˈ㽚Ⲫ⿟ᑺ䖒ࠄ᠔Ꮰᳯⱘ↨՟Ǆㅫ⊩㹿⿄Ўᑣ߫㽚Ⲫㅫ⊩ᰃ಴Ўᅗᣝ
⃵ᑣᄺдࠄϔ㒘㾘߭ˈᅗӀ݅ৠ㽚Ⲫњܼ䚼ℷ՟Ǆ᳔㒜ⱘ㾘߭䲚ৃ㹿ᥦᑣˈ䖭ḋߚ㉏ᮄᅲ՟
ᯊৃ‫ܜ‬ᑨ⫼㊒ᑺ᳔催ⱘ㾘߭Ǆᑣ߫㽚Ⲫㅫ⊩ⱘϔϾॳൟ೼㸼 Ё䰜䗄Ǆ
ᑣ߫㽚Ⲫㅫ⊩ᰃᑓ⊯Փ⫼ⱘᄺдᵤপ㾘߭䲚ㅫ⊩ⱘ݊ЁПϔǄᅗᇚᄺдᵤপ㾘߭䲚ⱘ䯂
乬࣪ㅔЎϔ㋏߫᳈ㅔऩⱘ䯂乬ˈ↣Ͼᄤ䯂乬া䳔ᄺࠄऩϾড়প㾘߭Ǆ⬅Ѣᅗᠻ㸠ⱘᰃϔ⾡䋾
ှ᧰㋶ˈᔶ៤ᑣ߫࣪ⱘ㾘߭Ϩ≵᳝ಲ⒃ˈ᠔ҹᅗϡ㛑ֱ䆕ᡒࠄ㛑㽚Ⲫḋ՟ⱘ᳔ᇣⱘ៪᳔Շⱘ
㾘߭Ǆ
བԩ䆒䅵 learnonerule ⿟ᑣҹ䖒ࠄᑣ߫㽚Ⲫㅫ⊩ⱘ㽕∖˛៥Ӏ䳔㽕ϔϾㅫ⊩㛑໳ᔶ៤

᳝䕗催㊒ᑺⱘ㾘߭ˈԚϡᖙ㽚Ⲫ᠔᳝ⱘℷ՟Ǆ೼ᴀ㡖Ёሩ⼎њ৘⾡ㅫ⊩ˈᑊᦣ䗄њᅗӀ೼ᄺ
ᴃⷨおϞᏆ᥶㋶ⱘЏ㽕Ꮒ߿Ǆᴀ㡖া㗗㰥ੑ乬㾘߭Ǆৢ䴶ⱘ㡖Ёᇚᡞ䖭ѯㅫ⊩ᠽሩࠄϔ䰊
HornᄤহǄ
㸼 10-1 ᄺдᵤপⱘ㾘߭䲚ⱘᑣ߫㽚Ⲫㅫ⊩Ǆ
learn-one-rule ᖙ乏䖨ಲऩϾⱘ㛑㽚Ⲫᶤѯ Examples ⱘ㾘߭Ǆperformance ᰃ⫼᠋ᦤկⱘᄤ⿟ᑣˈ

ҹ䆘Ԅ㾘߭ⱘ䋼䞣Ǆᔧㅫ⊩‫ݡ‬гϡ㛑ᄺдࠄϔϾᗻ㛑䍙䖛㒭ᅮ䯜ؐ Threshold ⱘ㾘߭ᯊˈ䆹ㅫ⊩㒜
ℶǄ
Sequential-covering(Target_attribute, Attributes, Examples, Threshold)

Learned_rulesĕ{}
Ruleĕlearn-one-rule(Target_attribute, Attributes, Examples)
ᔧ performance(Rule, Examples) > Thresholdˈ‫˖خ‬
Learned_rulesĕLearned_rules + Rule
ExamplesĕExamples-{㹿 Rule ℷ⹂ߚ㉏ⱘḋ՟}
Ruleĕlearn-one-rule(Target_attribute, Attributes, Examples)
Learned_rulesĕᣝ✻೼ Examples Ϟⱘ performance ᥦᑣⱘ Learned_rules
䖨ಲ Learned_rules
10.2.1 ж㡢ࡦ⢯⇀ḧ⣬ᩒ㍘
ᅲ⦄ learnonerule ⱘϔϾ᳝ᬜ䗨ᕘᰃᇚ‫؛‬䆒ぎ䯈᧰㋶䖛⿟䆒䅵ЎϢ ID ㅫ⊩ЁⳌԐⱘ

ᮍᓣˈԚ೼↣ϔℹা⊓ⴔ᳔᳝Ꮰᳯⱘߚᬃ৥ϟǄབ೒ ᠔⼎ⱘ᧰㋶ᷥˈ᧰㋶ᓔྟѢ᳔ϔ
㠀ⱘ㾘߭ࠡӊ˄े㛑ऍ䜡᠔᳝ᅲ՟ⱘぎ⌟䆩˅ˈ✊ৢ䋾ှഄࡴܹ䙷ѯ೼䆁㒗ḋ՟Ϟᗻ㛑ᬍ䖯
᳔໻ⱘሲᗻ⌟䆩Ǆϔᮺ䆹⌟䆩㹿ࡴܹˈ䆹䖛⿟䞡໡ˈ䋾ှഄࡴܹ㄀ѠϾሲᗻ⌟䆩ˈձℸ㉏᥼Ǆ
བ ID 䙷ḋˈ䆹䖛⿟䗮䖛䋾ှഄ๲ࡴᮄⱘሲᗻ⌟䆩ᴹ㦋ᕫ‫؛‬䆒ˈⳈࠄ䆹‫؛‬䆒ⱘᗻ㛑ࠄ䖒ϔ
ৃ᥹ফⱘ⿟ᑺǄϢ ID ϡৠⱘᰃˈℸ learnonerule ⱘᅲ⦄೼↣ϔℹ⊓ⴔऩϾߚᬃˉˉेѻ
⫳᳔Շᗻ㛑ⱘሲᗻؐᇍˈ㗠ϡᰃ⫼๲䭓ᄤᷥⱘࡲ⊩㽚Ⲫ᠔䗝ሲᗻⱘ᠔᳝ৃ㛑ؐǄ
䖭⾡ᅲ⦄ learnonerule ⱘ䗨ᕘᠻ㸠ⱘᰃᇍৃ㛑ᄬ೼ⱘ㾘߭ⱘϔ㠀ࠄ⡍⅞᧰㋶ˈҹᕫࠄ

ϔϾ᳝䕗催㊒ᑺԚϡϔᅮᅠܼ㽚Ⲫ᭄᥂ⱘ㾘߭Ǆབ೼‫އ‬ㄪᷥᄺдЁ䙷ḋˈ᳝䆌໮ᮍ⊩ৃҹᅮ
Н䗝ᢽĀ᳔Շāߚᬃⱘᑺ䞣ᷛ‫ޚ‬ǄϢ೼ ID Ё㉏Ԑˈ៥ӀৃᅮН᳔ՇߚᬃЎᅗ㽚Ⲫⱘḋ՟
᳔᳝Ԣⱘ➉˄ಲᖚᓣ ˅Ǆ
೒ 10-1 Learn-one-rule Ңϔ㠀ࠄ⡍⅞䖛⿟Ёⱘ㾘߭ࠡӊ᧰㋶
೼↣ϔℹˈ᳔Շ㾘߭ⱘࠡӊ㹿ҹ৘⾡ৃ㛑ᮍᓣ⡍࣪Ǆ㾘߭ৢӊᰃ⬅⒵䎇ࠡӊⱘḋ՟᠔‫އ‬ᅮⱘǄ
䆹೒ᰒ⼎ⱘᰃᆑᑺЎ 1 ⱘ᷅⢊᧰㋶Ǆ
Ϟ䴶᥼㤤ⱘϔ㠀ࠄ⡍⅞᧰㋶ᰃϔ⾡ϡᏺಲ⒃ⱘ䋾ှ⏅ᑺӬ‫᧰ܜ‬㋶Ǆབ݊Ҫ䋾ှ᧰㋶ϔ
ḋˈᅗ᠔ᏺᴹⱘॅ䰽ᰃ↣ϔℹৃ㛑԰ߎњ⃵Ӭⱘ䗝ᢽǄЎ‫ޣ‬ᇣ䖭⾡亢䰽ˈৃᇚℸㅫ⊩ᠽሩЎ
ϔ⾡᷅⢊᧰㋶˄beamsearch˅ˈे↣ϔℹㅫ⊩ֱ⬭ k Ͼ᳔Շ‫׭‬䗝ⱘ߫㸼ˈ೼↣ϔ᧰㋶ℹᇍ䖭
k Ͼ᳔Շ‫׭‬䗝⫳៤ߚᬃ˄⡍࣪˅ˈᑊϨ㒧ᵰ䲚‫ݡ‬㹿ࠞ‫ޣ‬㟇 k Ͼ᳔ৃ㛑៤ਬǄ᷅⢊᧰㋶䎳䏾ᔧ
᳔ࠡ催ߚؐ‫؛‬䆒ⱘ᳔᳝Ꮰᳯⱘ᳓ҷ㗙ˈҹՓ↣ϔℹЁᅗӀⱘ᠔᳝ৢ㒻䛑㹿㗗㰥ࠄǄ䆹ϔ㠀ࠄ
⡍⅞᷅⢊᧰㋶⫼Ѣ CN ⿟ᑣˈᅗ⬅ Clark Niblett˄˅ᦤߎǄ䆹ㅫ⊩೼㸼 Ёᦣ䗄Ǆ
㸼 10-2 learn-one-rule ⱘϔ⾡ᅲ⦄ᰃϔ㠀ࠄ⡍⅞᷅⢊᧰㋶Ǆ
ᔧࠡ‫؛‬䆒ⱘ䖍㓬㸼⼎Ўব䞣 Candidate_hypothesesǄ䆹ㅫ⊩Ϣ Clark & Niblett˄1989˅ᦣ䗄ⱘ

CN2 ⿟ᑣⳌ㉏ԐǄ
Learn-one-rule(Target_attribute, Attributes, Examples, k)

䖨ಲϔϾ㽚Ⲫ㢹ᑆḋ՟ⱘ㾘߭Ǆᅲᮑϔ㠀ࠄ⡍⅞䋾ှ᷅⢊᧰㋶ҹᕫࠄ᳔Շ㾘߭ˈ⬅ performance ᑺ䞣ᴹᓩᇐǄ
߱ྟ࣪ Best_hypothesis Ў᳔ϔ㠀ⱘ‫؛‬䆒
߱ྟ࣪ Candidate_hypotheses Ў䲚ড়{Best_hypothesis}
ᔧ Candidate_hypotheses ϡぎˈ‫خ‬ҹϟ᪡԰˖
1.⫳៤㋻䚏᳈⡍⅞ⱘ‫׭‬䗝‫؛‬䆒
All_constraintsĕ᠔᳝ᔶᓣЎ(a=v)ⱘ㑺ᴳ䲚ড়ˈ݊Ё a Ў Attributes ⱘ៤ਬˈ㗠 v Ўߎ⦄೼
ᔧࠡ Examples 䲚ড়Ёⱘ a ؐ
New_candidate_hypotheseĕ
ᇍ Candidate_hypotheses Ё↣Ͼ hˈ
ᇍ All_constraints Ё↣Ͼ c
䗮䖛ࡴܹ㑺ᴳ c ߯ᓎϔϾ h ⱘ⡍࣪ᓣ
Ң New_candidate_hypothese Ё⿏এӏᛣ䞡໡ⱘǃϡϔ㟈ⱘ៪䴲ᵕ໻⡍⅞࣪ⱘ‫؛‬䆒
2.᳈ᮄ Best_hypothesis
ᇍ New_candidate_hypotheses Ё᠔᳝ h ‫خ‬ҹϟ᪡԰˖
བᵰ(performance(h, Examples, Target_attribute)
> performance(Best_hypothesis, Examples, Target_attribute))
߭ Best_hypothesisĕh
3.᳈ᮄ Candidate_hypotheses
Candidate_hypothesesĕNew_candidate_hypotheses Ё k Ͼ᳔Շ៤ਬˈᣝ✻ performance ᑺ䞣
䖨ಲϔϾབϟᔶᓣⱘ㾘߭˖
Āབᵰ Best_hypothesisˈ߭ predictionā
݊Ё prediction Ў೼Ϣ Best_hypothesis ऍ䜡ⱘ Examples Ё᳔乥㐕ⱘ Target_attribute ؐ
performance(h, Examples, Target_attribute)

h_examplesĕϢ h ऍ䜡ⱘ Examples ᄤ䲚
䖨ಲ-Entropy(h_examples)ˈ݊Ё Entropy ᰃ݇Ѣ Target_attribute ⱘ➉
ϟ䴶ᰃᇍ㸼 Ёⱘ learnonerule ㅫ⊩ⱘϔѯ䇈ᯢǄ佪‫⊼ˈܜ‬ᛣ೼ㅫ⊩Џᕾ⦃Ё㗗㰥

ⱘ↣Ͼ‫؛‬䆒ᰃሲᗻؐ㑺ᴳⱘড়পǄ↣Ͼড়প‫؛‬䆒ᇍᑨѢᕙᄺд㾘߭ⱘ‫׭‬䗝ࠡӊ䲚ড়ˈᅗ⬅
݊㽚Ⲫⱘḋ՟ⱘ➉ᴹ䆘ԄǄ᧰㋶䖛⿟ϡᮁ⡍࣪‫׭‬䗝‫؛‬䆒ˈⳈࠄࠄ䖒ϔϾᵕ໻⡍⅞‫؛‬䆒ˈᅗࣙ
৿᠔᳝ৃ⫼ⱘሲᗻǄ⬅䆹ㅫ⊩䕧ߎⱘ㾘߭Ў᧰㋶䖛⿟Ё䘛ࠄⱘᗻ㛑᳔Շ˄performance ᳔໻˅
ⱘ㾘߭ˉˉϡϔᅮᰃ᧰㋶᳔㒜ѻ⫳ⱘ‫؛‬䆒Ǆ㾘߭ⱘৢӊ䕧ߎা೼ㅫ⊩ⱘ᳔ৢϔℹѻ⫳ˈ೼݊
ࠡӊ˄㸼⼎Ўব䞣 BestBhypothesis˅⹂ᅮПৢˈㅫ⊩ᵘ䗴ߎⱘ㾘߭ৢӊ⫼Ѣ乘⌟೼㾘߭ࠡӊ
᠔㛑㽚Ⲫⱘḋ՟Ё᳔ᐌ㾕ⱘⳂᷛሲᗻؐǄ᳔ৢˈ䖬ᑨ⊼ᛣሑㅵՓ⫼њ᷅⢊᧰㋶ҹ‫ޣ‬ᇣ亢䰽ˈ
䋾ှ᧰㋶ҡৃ㛑ѻ⫳⃵Ӭⱘ㾘߭Ǆ✊㗠ˈेՓ䖭ḋˈᑣ߫㽚Ⲫㅫ⊩ҡ㛑ᄺࠄϔ㒘㾘߭ˈᅗӀ
݅ৠ㽚Ⲫ䆁㒗ḋ՟ˈ಴Ўᅗᇍ࠽ԭⱘ᳾㽚Ⲫḋ՟䞡໡䇗⫼њ learnoneruleǄ
10.2.2 ࠖ〃਎ශ
ᑣ߫㽚Ⲫㅫ⊩ҹঞ learnonerule ㅫ⊩ৃᄺд ifthen 㾘߭䲚ҹ㽚Ⲫ䆁㒗ḋ՟Ǆ䆹䗨ᕘ᳝

䆌໮বൟǄ↨བˈᶤѯᚙ‫މ‬ϟৃ㛑Ꮰᳯ⿟ᑣাᄺд㽚Ⲫℷ՟ⱘ㾘߭ˈᑊϨᇍ䆹㾘߭≵᳝㽚Ⲫ
ⱘᅲ՟Ā咬䅸āഄ䌟ќ݊ড՟ߚ㉏Ǆ↨བˈ䖭⾡ᮍ⊩䗖⫼ѢᄺдⳂᷛὖᗉĀৃ㛑ᗔ᳝ঠ㚲㚢
ⱘᄩཛāǄ೼䖭⾡ᚙ‫މ‬ϟˈℷ՟೼ᭈϾ㕸ԧЁ᠔ऴ↨՟ᕜᇣˈ᠔ҹ㾘߭䲚བᵰাᷛᅮℷ՟ⱘ
㉏߿ˈ㗠ᇍ᠔᳝݊Ҫḋ՟咬䅸ߚ㉏Ўড՟ˈ㾘߭䲚Ӯ᳈Ўㅔ⋕ᯧពǄ䖭ϔᮍ⊩ᇍᑨѢ Prolog
ЁⱘĀ༅䋹৺ᅮāㄪ⬹ˈ݊Ёϡ㛑䆕ᯢЎⳳⱘ㸼䖒ᓣ䛑咬䅸Ў‫؛‬ǄЎњᄺд䖭ḋⱘা乘⌟ऩ
ϾⳂᷛؐⱘ㾘߭ˈ䳔㽕ׂᬍ learnonerule ㅫ⊩ҹ᥹ফ䰘ࡴⱘ䕧ܹব䞣ˈᣛᅮᛳ݈䍷ⱘⳂᷛ
ؐǄϔ㠀ࠄ⡍⅞᷅⢊᧰㋶བҹࠡϔḋ໘⧚ˈা㽕ׂᬍ䆘Ԅ‫؛‬䆒ⱘ performance ᄤ⿟ᑣǄ⊼ᛣ
performance ᅮНЎ䋳➉Ꮖϡ‫ݡ‬䗖⫼Ѣℸᮄⱘ䆒ᅮˈ಴Ўᅗᡞଃ⣀㽚Ⲫড՟ⱘ‫؛‬䆒䌟ќњ᳔
໻ߚؐˈϢଃ⣀㽚Ⲫℷ՟ⱘ‫؛‬䆒ϔḋǄ䖭⾡ᚙ‫މ‬ϟՓ⫼䆹‫؛‬䆒㽚Ⲫℷ՟↨՟ⱘᑺ䞣ᷛ‫߭ޚ‬᳈
Ў䗖ড়Ǆ
ㅫ⊩ⱘ঺ϔবൟᰃϔᮣ⿄Ў AQ ⱘㅫ⊩˄MichalskMichalskietal˅ˈᅗ
↨Ϟ䴶䅼䆎᠔෎Ѣⱘ CN ㅫ⊩᳈ᮽǄབ CN ϔḋˈAQ ᄺдᵤপ㾘߭䲚ˈҹ㽚ⲪⳂᷛߑ᭄Ǆ
✊㗠ˈAQ ϢϞ䴶㒭ߎⱘㅫ⊩᳝ҹϟϡৠ˖佪‫ ܜ‬AQ ⱘ㽚Ⲫㅫ⊩Ϣᑣ߫㽚Ⲫㅫ⊩ϡৠˈ಴Ў
ᅗᯢ⹂ഄᇏᡒ㽚Ⲫϔ⡍ᅮⳂᷛؐⱘ㾘߭ˈ✊ৢᇍ↣ϾⳂᷛؐᄺдϔᵤপ㾘߭䲚Ǆ㄀ѠˈAQ
ㅫ⊩ᄺдऩϾ㾘߭ⱘᮍ⊩гϡৠѢ learnoneruleǄᔧᅗᇍ↣Ͼ㾘߭ᠻ㸠ϔ㠀ࠄ⡍⅞᷅⢊᧰
㋶ᯊˈᅗೈ㒩ऩϾℷ՟ᴹ䖯㸠᧰㋶Ǆ⹂ߛഄ䇈ˈᅗ೼᧰㋶Ёা㗗㰥㹿䆹ℷ՟⒵䎇ⱘሲᗻˈҹ
ᕫࠄ䗤⏤⡍⅞ⱘ‫؛‬䆒Ǆ↣⃵ᄺϔϾᮄ㾘߭ᯊˈᅗҢ䙷ѯ᳾㽚Ⲫⱘḋ՟Ёг䗝ᢽϔϾᮄⱘℷ՟ˈ
԰Ў⾡ᄤҹᣛᓩᮄᵤপ乍ⱘ᧰㋶Ǆ
10.3 ᆜҖ㿺ࡏ䳼φቅ㔉
Ϟ䴶䅼䆎ⱘᑣ߫㽚Ⲫㅫ⊩੠㄀ゴЁⱘ‫އ‬ㄪᷥᄺдㅫ⊩ᦤկњ޴⾡ᄺд㾘߭䲚ⱘᮍ⊩Ǆ
ᴀ㡖㗗㰥䖭ѯ㾘߭ᄺдㅫ⊩䆒䅵Ёⱘ݇䬂ᗱᛇǄ
佪‫ˈܜ‬ᑣ߫㽚Ⲫㅫ⊩↣⃵ᄺдϔϾ㾘߭ˈ⿏এ㽚Ⲫⱘḋ՟✊ৢ೼࠽ԭḋ՟Ϟ䞡໡䖭ϔ䖛
⿟ǄⳌডˈབ ID 䙷ḋⱘ‫އ‬ㄪᷥㅫ⊩Փ⫼ऩϾ᧰㋶䖛⿟ᴹ᧰㋶ৃ᥹ফ‫އ‬ㄪᷥˈ↣ℹᑊ㸠ᄺ
дᭈϾᵤপ乍ⱘ䲚ড়Ǆ಴ℸˈ៥Ӏгৃᇚ ID 䖭ḋⱘㅫ⊩⿄Ўᑊ㸠㽚Ⲫㅫ⊩ˈᇍᑨѢ CN
䖭ḋⱘᑣ߫㽚Ⲫㅫ⊩Ǆાϔ⾡ㅫ⊩↨䕗ད˛ㄨḜ݇䬂೼Ѣ᧰㋶Ё᳔෎ᴀℹ偸П䯈ⱘᏂ߿Ǆ
ID ೼↣ϔ᧰㋶ℹЁḍ᥂ᅗᇍ᭄᥂ѻ⫳ⱘߦߚ䗝ᢽϡৠⱘሲᗻǄⳌডˈCN 䗝ᢽⱘᰃϡৠⱘ
ሲᗻؐᇍˈᮍ⊩ᰃ䗮䖛↨䕗ᅗӀ㽚Ⲫⱘ᭄᥂ᄤ䲚Ǆ㽕ⳟߎ䖭⾡Ꮒ߿ⱘᛣН᠔೼ˈ䳔㽕↨䕗
ϸ⾡ㅫ⊩ЎᄺдࠄⳌৠⱘ㾘߭䲚ড়᠔԰ߎⱘϡৠ䗝ᢽⱘ⃵᭄ǄЎњᄺдࠄ n Ͼ㾘߭ⱘ䲚ড়ˈ
↣Ͼ㾘߭ࠡӊࣙড় k Ͼሲᗻؐ⌟䆩ˈᑣ߫㽚Ⲫㅫ⊩䳔㽕ᠻ㸠 ngk ⃵෎ᴀ᧰㋶ℹˈЎ↣Ͼ㾘
߭ⱘ↣Ͼࠡӊ‫خ‬⣀ゟⱘ䗝ᢽˈ㗠ᑊ㸠㽚Ⲫㅫ⊩ⱘ⣀ゟ䗝ᢽ⃵᭄䖰䖰ᇥѢℸˈ಴Ў೼‫އ‬ㄪᷥЁ
↣Ͼ‫އ‬ㄪ㒧⚍ⱘ䗝ᢽ䛑ᇍᑨњϢ䆹㒧⚍Ⳍ݇㘨ⱘ໮Ͼ㾘߭ⱘࠡӊ䗝ᢽǄᤶ㿔Пˈབᵰ‫އ‬ㄪ㒧
⚍⌟䆩ϔϾ᳝ m ⾡ৃ㛑ؐⱘሲᗻˈ↣⃵‫އ‬ㄪ㒧⚍ⱘ䗝ᢽ䛑ᇍᑨњᇍ m ϾⳌᑨⱘ㾘߭Ё↣Ͼ
㾘߭ⱘࠡӊ䗝ᢽ˄㾕д乬 ˅Ǆ䖭ḋˈᑣ߫㽚Ⲫㅫ⊩˄བ CN˅԰ߎⱘ⣀ゟ䗝ᢽ⃵᭄催Ѣ
ID 䖭ḋⱘᑊ㸠㽚Ⲫㅫ⊩ǄԚાϔ⾡ㅫ⊩᳈དⱘ䯂乬ձ✊ᄬ೼Ǆ݊㾷ㄨձ䌪Ѣ᳝໮ᇥ䆁㒗᭄
᥂ᰃৃ⫼ⱘǄབᵰ᭄᥂䴲ᐌЄᆠˈ䙷Мᅗৃҹᬃᣕᑣ߫㽚Ⲫㅫ⊩᠔㽕∖ⱘ䕗໻᭄䞣ⱘ⣀ゟ䗝
ᢽǄ✊㗠㢹᭄᥂䕗㔎УˈᇍѢϡৠ㾘߭ࠡӊⱘ‫އ‬ㄪĀ݅ѿā߭᳈᳝ᬜǄ঺ϔ㗗㰥೼Ѣ⡍ᅮⱘ
ӏࡵЁᰃ৺Ꮰᳯϡৠⱘ㾘߭⌟䆩ⳌৠⱘሲᗻǄ೼ᑊ㸠㽚Ⲫ‫އ‬ㄪᷥᄺдㅫ⊩ЁӮߎ⦄䖭ḋⱘᚙ
‫މ‬Ǆ೼ᑣ߫㽚Ⲫㅫ⊩Ё߭ϡᄬ೼Ǆ
ϡৠᮍ⊩ⱘ㄀ѠϾⳌᓖП໘೼Ѣ learnonerule ᧰㋶ⱘᮍ৥Ǆ೼Ϟ䴶ᦣ䗄ⱘㅫ⊩Ёˈ᧰

㋶ᰃҢϔ㠀ࠄ⡍⅞ⱘǄ݊ҪᏆ䅼䆎ⱘㅫ⊩˄བ㄀ゴЁⱘ FindS˅ᰃҢ⡍⅞ࠄϔ㠀ⱘǄ೼
ℸᚙ‫މ‬ϟˈҢϔ㠀ࠄ⡍⅞᧰㋶ⱘϔϾӬ⚍೼Ѣা᳝ϔϾᵕ໻ϔ㠀‫؛‬䆒ৃ԰Ў᧰㋶䍋ྟ⚍ˈ㗠
೼໮᭄‫؛‬䆒ぎ䯈Ё᳝ᕜ໮⡍⅞‫؛‬䆒˄བᇍ↣Ͼᅲ՟᳝ϔ‫؛‬䆒˅Ǆ಴Ў᳝䆌໮ᵕ໻⡍⅞‫؛‬䆒ˈ
ህϡ㛑⹂ⶹ䗝ᢽાϔϾ԰Ў᧰㋶ⱘᓔྟ⚍Ǆᠻ㸠Ң⡍⅞ࠄϔ㠀᧰㋶ⱘϔϾ⿄Ў Golem
Muggleton Fengⱘ⿟ᑣ㾷‫އ‬ℸ䯂乬ⱘᮍ⊩ᰃ䱣ᴎ䗝ᢽ໮Ͼℷ՟ˈҹℸЎ߱ྟᴹ䖯
㸠᧰㋶Ǆ೼໮Ͼ䱣ᴎ䗝ᢽЁⱘ᳔Շ‫؛‬䆒԰Ў᳔ৢ㒧ᵰǄ
㄀ϝϾ㽕㗗㰥ⱘᰃ learnonerule ᰃЎϔϾ⫳៤‫⌟ݡ‬䆩˄generatethentest˅᧰㋶ˈ㣗ೈ

Ў᠔᳝ড়⊩ⱘ‫؛‬䆒ˈབ៥Ӏ᥼㤤ⱘᅲ⦄Ё䙷ḋ˗䖬ᰃϔϾḋ՟偅ࡼ˄exampledriven˅᧰㋶ˈ
ҹՓ䆁㒗ḋ՟Ͼԧ㑺ᴳ‫؛‬䆒ⱘ⫳៤Ǆḋ՟偅ࡼ᧰㋶ㅫ⊩ࣙᣀ㄀ゴⱘ FindSǃ‫׭‬䗝⍜䰸ǃ
AQ ㅫ⊩ˈҹঞᴀゴৢ䴶㽕䅼䆎ⱘ Cigol ㅫ⊩Ǆ೼䖭ѯㅫ⊩Ёˈᇍ‫؛‬䆒ⱘ⫳៤៪ׂℷᰃ⬅ऩ
⣀ⱘ䆁㒗ḋ՟偅ࡼⱘˈ㗠Ϩ㒧ᵰᰃϔϾᏆׂℷⱘ‫؛‬䆒ˈՓᇍℸऩϾḋ՟ⱘᗻ㛑ᕫࠄᬍ୘Ǆ䖭
ϡৠѢ㸼 Ё learnonerule ㅫ⊩ⱘ⫳៤‫⌟ݡ‬䆩᧰㋶ˈ݊Ёৢ㓁ⱘ‫؛‬䆒ⱘ⫳៤া෎Ѣ‫؛‬䆒
㸼⼎ⱘ䇁⊩Ǆ೼䖭ѯ‫׭‬䗝‫؛‬䆒⫳៤Пৢ‫ߚݡ‬ᵤ䆁㒗᭄᥂ˈ✊ৢ෎Ѣ䖭ѯ‫؛‬䆒೼ܼ䚼ḋ՟Ϟⱘ
ᗻ㛑ᴹ䖯㸠䗝ᢽǄ⫳៤‫⌟ݡ‬䆩ⱘϔϾ䞡㽕Ӭ⚍ᰃ᧰㋶Ё↣ϔℹⱘ䗝ᢽ䛑෎Ѣ೼䆌໮ḋ՟Ϟⱘ
‫؛‬䆒ᗻ㛑ˈ಴ℸాໄ᭄᥂ⱘᕅડ㹿᳔ᇣ࣪ǄⳌডˈḋ՟偅ࡼㅫ⊩෎ѢऩϾⱘḋ՟ᬍ䖯‫؛‬䆒ˈ
ᅗ᳈ᆍᯧ㹿ϔѯాໄ䆁㒗ḋ՟ᕅડˈ಴ℸᇍ䆁㒗᭄᥂ЁᏂ䫭ⱘ剕Ầᗻ䕗ᏂǄ
㄀ಯϾ㽕㗗㰥ⱘᰃᰃ৺䳔㽕ᇍ㾘߭䖯㸠ৢׂ࠾ҹঞᗢḋׂ࠾Ǆབ೼‫އ‬ㄪᷥᄺдЁϔḋˈ
learnonerule г᳝ৃ㛑ᔶ៤೼䆁㒗᭄᥂Ϟᗻ㛑ᕜདˈԚ೼ҹৢⱘ᭄᥂ЁᕜᏂⱘ㾘߭Ǆ㾷‫އ‬
ⱘࡲ⊩гᰃ೼ᕫࠄ↣Ͼ㾘߭ৢ䖯㸠ৢׂ࠾Ǆ⹂ߛഄ䆆ˈৃҹ⿏এᶤѯࠡӊˈা㽕䖭ᇐ㟈ϡৠ
Ѣ䆁㒗ḋ՟ⱘ⫼Ѣׂ࠾ⱘϔϾḋ՟䲚ড়Ϟⱘᗻ㛑ᦤ催ˈᇍѢৢׂ࠾᳈䆺㒚ⱘ䅼䆎㾕㄀
㡖Ǆ
᳔ৢ㽕㗗㰥ⱘᰃ೼ learnonerule ᯊᣛᓩ᧰㋶ᮍ৥ⱘ㾘߭ᗻ㛑˄performance˅ⱘᅮНǄ

Ꮖ᳝њ৘⾡ϡৠⱘ䆘Ӌߑ᭄ˈᶤѯᐌ⫼ⱘ䆘Ԅߑ᭄ࣙᣀ˖
x Ⳍᇍ乥⥛˄Relative frequency˅ǄҸnҷ㸼㾘߭᠔ऍ䜡ⱘḋ՟᭄ⳂˈҸncҷ㸼݊Ё
ᅗ㛑ℷ⹂ߚ㉏ⱘ᭄ⳂǄ㾘߭ᗻ㛑ⱘⳌᇍ乥⥛Ԅ䅵Ў˖
nc
n
Ⳍᇍ乥⥛㹿⫼Ѣ೼ AQ ⿟ᑣЁ䆘Ԅ㾘߭Ǆ
x ㊒ᑺⱘm-Ԅ䅵˄m-estimate of accuracy˅Ǆ䆹㊒ᑺԄ䅵‫أ‬৥Ѣ㾘߭᠔ᳳᳯⱘ咬䅸
㊒ᑺǄᅗ೼᭄᥂↨䕗㔎УˈᑊϨ㾘߭ᖙ乏೼ᕜᇥⱘḋ՟Ϟ䆘Ԅᯊᐌ⫼ǄҸnc੠n
བϞᅮНˈᑊҸpЎҢᭈϾ᭄᥂䲚Ё䱣ᴎᢑপⱘḋ՟Ϣ䆹㾘߭䌟ќⱘߚ㉏Ⳍৠⱘ
‫ܜ‬偠ὖ⥛˄՟བˈབᵰ 100 Ͼḋ՟Ё᳝ 12 ϾϢ䆹㾘߭ⱘ乘⌟ؐⳌৠˈ䙷Мp=0.12˅Ǆ
᳔ৢˈҸmЎᴗ䞡ˈ៪⿄ᇍℸ‫ܜ‬偠ὖ⥛p䖯㸠ࡴᴗⱘㄝᬜḋ՟᭄ⳂǄᇍ㾘߭㊒ᑺ
ⱘm-Ԅ䅵Ў˖
nc mp
nm
⊼ᛣབᵰ m 㹿䆒Ў 0ˈ߭ m Ԅ䅵বЎϞ䴶ⱘⳌᇍ乥⥛Ԅ䅵Ǆᔧ m Ϟछᯊˈ䳔㽕
᳈໮ⱘḋ՟ᴹ‫ܟ‬᳡䖭Ͼ乘‫؛ܜ‬ᅮⱘ㊒ᑺ pǄm-Ԅ䅵ᑺ䞣⬅ Cestnik & Bratko
˄1991˅ᦤߎˈᅗᏆ⫼Ѣᶤѯ⠜ᴀⱘ CN2 ㅫ⊩Ǆᅗг⫼Ѣ㄀ 6.9.1 㡖䅼䆎ⱘᴈ
㋴䋱৊ᮃߚ㉏఼Ǆ
x ➉(entropy)Ǆ䖭ᰃ೼㸼 10-2 ЁՓ⫼ⱘ performance ᄤ⿟ᑣЁՓ⫼ⱘᑺ䞣ǄҸ S Ў
ऍ䜡㾘߭ࠡӊⱘḋ՟䲚ড়Ǆ➉㸵䞣ⱘᰃ䆹ḋ՟䲚ড়ЁⳂᷛߑ᭄ⱘഛϔᗻǄ䖭䞠
Փ⫼➉ⱘ䋳ؐˈҹՓ䕗Շⱘ㾘߭ᢹ᳝䕗催ⱘߚؐ˖
c
Entropy ( S ) ¦p
i 1
i log 2 pi
݊ЁcЎⳂᷛߑ᭄ৃপⱘϡৠ᭄ؐ䞣ˈpiЎSЁⳂᷛߑ᭄প㄀iϾؐⱘḋ՟᠔ऴ↨
՟ǄϢ㒳䅵ᛣН⌟䆩Ⳍ㒧ড়ˈ➉ᑺ䞣⫼ѢCN2 ㅫ⊩˄Clark & Niblett 1989˅ˈᅗ
гᰃ䆌໮‫އ‬ㄪᷥᄺдㅫ⊩Ёֵᙃ๲Ⲟᑺ䞣ⱘ෎⸔Ǆ
10.4 ᆜҖж䱬㿺ࡏ
ࠡ䴶䅼䆎ⱘㅫ⊩䩜ᇍᄺдੑ乬㾘߭䲚˄े᮴ব䞣ⱘ㾘߭˅Ǆᴀ㡖Ёᇚ㗗㰥ᏺ᳝ব䞣ⱘ㾘
߭ˈ⹂ߛഄ䆆Ўϔ䰊 Horn ᄤহǄП᠔ҹ㗗㰥䖭ḋⱘ㾘߭ˈᰃ಴ЎᅗӀ↨ੑ乬㾘߭᳈᳝㸼ᕕ
㛑࡯ǄᇍѢϔ䰊↉㾘߭ⱘᔦ㒇ᄺд䗮ᐌ㹿⿄Ўᔦ㒇䘏䕥㓪⿟˄InductiveLogicProgrammingˈ
ㅔ‫ ݭ‬ILP˅ˈ಴Ў䖭ϔ䖛⿟ৃⳟ԰Ңḋ՟Ё㞾ࡼ᥼䆎ߎ Prolog ⿟ᑣǄProlog ᰃϔϾ䗮⫼ⱘǃ
೒♉ㄝӋⱘ㓪⿟䇁㿔ˈ݊Ё⿟ᑣ㹿㸼⼎Ўϔ㒘 Horn ᄤহǄ
10.4.1 ж䱬 Horn ᆆਛ
Ў䇈ᯢϔ䰊㸼⼎↨ੑ乬˄᮴ব䞣˅㸼⼎ⱘӬ䍞П໘ˈ㗗㰥ϔϾᄺдӏࡵˈⳂᷛὖᗉᕜㅔ
ऩˈЎ DaughterxyˈᅮН೼᠔᳝ⱘҎ x ੠ y ϞǄDanghterxyⱘؐ೼ x ᰃ y ⱘཇ‫ܓ‬ᯊЎ
ⳳˈ৺߭Ў‫؛‬Ǆ‫؛‬ᅮ↣ϾҎ㹿ᦣ䗄Ўሲᗻ NameMotherFatherMale ੠ FemaleǄ಴ℸ↣
Ͼ䆁㒗ḋ՟ᇚࣙ৿ҹ䖭ѯሲᗻ䖯㸠ⱘᦣ䗄ⱘϸϾҎˈҹঞⳂᷛሲᗻ Daughter ⱘؐǄ՟བˈ
ϟ䴶ЎϔϾℷ՟ˈ݊Ё Sharon Ў Bob ⱘཇ‫ܓ‬Ǆ
Name Sharon Mother Louise Father Bob

Male False Female True
Name Bob Mother Nora Father Victor
Male True Female False Daughter True!
݊Ё↣ϾሲᗻৡϞⱘϟᷛᰃЎњऎߚ䖭ϸϾҎǄ⦄೼ˈབᵰ᧰䲚䆌໮䖭ḋⱘⳂᷛὖᗉ
Daughterⱘ䆁㒗ḋ՟ˈᑊᇚᅗӀᦤկ㒭ϔϾੑ乬㾘߭ᄺд఼ˈབCN ੠Cˈ㒧ᵰᇚЎϔ
㒘䴲ᐌ⡍⅞ⱘ㾘߭བ˖
IF Father BobġName BobġFemale True

THEN Daughter True
㱑✊䖭Ͼ㾘߭ᰃℷ⹂ⱘˈԚᅗ䖛Ѣ⡍⅞њˈ಴ℸᅗᇍҞৢⱘߚ㉏޴Т↿᮴⫼໘Ǆ䯂乬೼
Ѣˈੑ乬㸼⼎ᮍ⊩ϡ㛑໳ᦣ䗄ሲᗻؐП䯈ᅲ䋼݇㋏ǄϢℸϡৠˈՓ⫼ϔ䰊㸼⼎ⱘ⿟ᑣᇚᄺࠄ
ϟ䴶ⱘϔ㠀㾘߭˖
IF FatheryxġFemaleyTHENDaughterxy
݊Ё x ੠ y Ўব䞣ˈᅗӀৃᣛҷӏᛣҎǄ
ϔ䰊 Horn ᄤহ䖬ৃᣛᅮࠡӊЁⱘব䞣ϡߎ⦄೼ৢӊЁⱘ㾘߭Ǆ՟བᇍ GrandDaughter

ⱘ㾘߭Ў˖
IF FatheryzġMotherzyġFemaley

THEN GrandDaughterxy
⊼ᛣ䆹㾘߭Ёⱘব䞣 zˈᅗᣛҷ y ⱘ⠊҆ˈ೼㾘߭ৢӊЁ≵᳝ߎ⦄ǄᔧϔϾব䞣া೼ࠡ

ӊЁߎ⦄ᯊˈ‫؛‬ᅮᅗᰃ㹿ᄬ೼䞣࣪˄existentiallyquantified˅ⱘˈेা㽕ᄬ೼䆹ব䞣ⱘϔϾ
㑺ᴳ㛑⒵䎇ᇍᑨⱘ᭛ᄫˈ䙷М㾘߭ࠡӊህ⒵䎇Ǆ
䖬ৃ㛑೼㾘߭ⱘৢӊ੠ࠡӊЁՓ⫼Ⳍৠⱘ䇧䆡ˈᦣ䗄䗦ᔦⱘ㾘߭Ǆབ೼ᴀゴᓔ༈ⱘϸϾ
㾘߭ᦤկњὖᗉ Ancestorxyⱘ䗦ᔦᅮНǄҹϟᇚᦣ䗄ⱘ ILP ᄺдᮍ⊩Ꮖৃҹᄺд޴⾡ㅔ
ऩⱘ䗦ᔦߑ᭄ˈࣙᣀབϞ䴶ⱘ Ancestor ߑ᭄ҹঞ݊Ҫϔѯߑ᭄ˈབᇍ߫㸼Ё‫ܗ‬㋴䖯㸠ᥦᑣ˗
Ң߫㸼Ё⿏এϔ⡍ᅮ‫ܗ‬㋴˗ᣐ᥹ϸϾ߫㸼Ǆ
10.4.2 ᵥ䈣
೼㒻㓁ҟ㒡ᄺд Horn ᄤহ䲚ⱘㅫ⊩Пࠡˈ‫ܜ‬ҟ㒡ϔѯᔶᓣ䘏䕥Ёⱘ෎ᴀᴃ䇁Ǆ᠔᳝ⱘ

㸼䖒ᓣ⬅ᐌ䞣˄བ BobLouise˅ǃব䞣˄བ xy˅
ǃ䇧䆡ヺো˄བ MarriedGreaterBThan˅
ҹঞߑ᭄ヺো˄བ age˅㒘៤Ǆ䇧䆡੠ߑ᭄ⱘऎ߿೼Ѣ䇧䆡া㛑পؐⳳ៪‫ˈ؛‬㗠ߑ᭄ⱘপؐ
ৃЎӏᛣᐌ䞣Ǆ䖭䞠Փ⫼ᇣ‫ݭ‬ヺো㸼⼎ߑ᭄ˈ໻‫ݭ‬ヺো㸼⼎䇧䆡Ǆ
ᇍѢ䖭ѯヺোˈৃབϟᵘ䗴㸼䖒ᓣ˖乍˄term˅ᰃӏᛣᐌ䞣ǃӏᛣব䞣ǃ៪ᑨ⫼ࠄӏᛣ
乍Ϟⱘӏᛣߑ᭄˄՟བ˖BobxageBobㄝ˅ǄϔϾ᭛ᄫ˄literal˅ᰃᑨ⫼ࠄ乍Ϟⱘӏᛣ䇧
䆡៪݊৺ᅮǄབ MarriedBobLouise GreaterBThanageSueㄝǄབᵰϔ᭛ᄫࣙ৿
ϔ৺ᅮヺো˄˅ˈᇚ݊⿄Ў䋳᭛ᄫ˄negativeliteral˅ˈ৺߭Ўℷ᭛ᄫ˄positiveliteral˅Ǆ
ϔϾᄤহ˄clause˅ᰃ໮Ͼ᭛ᄫⱘӏᛣᵤপˈ݊Ё᠔᳝ⱘব䞣‫؛‬ᅮЎܼ⿄䞣࣪ⱘǄHorn
ᄤহ˄Hornclause˅Ўࣙ৿㟇໮ϔϾℷ᭛ᄫⱘᄤহˈ՟བ˖
H L1 Ln
݊ Ё H Ў ᭛ ᄫ ˈ 㗠 L1 Ln Ў 䋳 ᭛ ᄫ Ǆ ⬅ Ѣ ㄝ ᓣ ( B A) ( B m A) ੠

( A B ) (A B ) ˈϞ䴶ⱘ Horn ᄤহৃ㹿‫ݭ‬Ўབϟᔶᓣ˖
H m ( L1 Ln )
ᅗϢ៥Ӏࠡ䴶ⱘ㾘߭ㄝӋˈᣝ✻ ifthen ⱘ‫⊩ݭ‬བϟ˖
IF L1 Ln , THEN H
᮴⊩‫⊩ݭ‬བԩˈHorn ᄤহⱘࠡӊ L1 Ln 㹿⿄Ўᄤহԧ˄body˅៪㗙ᄤহ‫ܜ‬㸠䆡

˄antecedents˅Ǆ᭛ᄫ H ৢӊ⿄Ўᄤহ༈˄head˅៪ᄤহ᥼䆎˄consequent˅ǄЎখ㗗ᮍ֓ˈ
䖭ѯᅮНҹঞᴀゴৢᇚҟ㒡ⱘὖᗉ೼㸼 Ё߫ߎǄ
㸼 10-3 ϔ䰊䘏䕥Ёⱘ෎ᴀᅮН
↣Ͼড়ᓣ݀ᓣ⬅ᐌ䞣˄constantˈབ Maryǃ23ǃ៪ Joe˅ǃব䞣˄variableˈབ x˅ǃ䇧䆡˄predicateˈ

བ೼ Female(Mary)Ёⱘ Female˅੠ߑ᭄˄functionˈབ ageˈ೼ age(Mary)˅ᵘ៤Ǆ
乍˄term˅Ўӏᛣᐌ䞣ǃӏᛣব䞣ǃ៪ӏᛣᑨ⫼ࠄ乍䲚ড়Ϟⱘߑ᭄Ǆ՟བ Mary, x, age(Mary), age(x)Ǆ
᭛ ᄫ ˄ literal ˅ ᰃ ᑨ ⫼ ࠄ 乍䲚 ড় Ϟ ⱘ ӏ ᛣ 䇧䆡 ៪ ݊ ৺ ᅮ Ǆ ՟ བ Female(Mary), Female(x),
Greater_than(age(Mary), 20)Ǆ
෎ᴀ᭛ᄫ˄ground literal˅ᰃϡࣙ৿ӏԩব䞣ⱘ᭛ᄫ˄བFemale(Joe)˅Ǆ
䋳᭛ᄫ˄negative literal˅ᰃࣙ৿৺ᅮ䇧䆡ⱘ᭛ᄫ˄བFemale(Joe)˅Ǆ
ℷ᭛ᄫ˄positive literal˅ᰃϡࣙ৿৺ᅮヺোⱘ᭛ᄫ˄བ Female(Joe)˅Ǆ
ᄤহ˄clause˅ᰃ໮Ͼ᭛ᄫⱘᵤপᓣˈM1Ģ…Mnˈ݊Ёⱘ᠔᳝ব䞣ᰃܼ⿄䞣࣪ⱘǄ
Horn ᄤহᰃϔϾབϟᔶᓣⱘ㸼䖒ᓣ˖
H m ( L1 Ln )
݊ Ё H LĂLn Ў ℷ ᭛ ᄫ Ǆ H 㹿⿄ Ў Horn ᄤ হ ⱘ ༈ head ៪ ᥼ 䆎 ˄ consequent ˅ Ǆ ᭛ ᄫ ড় প ᓣ
L1 L2 Ln 㹿⿄ЎHornᄤহⱘԧ˄body˅៪㗙‫ܜ‬㸠䆡˄antecedents˅Ǆ
ᇍӏᛣ᭛ᄫ A ੠ Bˈ㸼䖒ᓣ(AĕB)ㄝӋѢ(AĢB)ˈ㗠㸼䖒ᓣ (AġB)ㄝӋѢ(AĢB)Ǆ಴ℸˈϔϾ
Horn ᄤহৃ㹿ㄝᬜഄ‫ݭ‬԰ϟ䴶ⱘᵤপᓣ˖
H L1 Ln
㕂ᤶ˄substitution˅ᰃϔϾᇚᶤѯব䞣᳓ᤶЎᶤѯ乍ⱘߑ᭄Ǆ՟བ㕂ᤶ{x/3, y/z}ᡞব䞣 x ᳓ᤶЎ乍 3
ᑊϨᡞব䞣 y ᳓ᤶЎ乍 zǄ㒭ᅮϔϾ㕂ᤶ©੠ϔ᭛ᄫ Lˈ៥ӀՓ⫼ L©ҷ㸼ᑨ⫼㕂ᤶ©ࠄ L ᕫࠄⱘ㒧ᵰǄ
ϸϾ᭛ᄫL1੠L2ⱘড়ϔ㕂ᤶ˄unifying substitution˅ЎϔϾ㕂ᤶ©ˈՓᕫL1©=L2©Ǆ
10.5 ᆜҖж䱬㿺ࡏ䳼φFOIL
᳝䆌໮ㅫ⊩Ꮖ㹿ᦤߎ⫼Ѣᄺдϔ䰊㾘߭៪ Horn ᄤহǄᴀ㡖Ёᇚҟ㒡 FOIL ⿟ᑣ˄Quinlan

˅ˈᅗՓ⫼ⱘᮍ⊩䴲ᐌ㉏ԐѢࠡ䴶ҟ㒡ⱘᑣ߫㽚Ⲫ੠ learnonerule ㅫ⊩Ǆᅲ䰙ϞˈFOIL
ᰃ䖭ѯ䕗ᮽⱘㅫ⊩೼ϔ䰊㸼⼎Ϟⱘ㞾✊ᠽሩǄᔶᓣ࣪ഄ䆆ˈ⬅ FOIL ᄺдⱘ‫؛‬䆒Ўϔ䰊㾘߭
䲚ˈ݊Ёⱘ㾘߭㉏ԐѢ Horn ᄤহˈԚ᳝ϸϾϡৠ˖佪‫ ⬅ܜ‬FOIL ᄺдⱘ㾘߭↨ϔ㠀ⱘ Horn
ᄤহ᳈ফ䰤ˈ಴Ў᭛ᄫϡ‫ܕ‬䆌ࣙ৿ߑ᭄ヺো˄䖭‫ޣ‬ᇣњ‫؛‬䆒ぎ䯈᧰㋶ⱘ໡ᴖᑺ˅Ǆ݊⃵ˈFOIL
㾘߭↨ Horn ᄤহ᳈᳝㸼ᕕ࡯ˈ಴Ў㾘߭ԧЁⱘ᭛ᄫгৃЎ䋳᭛ᄫǄFOIL Ꮖ㹿ᑨ⫼Ѣ໮⾡䯂
乬乚ඳǄ՟བˈᅗᏆ⫼Ѣᄺдᖿ䗳ᥦᑣㅫ⊩ Quicksort ⱘ䗦ᔦᅮНˈҹঞᄺдҢড়⊩ẟⲬ⢊
ᗕЁऎߚߎ䴲⊩⢊ᗕǄ
FOIL ㅫ⊩೼㸼 Ё߫ߎǄ⊼ᛣ໪ሖᕾ⦃ᇍᑨѢࠡ䴶ᦣ䗄ⱘᑣ߫㽚Ⲫㅫ⊩Ǆᅗ↣⃵ᄺ

дϔϾᮄ㾘߭ˈ✊ৢᇚℸ㾘߭㽚Ⲫⱘℷ՟⿏এˈ✊ৢᄺдϟϔ㾘߭Ǆㅫ⊩ⱘ‫ݙ‬ሖᕾ⦃ᰃࠡ䴶
ⱘ learnonerule ⱘ঺ϔ⾡ᔶᓣˈᅗᏆ㹿ᠽሩҹ䗖ড়໘⧚ϔ䰊㾘߭Ǆ䖬㽕⊼ᛣ FOIL ੠ࠡ䴶
ㅫ⊩ⱘϔѯᖂᇣⱘϡৠǄ⹂ߛഄ䆆ˈFOIL া᧰ᇏ䙷ѯ乘⌟Ⳃᷛ᭛ᄫԩᯊЎ True ⱘ㾘߭ˈ㗠
ࠡ䴶ⱘㅫ⊩᮶᧰ᇏ乘⌟ԩᯊЎ True ⱘ㾘߭ˈг᧰ᇏ乘⌟ԩᯊЎ False ⱘ㾘߭ǄFOIL 䖬ᑨ⫼
њϔϾㅔऩⱘ⠀ቅ᧰㋶ˈ㗠ϡᰃ᷅⢊᧰㋶˄ेᅗᠻ㸠ⱘ᧰㋶ㄝӋѢᆑᑺЎ ⱘ᷅⢊᧰㋶˅Ǆ
㸼 10-4 ෎ᴀⱘ FOIL ㅫ⊩Ǆ
݊Ё㒭ߎњ⫳៤‫׭‬䗝᭛ᄫ Candidate-literal ⱘᮍ⊩੠ FOIL ๲Ⲟ Foil_Gain ⱘᅮНǄ䆹෎ᴀㅫ

⊩ৃ⿡‫ׂخ‬ᬍҹ᳈དഄ໘⧚᳝ాໄ᭄᥂ˈབ᭛Ё᠔ᦣ䗄ⱘǄ
FOIL(Target_predicate, Predicates, Examples)

PosĕExamples Ё Target_predicate Ў True ⱘ៤ਬ
NegĕExamples Ё Target_predicate Ў False ⱘ៤ਬ
Learned_rulesĕ{}
ᔧ Pos ϡぎˈ‫خ‬ҹϟ᪡԰
ᄺдϔϾᮄ㾘߭ NewRule
NewRuleĕ≵᳝ࠡӊⱘ䇧䆡 Target_predicate 㾘߭
NewRuleNegĕNeg
ᔧ NewRuleNeg ϡぎˈ‫خ‬ҹϟ᪡԰
๲ࡴϔϾᮄ᭛ᄫҹ⡍࣪ NewRule
Candidate_literalsĕᇍ NewRule ⫳៤‫׭‬䗝ᮄ᭛ᄫˈ෎Ѣ Predicates
Best_literalĕ arg max Foil _ Gain( L, NewRule)
LCandidate _ literals
ᡞ Best_literal ࡴܹࠄ NewRule ⱘࠡӊ

NewRuleNegĕNewRuleNeg Ё⒵䎇 NewRule ࠡӊⱘᄤ䲚
Learned_rulesĕLearned_rules+NewRule
PosĕPos-{㹿 NewRule 㽚Ⲫⱘ Pos ៤ਬ}
䖨ಲ Learned_rules
Ў⧚㾷⬅ FOIL ᠻ㸠ⱘ‫؛‬䆒ぎ䯈᧰㋶ˈ᳔དᇚ݊ⳟ԰ᰃሖ⃵࣪ⱘǄFOIL ໪ሖᕾ⦃Ё↣

⃵ᇚࡴܹϔϾᮄⱘ㾘߭ࠄ݊ᵤপᓣ‫؛‬䆒 LearnedBrules ЁএǄ↣Ͼᮄ㾘߭ⱘᬜᵰᰃ䗮䖛ࡴܹ
ϔϾᵤপ乍⊯࣪ᔧࠡⱘᵤপ‫؛‬䆒˄े๲ࡴ݊ߚ㉏Ўℷ՟ⱘᅲ՟᭄˅Ǆ೼䖭ϔሖ⃵Ϟⳟˈ䖭ᰃ
ϔϾ‫؛‬䆒ぎ䯈Ёⱘ⡍⅞ࠄϔ㠀ⱘ᧰㋶䖛⿟ˈᅗᓔྟѢ᳔⡍⅞ⱘぎᵤপᓣˈ೼‫؛‬䆒䎇໳ϔ㠀ҹ
㟇㽚Ⲫ᠔᳝ℷ՟ᯊ㒜ℶǄFOIL ⱘ‫ݙ‬ሖᕾ⦃ᠻ㸠ⱘᰃϔ䕗㒚㉦ᑺⱘ᧰㋶ˈҹ⹂ᅮ↣Ͼᮄ㾘߭
ⱘ⹂ߛᅮНǄ䆹‫ݙ‬ሖᕾ⦃೼঺ϔ‫؛‬䆒ぎ䯈Ё᧰㋶ˈᅗࣙ৿᭛ᄫⱘড়পˈҹᡒࠄϔϾড়পᓣᔶ
៤ᮄ㾘߭ⱘࠡӊǄ೼䖭Ͼ‫؛‬䆒ぎ䯈Ёˈᅗᠻ㸠ⱘᰃϔ㠀ࠄ⡍⅞ⱘ⠀ቅ᧰㋶ˈᓔྟѢ᳔ϔ㠀ⱘ
ࠡӊ˄ぎࠡӊ˅ˈ✊ৢ๲ࡴ᭛ᄫҹՓ㾘߭⡍࣪Ⳉࠄ݊䙓ᓔ᠔᳝ⱘড՟Ǆ
೼ FOIL ੠ࠡ䴶ⱘᑣ߫㽚Ⲫ੠ learnonerule ㅫ⊩П䯈᳝ϸϾ᳔ᅲ䋼ⱘϡৠˈᅗᴹ⑤Ѣ

ℸㅫ⊩ᇍϔ䰊㾘߭໘⧚ⱘ䳔∖Ǆ䖭ѯϡৠ೼Ѣ˖
ˊ೼ᄺд↣Ͼᮄ㾘߭ⱘϔ㠀ࠄ⡍⅞᧰㋶ЁˈFOIL Փ⫼њϡৠⱘ㒚㡖ℹ偸ᴹ⫳៤㾘߭ⱘ
‫׭‬䗝⡍࣪ᓣǄ䖭ϔϡৠᰃЎњ໘⧚㾘߭ࠡӊЁ৿᳝ⱘব䞣Ǆ
ˊFOIL Փ⫼ⱘᗻ㛑ᑺ䞣 FoilBGain ϡৠѢ㸼 Ёⱘ➉ᑺ䞣Ǆ䖭ᰃЎњऎߚ㾘߭ব

䞣ⱘϡৠ㑺ᴳˈҹঞ⬅Ѣ FOIL া᧰ᇏ㽚Ⲫℷ՟ⱘ㾘߭Ǆ
ϟ䴶ϸ㡖ᇚ᳈䆺㒚ഄ㗗㰥䖭ϸϾϡৠП໘Ǆ
10.5.1 FOIL ѣⲺُ䘿⢯ौᕅ⭕ᡆ
Ўњ⫳៤ᔧࠡ㾘߭ⱘ‫׭‬䗝⡍࣪ᓣˈFOIL ⫳៤᭄Ͼϡৠⱘᮄ᭛ᄫˈ↣Ͼৃ㹿ऩ⣀ഄࡴࠄ
㾘߭ࠡӊЁǄ᳈㊒⹂ഄ䆆ˈ‫؛‬ᅮᔧࠡ㾘߭Ў˖
PxxĂxkĕLĂLn
݊ЁLĂLnЎᔧࠡ㾘߭ࠡӊЁⱘ᭛ᄫˈ㗠PxxĂxkЎ㾘߭༈˄៪ৢӊ˅ǄFOIL⫳

៤䆹㾘߭ⱘ‫׭‬䗝⡍࣪ᓣⱘᮍ⊩ᰃ㗗㰥ヺড়ϟ߫ᔶᓣⱘᮄ᭛ᄫLn˖
x Q(v1,…,vr)ˈ݊ЁQЎ೼PredicatesЁߎ⦄ⱘӏᛣ䇧䆡ৡˈᑊϨvi᮶ৃЎᮄব䞣ˈ
гৃЎᏆ೼㾘߭Ё᳝ⱘব䞣ǄviЁ㟇ᇥϔϾᖙ乏ᰃᔧࠡ㾘߭ЁᏆ᳝ⱘব䞣Ǆ
x Equal(xj, xk)ˈ݊Ёxj੠xkЎ㾘߭ЁᏆ᳝ⱘব䞣Ǆ
x Ϟ䗄ϸ⾡᭛ᄫⱘ৺ᅮǄ
Ў䇈ᯢ䖭ϔ⚍ˈ㗗㰥ᕙᄺдⱘ㾘߭ᰃ乘⌟Ⳃᷛ᭛ᄫ GrandDanghterxyˈ݊Ёᦣ䗄ḋ

՟ⱘ݊Ҫ䇧䆡ࣙᣀ Father ੠ FemaleǄFOIL Ёⱘϔ㠀ࠄ⡍⅞᧰㋶ᓔྟѢ᳔ϔ㠀ⱘ㾘߭˖
GrandDaughterxyĕ
ᅗᮁ㿔ᇍӏᛣ x ੠ yˈGrandDaughter 䛑ЎⳳǄЎ⡍࣪䖭ϔ߱ྟ㾘߭ˈϞ䴶ⱘ䖛⿟⫳៤

ϟ ߫ ᭛ ᄫ ԰ Ў ᇚ ⏏ ࡴ ࠄ 㾘 ߭ ࠡ ӊ Ё ⱘ ‫ ׭‬䗝 ᭛ ᄫ ˖ Equalxy Femalex Femaley
FatherxyFatheryxFatherxzFatherzxFatheryzFatherzyˈҹঞ䖭
ѯ᭛ᄫⱘ৺ᅮ˄՟བ˖Equalxy˅ Ǆ⊼ᛣ䖭䞠 z ᰃϔᮄব䞣ˈ㗠 x ੠ y ᰃᔧࠡ㾘߭ЁᏆ᳝
ⱘǄ
⦄೼‫؛‬ᅮ೼Ϟ䗄᭛ᄫЁ FOIL 䋾ှഄ䗝ᢽњ Fatheryz԰Ў᳔᳝Ꮰᳯⱘ᭛ᄫˈᕫࠄϔ

Ͼ䕗⡍⅞ⱘ㾘߭˖
GrandDaughterxyĕFatheryz
೼⫳៤Ў䖯ϔℹ⡍࣪䆹㾘߭ⱘ‫׭‬䗝᭛ᄫᯊˈFOIL ⦄㽕㗗㰥ⱘ᭛ᄫ䰸Ϟϔℹ᠔᳝᭛ᄫП
໪ˈ䖬㽕ࡴϞ᭛ᄫˈFemalezEqualzxEqualzyFatherzwFatherwzˈҹ
ঞᅗӀⱘ৺ᅮǄП᠔ҹࡴϞ䖭ѯ᭛ᄫᰃ಴Ў೼ࠡϔℹব䞣 z 㹿ࡴࠄ㾘߭Ёˈ᠔ҹ FOIL 㽕㗗
㰥๲ࡴ঺ϔϾᮄব䞣 wǄ
བᵰ FOIL 䖭ᯊ䗝ᢽњ Fatherzxˈ✊ৢ೼ϟϔᕾ⦃䗝ᢽњ᭛ᄫ Femaleyˈᇚᕫࠄ

ϟ䴶ⱘ㾘߭Ǆᅗা㽚Ⲫℷ՟ˈ಴ℸˈ㒜ℶњ䖯ϔℹ᧰㋶䆹㾘߭ⱘ⡍࣪ᓣⱘ䖛⿟˖
GrandDaughterxyĕFatheryzġFatherzxġFemaley
䖭ᯊˈFOIL ᇚӮ⿏এ㹿䆹ᮄ㾘߭㽚Ⲫⱘ᠔᳝ḋ՟Ǆབᵰ䖬᳝᳾㽚Ⲫⱘℷ՟ˈㅫ⊩ᇚᓔ
ྟϟϔϾϔ㠀ࠄ⡍⅞᧰㋶ҹ㦋ᕫᮄⱘ㾘߭Ǆ
10.5.2 ᕋሲ FOIL Ⲻᩒ㍘
㽕೼↣ϔℹЁҢ‫׭‬䗝᭛ᄫЁ䗝ᢽ᳔᳝Ꮰᳯⱘ᭛ᄫˈFOIL ೼䆁㒗᭄᥂Ϟ⌟䞣㾘߭ⱘᗻ㛑Ǆ
೼ℸ䖛⿟Ёˈᅗ㗗㰥ᔧࠡ㾘߭Ё↣Ͼব䞣ⱘৃ㛑ⱘ㑺ᴳǄЎ䇈ᯢ䖭ϔ䖛⿟ˈ‫⃵ݡ‬㗗㰥ᄺдⳂ
ᷛ᭛ᄫ GrandDaughterxyⱘ㾘߭䲚ⱘ՟ᄤǄ‫؛‬ᅮ䆁㒗᭄᥂ࣙ৿ϟ߫ⱘㅔऩⱘᮁ㿔䲚ড়ˈ
݊ЁՓ⫼㑺ᅮⱘ Pxyৃ㹿䇏԰Āx ⱘ P ᰃ yāǄ
GrandDaughterVictorSharonFatherSharonBob FatherTomBob

FemaleSharon FatherBobVictor
ᇍ䖭Ͼᇕ䯁ⱘϪ⬠䖬㽕԰ϔ‫؛‬ᅮˈेӏԩ⍝ঞࠄ䇧䆡 GrandDaughterFatherFemale
ঞᐌ䞣 VictorˈSharonˈBob ੠ Tom ⱘ᭛ᄫˈ㢹ᅗӀ≵᳝೼Ϟ䴶߫ߎˈ߭㹿‫؛‬ᅮЎ False˄བˈ
៥Ӏৃҹ䱤৿ഄᮁ㿔GrandDaughterTomBobˈGrandDaughterVictorVictorㄝ˅Ǆ
Ў䗝ᢽᔧࠡ㾘߭ⱘ᳔Շ⡍࣪ᓣˈFOIL 㗗㰥㾘߭ব䞣㑺ᴳࠄ䆁㒗ḋ՟Ё৘ᐌ䞣ⱘ↣⾡ϡ
ৠⱘᮍᓣǄ՟བˈ೼߱ྟℹ㾘߭Ў˖
GrandDanghterxyĕ
㾘߭ব䞣≵᳝㹿ӏԩࠡӊ㑺ᴳˈ಴ℸৃҹ㑺ᴳࠄಯϾᐌ䞣 VictorSharonBob ੠ Tom

ⱘӏᛣ㒘ড়Ǆ䖭䞠Փ⫼䆄ো^xBobySharon`ҷ㸼⡍ᅮⱘব䞣㑺ᴳˈेᇚ↣Ͼব䞣᯴ᇘࠄ
ϔᐌ䞣ⱘ㕂ᤶǄ Ͼᐌ䞣ᇍℸ߱ྟ㾘߭ৃѻ⫳ ⾡ৃ㛑ⱘ㑺ᴳǄ㗠㑺ᴳ^xVictorySharon`
ᇍᑨⱘᰃℷ՟㑺ᴳˈ಴Ў䆁㒗᭄᥂Ёࣙ৿ᮁ㿔 GrandDaughterVictorSharonǄ೼ℸ՟Ёˈ
݊Ҫ ⾡㾘߭‫ܕ‬䆌ⱘ㑺ᴳ˄՟བ㑺ᴳ^xBobyTom`˅㒘៤њ㾘߭ⱘ৺ᅮ䆎᥂ˈ಴Ў䆁㒗
᭄᥂Ё≵᳝ᅗӀⳌᑨⱘᮁ㿔Ǆ
೼↣ϔ䰊↉ˈ㾘߭ⱘ䆘Ԅ෎Ѣ䖭ѯℷ՟੠ড՟ব䞣㑺ᴳˈ㗠៥Ӏؒ৥Ѣ䗝ᢽⱘᰃᢹ᳝䕗
໮ℷ՟㑺ᴳ㗠䕗ᇥড՟㑺ᴳⱘ㾘߭Ǆᔧᮄ᭛ᄫࡴܹࠄ㾘߭Ёˈ㑺ᴳⱘ䲚ড়ᇚᬍবǄ⊼ᛣᔧϔ
᭛ᄫࡴܹৢˈᅗᓩܹњϔϾᮄব䞣ˈ䙷М㾘߭ⱘ㑺ᴳ䭓ᑺᇚ๲䭓Ǆ՟བˈ㢹Fatheryzࡴ
ܹࠄϞ䗄㾘߭ˈ䙷М߱ྟⱘ㑺ᴳ^xVictorySharon`ᇚবЎ᳈䭓ⱘ^xVictorySharon
zBob`Ǆ䖬㽕⊼ᛣབᵰᮄব䞣ৃ㑺ᴳࠄ໮Ͼϡৠⱘᐌ䞣ˈ䙷МϢᠽሩৢ㾘߭Ⳍऍ䜡ⱘ㑺ᴳⱘ
᭄Ⳃᇚ໻ѢϢॳྟ㾘߭ऍ䜡ⱘ᭄ⳂǄ
FOIL Փ⫼䆘Ԅߑ᭄ҹԄ䅵๲ࡴᮄ᭛ᄫⱘᬜ⫼ˈᅗ෎Ѣࡴܹᮄ᭛ᄫࠡৢⱘℷ՟੠ড՟㑺
ᴳ᭄ⳂǄ᳈㊒⹂ഄ䆆ˈ㗗㰥ᶤ㾘߭ Rˈ੠ϔϾৃ㛑㹿ࡴࠄ R ⱘ㾘߭ԧⱘ‫׭‬䗝᭛ᄫ LǄҸ R嘔
Ўࡴܹ᭛ᄫ L ࠄ㾘߭ R ৢ⫳៤ⱘ㾘߭ǄFoilBGainLRⱘؐᅮНЎ˖
§ p1 p0 ·
Foil _ Gain( L, R) { t ¨¨ log 2 log 2 ¸¸ ˄˅
© p1 n1 p 0 n0 ¹
݊ЁpЎ㾘߭Rⱘℷ՟㑺ᴳ᭄ⳂˈnЎRⱘড՟㑺ᴳ᭄Ⳃˈpᰃ㾘߭R嘔ⱘℷ՟㑺ᴳ᭄ˈnЎ
㾘߭R嘔ⱘড՟㑺ᴳ᭄ⳂǄ᳔ৢˈtᰃ೼ࡴܹ᭛ᄫLࠄRৢҡᮻ㛑㽚Ⲫⱘ㾘߭Rⱘℷ՟㑺ᴳ᭄Ǆ
ᔧࡴܹLᓩܹњϔϾᮄব䞣ࠄRЁᯊˈা㽕೼R嘔ⱘ㑺ᴳЁⱘᶤѯ㑺ᴳᠽሩњॳྟⱘ㑺ᴳˈᅗ
Ӏҡ✊㛑㹿㽚ⲪǄ
p0
䆹 FoilBGain ߑ᭄ৃҹ⫼ֵᙃ䆎ᴹㅔऩഄ㾷䞞Ǆᣝ✻ֵᙃ䆎ⱘ⧚䆎ˈ log 2 ᰃ
p 0 n0
p1
Ўњᇍ㾘߭ R 㛑㽚Ⲫⱘӏᛣℷ՟㑺ᴳ㓪ⷕ᠔䳔ⱘ᳔ᇣԡ᭄ǄⳌԐⱘˈ log 2 ᰃᇍ㾘
p1 n1
߭ R嘔㛑㽚Ⲫⱘӏᛣℷ՟㑺ᴳ㓪ⷕⱘ᳔ᇣԡ᭄Ǆ⬅Ѣ t ᰃ R 㛑㽚Ⲫⱘℷ՟㑺ᴳЁҡֱ⬭೼ R嘔
Ёⱘ㑺ᴳˈFoilBGainLRৃ㹿ⳟ԰˖Ўњ㓪ⷕ R ⱘ᠔᳝ℷ՟㑺ᴳⱘߚ㉏᠔䳔ⱘܼ䚼ԡ᭄⬅
Ѣ L ᏺᴹⱘ‫ޣ‬ᇥǄ
10.5.3 ᆜҖ䙈ᖈ㿺ࡏ䳼
೼Ϟ䴶ⱘ䅼䆎Ёˈ៥Ӏᗑ⬹њࡴܹࠄ㾘߭ԧЁⱘᄤহЎⳂᷛ䇧䆡ᴀ䑿˄े೼㾘߭༈Ёߎ
⦄ⱘ䇧䆡˅ⱘৃ㛑ᗻǄ✊㗠ˈབᵰ೼ Predicates ⱘ䕧ܹ߫㸼Ёࣙ৿Ⳃᷛ䇧䆡ˈFOIL ೼⫳៤
‫׭‬䗝᭛ᄫᯊᖙ乏㗗㰥ᅗǄ䖭‫ܕ‬䆌ᅗѻ⫳䗦ᔦⱘ㾘߭üüे೼㾘߭༈੠㾘߭ԧЁՓ⫼Ⳍৠ䇧䆡
ⱘ㾘߭Ǆ՟བˈಲᖚ Ancestor ݇㋏ⱘ䗦ᔦᅮНǄ⬅ϟ䴶ⱘ㾘߭䲚㸼⼎˖
IF Parentxy THENAncestorxy

IF ParentxzġAncestorzyTHENAncestorxy
㒭ᅮ䗖ᔧⱘ䆁㒗ḋ՟䲚ˈ䖭ϸϾ㾘߭ⱘᄺдৃᣝ✻㉏ԐѢϞ䴶 GrandDaughter ⱘℹ偸Ǆ

⊼ᛣা㽕 Ancestor ࣙ৿೼ Predicates ߫㸼Ёˈৢ㗙‫އ‬ᅮњ೼⫳៤ᮄ᭛ᄫᯊ㽕㗗㰥ⱘ䇧䆡ˈϞ
䴶㄀ѠϾ㾘߭ህࣙ৿೼ FOIL ⱘ↣⃵ⱘ᧰㋶ЁǄᔧ✊䆹⡍ᅮ㾘߭ᰃ৺㛑㹿ᄺдࠄপ‫އ‬Ѣ䖭ѯ
⡍ᅮⱘᄤহ೼ FOIL ⱘ䋾ှ᧰㋶⏤䖯⡍⅞ⱘ㾘߭Ё㛑↨݊Ҫ‫׭‬䗝䆘ߚ᳈催ǄCameronJones
Quinlan˄˅䅼䆎њ޴Ͼ՟ᄤˈ݊Ё FOIL 㛑៤ࡳഄথ⦄䗦ᔦⱘ㾘߭䲚ǄҪӀ䖬䅼䆎њৃ
㛑ѻ⫳ⱘ䞡㽕䯂乬ˈ↨བབԩ䙓‫ܡ‬೼ᄺд㾘߭䲚Ёѻ⫳᮴䰤䗦ᔦǄ
10.5.4 FOIL ቅ㔉
ὖᣀⱘ䇈ˈFOIL ᠽሩњ CN ⱘᑣ߫㽚Ⲫㅫ⊩ˈҹ໘⧚㉏ԐѢ Horn ᄤহⱘϔ䰊㾘߭ᄺ

д䯂乬ǄЎᄺд䖭ḋⱘ㾘߭ˈFOIL ᠻ㸠ϔ㠀ࠄ⡍⅞᧰㋶ˈ↣ℹ๲ࡴϔϾᮄⱘ᭛ᄫࠄ㾘߭ࠡ
ӊЁএǄᮄⱘ᭛ᄫৃЎ㾘߭ࠡӊ៪ৢӊЁᏆ᳝ⱘব䞣ˈ៪㗙ৃᰃϔᮄব䞣Ǆ೼↣ϔℹˈᅗՓ
⫼ᓣ Ёⱘ FoilBGrain ߑ᭄೼‫׭‬䗝ᮄ᭛ᄫЁ䖯㸠䗝ᢽǄབᵰᮄ᭛ᄫৃᣛ৥Ⳃᷛ䇧䆡ˈ䙷
Мॳ߭Ϟ FOIL ৃᄺдࠄ䗦ᔦ㾘߭䲚Ǆ㱑✊䖭ѻ⫳њ঺ϔ໡ᴖᗻˈे䙓‫ܡ‬㾘߭䲚ⱘ᮴䰤䗦ᔦˈ
Ԛ FOIL Ꮖ೼ᶤѯᚙ‫މ‬ϟ៤ࡳഄ⫼Ѣᄺд䗦ᔦ㾘߭䲚Ǆ
೼䆁㒗᭄᥂᮴ాໄⱘᚙ‫މ‬ϟˈFOIL ৃᣕ㓁ഄ๲ࡴᮄ᭛ᄫࠄ㾘߭ЁˈⳈࠄᅗϡ㽚Ⲫӏԩ
ড՟ǄЎ໘⧚᳝ాໄ᭄᥂ˈ᧰㋶ⱘ㒜ℶ䳔㽕೼㾘߭㊒ᑺǃ㽚Ⲫᑺ੠໡ᴖᗻП䯈԰ߎᡬЁǄ
FOIL Փ⫼᳔ᇣᦣ䗄䭓ᑺⱘᮍ⊩ᴹՓ㾘߭๲䭓㒜ℶˈᮄⱘ᭛ᄫা೼ᅗӀⱘᦣ䗄䭓ᑺⷁѢᅗӀ
᠔㾷䞞ⱘ᭄᥂ⱘᦣ䗄䭓ᑺᯊᠡ㹿ࡴܹǄ䆹ㄪ⬹ⱘ㒚㡖⬅ Quinlan˄˅㒭ߎǄ঺໪ǄFOIL
ᇍ↣Ͼᄺࠄⱘ㾘߭䖯㸠ৢׂ࠾ˈՓ⫼㄀ゴ‫އ‬ㄪᷥЁⳌৠⱘ㾘߭ৢׂ࠾ㄪ⬹Ǆ
10.6 ֒Ѱ䘼╊㔄Ⲻᖈ㓩
ᔦ㒇䘏䕥㓪⿟᳝঺ϔ⾡ᅠܼϡৠⱘ䗨ᕘˈᅗ෎ѢϔϾㅔऩⱘџᅲ˖ेᔦ㒇ᰃⓨ㒢ⱘ䗚䖛
⿟Ǆϔ㠀ᴹ䇈ˈᴎ఼ᄺд⍝ঞⱘᰃབԩᓎゟ㛑㾷䞞㾖ᆳ᭄᥂ⱘ⧚䆎Ǆ㒭ᅮᶤѯ᭄᥂D੠ϔѯ
ϡᅠᭈⱘ㚠᱃ⶹ䆚Bˈᄺд䖛⿟ৃ㹿ᦣ䗄Ў⫳៤ϔϾ‫؛‬䆒hˈᅗϢBϔ䍋㾷䞞њDǄ᳈㊒⹂ഄ
䆆ˈ‫؛‬ᅮབ䗮ᐌ䙷ḋ䆁㒗᭄᥂DЎ䆁㒗ḋ՟ⱘ䲚ড়ˈ↣Ͼḋ՟ᔶᓣЎǇxifxiǈǄ䖭䞠xiҷ
㸼㄀iϾ䆁㒗ᅲ՟ˈ㗠fxiҷ㸼ᅗⱘⳂᷛؐǄ䙷Мᄺд䖛⿟ህᰃЎњথ⦄ϔϾ‫؛‬䆒hˈՓ↣Ͼ
䆁㒗ᅲ՟xiⱘߚ㉏fxiᰃҢ‫؛‬䆒hǃxiⱘᦣ䗄ǃঞ㋏㒳ⶹ䘧ⱘӏᛣ㚠᱃ⶹ䆚BЁⓨ㒢⌒⫳Ǆ
(¢ xi , f ( xi )² D)( B h xi ) ƅfxi ˄˅
㸼䖒ᓣXƅY䇏԰ĀYҢXЁⓨ㒢⌒⫳ā
ˈ៪㗙ЎĀX⎉㭈˄entail˅YāǄ㸼䖒ᓣ ᦣ䗄
њᄺдࠄⱘ‫؛‬䆒hᖙ乏⒵䎇ⱘ㑺ᴳˈेᇍ↣Ͼ䆁㒗ᅲ՟xiˈⳂᷛߚ㉏fxiᖙ乏ҢBǃh੠xiЁⓨ
㒢⌒⫳Ǆ
㗗㰥ϔϾ՟ᄤˈ݊ЁᕙᄺдⱘⳂᷛὖᗉᰃĀϸϾҎuv!Ё u ⱘᄽᄤᰃ vāˈᅗ㸼⼎њ䇧

䆡 ChilduvǄ‫؛‬ᅮ㒭ߎњऩϾℷ՟ ChildBobSharonˈ݊Ёᅲ՟ᦣ䗄Ў᭛ᄫ MaleBob
FemaleSharon੠ FatherSharonBobǄ䖯ϔℹ‫؛‬ᅮ᳝㚠᱃ⶹ䆚 ParentuvĕFatheruvǄ
ৃᇚℸᚙᔶᣝᓣ ᦣ䗄བϟ˖
xi MaleBobFemaleSharonFatherSharonBob

fxi ChildBobSharon
B ParentuvĕFatheruv
೼ℸᚙ‫މ‬ϟˈ䆌໮‫؛‬䆒Ё⒵䎇㑺ᴳ ( B h xi ) ƅfxiⱘϸϾ‫؛‬䆒Ў˖
hChilduvĕFathervu

hChilduvĕParentvu
⊼ᛣⳂᷛ᭛ᄫChildBobSharonᰃ⬅hġxi⎉㭈ˈϡ䳔㽕㚠᱃ⶹ䆚BǄ✊㗠ᇍѢ‫؛‬䆒hˈ

ᚙ‫މ‬᳝ѯϡৠǄⳂᷛChildBobSharonᰃҢBġhġxiЁ⌒⫳ˈ㗠ϡᰃऩ⣀ⱘhġxiЁ⌒⫳Ǆ
䆹՟䇈ᯢњ㚠᱃ⶹ䆚ⱘ԰⫼ˈे䩜ᇍ㒭ᅮⱘ䆁㒗᭄᥂ᠽሩৃ᥹ফⱘ‫؛‬䆒䲚ড়Ǆᅗ䖬䇈ᯢᮄⱘ
䇧䆡˄བParent˅ᗢḋᓩܹࠄ‫؛‬䆒˄བh˅ЁˈेՓℸ䇧䆡ϡ೼ॳᴹⱘᅲ՟xiᦣ䗄ЁǄ䖭ϔ
෎Ѣ㚠᱃ⶹ䆚ᠽሩ䇧䆡䲚ড়ⱘ䖛⿟ˈ䗮ᐌ⿄Ўᓎ䆒ᗻᔦ㒇constructiveinductionǄ
ᓣ ⱘᛣН೼Ѣᅗᡞᄺд䯂乬㕂Ѣⓨ㒢᥼⧚੠ᔶᓣ䘏䕥ⱘḚᶊПϟǄᇍѢੑ乬䘏䕥
੠ϔ䰊䘏䕥ˈ᳝ϔѯᏆ⧚㾷ᕫᕜདⱘㅫ⊩ৃ㞾ࡼⓨ㒢Ǆ᳝䍷ⱘᰃˈ᳝ৃ㛑߽⫼ⓨ㒢᥼⧚ⱘ䗚
䖛⿟ˈҹՓᔦ㒇⊯࣪ⱘ䖛⿟㞾ࡼ࣪ǄᇍĀᔦ㒇ৃ⬅ড䕀ⱘⓨ㒢ᅲ⦄ā䖭ϔ㾖⚍ⱘ⋲ᙝ佪‫ߎܜ‬
⦄Ѣ Ϫ㑾ⱘ㒣⌢ᄺᆊ WSJevonsˈҪ‫˖ࠄݭ‬
ᔦ㒇ᅲ䰙Ϟᰃⓨ㒢ⱘ䗚᪡԰ˈ㗠Ϩϡ㛑ᛇ䈵≵᳝݊ЁϔϾᯊˈ঺ϔϾӮᄬ೼Ǆ಴ℸϡӮ᳝ાϔ
Ͼ᳈䞡㽕ⱘ䯂乬Ǆ䇕Ӯ䯂ࡴ⊩੠‫⊩ޣ‬ЁાϔϾᰃ↨䕗䞡㽕ⱘ᭄ᄺ᪡԰ਸ਼˛Ԛৠᯊˈ೼ϔ᪡԰੠ᅗⱘ䗚
᪡԰П䯈ˈ݊䲒ᑺ᳝ᕜ໻ⱘᏂᓖ˗ĂĂᖙ乏ᡓ䅸ˈᔦ㒇ߚᵤ೼䲒ᑺ੠໡ᴖᑺᮍ䴶䛑䖰䖰໻Ѣӏԩⓨ㒢
䯂乬Ǆ˄Jevons 1874˅
ᴀゴⱘ࠽ԭ䚼ߚᇚ᥶お䖭⾡ᡞᔦ㒇ⳟ៤䗚ⓨ㒢ⱘ㾖⚍Ǆ៥Ӏ೼䖭䞠᠔ᛳ݈䍷ⱘᰃϔ㠀䯂
乬ᰃ䆒䅵ϔϾ䗚⎉㭈ㅫᄤinverseentailmentoperatorǄϔϾ䗚⎉㭈ㅫᄤOBDՓ⫼䆁㒗
᭄᥂D ^xifxi!`੠㚠᱃ⶹ䆚B԰Ў䕧ܹˈᑊϨ䕧ߎϔ‫؛‬䆒h⒵䎇ᓣ Ǆ
OBD h݊Ё (¢ xi , f ( xi )² D)( B h xi ) ƅfxi
ᔧ✊Ӯ᳝ᕜ໮ϡৠⱘ‫؛‬䆒h⒵䎇 (¢ xi , f ( xi )² D)( B h xi ) ƅfxiǄ೼ILPЁ䗝ᢽ

‫؛‬䆒ⱘᐌ⫼ਃথᓣ㾘߭Ўձ䌪Ѣ᳔ᇣᦣ䗄䭓ᑺ‫˄߭ޚ‬㾕㡖˅Ǆ
ᇚᄺдӏࡵᔶᓣ࣪ЎᇏᡒϔϾ‫؛‬䆒hՓ݊⒵䎇 (¢ xi , f ( xi )² D)( B h xi ) ƅfxiˈ

᳝䆌໮᳝਌ᓩ࡯ⱘ⡍⚍˖
x 䖭⾡݀ᓣࣙ৿њϔ⾡᱂䘡ⱘᄺдᅮНᮍ⊩ˈेᇏᡒᶤϾϔ㠀ὖᗉˈᅗϢ㒭ᅮⱘ
䆁㒗ḋ՟Ⳍᢳড়Ǆ݊Ё䆁㒗ḋ՟ᇍᑨ≵᳝㚠᱃ⶹ䆚 B ᯊⱘ⡍⅞ᚙ‫މ‬Ǆ
x ℸ݀ᓣ䗮䖛ᓩܹ㚠᱃ⶹ䆚BˈৃҹᇍԩᯊϔϾ‫؛‬䆒ৃ㹿⿄԰Āᢳড়ā䆁㒗᭄᥂䖯
㸠᳈‫ⱘߚܙ‬ᅮНǄ㟇ℸЎℶˈ៥ӀϔⳈ䛑ҙҙ෎Ѣ‫؛‬䆒੠᭄᥂ⱘᦣ䗄ᴹ⹂ᅮϔ
‫؛‬䆒˄བ⼲㒣㔥㒰˅ᰃ৺ᢳড়᭄᥂ˈ㗠ϡձ䌪Ѣᕙᄺдⱘӏࡵ乚ඳǄⳌডˈ䖭
⾡ᔶᓣ‫ܕ‬䆌乚ඳ⡍ᅮⱘ㚠᱃ֵᙃB៤ЎĀᢳড়āᅮНⱘϔ䚼ߚǄ⹂ߛഄ䆆ˈhা
೼f(xi)ҢBġhġxiЁⓨ㒢⌒⫳ᯊˈᢳড়䆁㒗ḋ՟<xi,f(xi)>Ǆ
x 䗮䖛ᓩܹ㚠᱃ⶹ䆚 Bˈ䆹݀ᓣ㽕∖ᄺдㅫ⊩Փ⫼䖭ϔ㚠᱃ֵᙃᴹᓩᇐ h ⱘ᧰㋶ˈ
㗠ϡᰃা᧰㋶䇁⊩ড়⊩ⱘ‫؛‬䆒ぎ䯈Ǆϟ䴶ゴ㡖Ёᦣ䗄ⱘ䗚ᔦ㒧䖛⿟ህҹ䖭⾡ᔶ
ᓣՓ⫼њ㚠᱃ⶹ䆚Ǆ
ৠᯊˈᣝ✻䖭⾡݀ᓣⱘᔦ㒇䘏䕥㓪⿟䘛ࠄњ޴⾡ᅲ䏉Ϟⱘೄ䲒Ǆ
x ᇍ (¢ xi , f ( xi )² D)( B h xi ) ƅfxiⱘ㽕∖ᅲ䋼Ϟϡ㛑໘⧚᳝ాໄ᭄᥂Ǆ

䯂乬೼Ѣˈ䆹㸼䖒ᓣϡ‫ܕ‬䆌೼㾖ᆳࠄᅲ՟xi੠݊Ⳃᷛؐf(xi)Ёߎ⦄Ꮒ䫭ⱘৃ㛑ᗻǄ
䖭ḋⱘᏂ䫭ৃ㛑ѻ⫳ᇍhⱘϡϔ㟈㑺ᴳǄϡᑌⱘᰃˈ໮᭄ᔶᓣ䘏䕥Ḛᶊᅠܼ≵᳝
㛑࡯೼㒭ᅮϡϔ㟈ᮁ㿔ᯊऎߚߎⳳ੠‫؛‬ᴹǄ
x ϔ䰊䘏䕥䇁㿔ⱘ㸼ᕕ࡯໾ᔎˈ㗠Ϩ⒵䎇 (¢ xi , f ( xi )² D)( B h xi ) ƅfxi
ⱘ‫؛‬䆒᭄䞣໾໮ˈҹ㟇Ѣ‫؛‬䆒ぎ䯈ⱘ᧰㋶೼ϔ㠀ᚙᔶϟᰃ䲒ҹᠻ㸠ⱘǄ䆌໮䖥
ᳳⱘᎹ԰Ꮖᇏ∖ফ䰤ᔶᓣⱘϔ䰊㸼䖒ᓣ៪݊ҪѠ䰊ⶹ䆚ˈҹᬍ䖯‫؛‬䆒ぎ䯈᧰㋶
ⱘᯧ໘⧚ᗻǄ
x ሑㅵⳈ㾝Ϟ㚠᱃ⶹ䆚ৃ᳝ࡽѢ䰤ࠊ‫؛‬䆒ⱘ᧰㋶ˈ೼໮᭄ ILP ㋏㒳Ё˄ࣙᣀ᠔᳝
ᴀゴ䅼䆎ⱘ˅ˈ‫؛‬䆒ぎ䯈᧰㋶ⱘ໡ᴖᑺӮ䱣ⴔ㚠᱃ⶹ䆚ⱘ๲ࡴ㗠๲催Ǆ˄✊㗠ˈ
ৃ㾕㄀ 11 ੠ 12 ゴЁϔѯㅫ⊩Փ⫼㚠᱃ⶹ䆚ᴹ‫ޣ‬ᇣ㗠ϡᰃ๲ࡴḋᴀ໡ᴖᑺǄ˅
೼ϟϔ㡖ˈ៥Ӏ㗗ᶹњϔϾᕜ᱂䘡ⱘ䗚⎉㭈ㅫᄤˈᅗ䗮䖛ড䕀ⱘⓨ㒢᥼⧚㾘߭ᴹᵘ䗴‫؛‬
䆒Ǆ
10.7 䘼ᖈ㔉
㞾ࡼⓨ㒢ⱘϔ㠀ᮍ⊩ᰃ⫼ Robinson˄˅ᦤߎⱘᔦ㒧㾘߭˄resolutionrule˅Ǆᔦ㒧㾘
߭ᰃϔ䰊䘏䕥ЁϔϾড়⧚Ϩᅠ໛ⱘⓨ㒢᥼⧚㾘߭Ǆ಴ℸˈৃҹᛇࠄ䖭ḋⱘ䯂乬˖ᰃ৺ৃҹ䗮
䖛ড䕀ᔦ㒧㾘߭ᴹᔶ៤䗚⎉㭈ㅫᄤǄಲㄨᰃ㚃ᅮⱘˈ㗠Ϩℷᰃ䖭Ͼㅫᄤᔶ៤њ Cigol ⿟ᑣⱘ
෎⸔Ǆ˄Muggleton Buntine˅Ǆ
ҟ㒡ᔦ㒧㾘᳔߭ᆍᯧⱘᮍ⊩ᰃҹੑ乬㸼⼎ⱘᔶᓣˈᅗৃҹ㹿ᠽሩࠄϔ䰊㸼⼎ЁǄҸ L
ЎӏᛣϔϾੑ乬᭛ᄫˈᑊҸ P ੠ R Ўӏᛣੑ乬ᄤহǄᔦ㒧㾘߭Ў˖
P Ģ L
L Ģ R
P Ģ R
ᅗৃ⧚㾷Ў˖㒭ᅮ㒓ϞⱘϸϾᄤহˈᕫࠄ㒓ϟⱘᄤহǄⳈ㾝Ϟᔦ㒧㾘߭ᰃ⧚᠔ᔧ✊ⱘǄ
㒭ᅮϸϾᮁ㿔 PĢL ੠LĢRˈᰒ✊ L ៪L Ёᖙ᳝ϔϾЎ‫؛‬Ǆ಴ℸˈP ៪ R Ёᖙ᳝ϔϾЎ
ⳳǄ಴ℸ㒧䆎 PĢR 㚃ᅮᰃ⒵䎇ⱘǄ
ੑ乬ᔦ㒧ㅫᄤⱘϔ㠀ᔶᓣ೼㸼 Ёᦣ䗄Ǆ㒭ᅮϸϾᄤহC੠Cˈᔦ㒧ㅫᄤ佪‫⹂ܜ‬ᅮ

᭛ᄫLᰃ৺ҹℷ᭛ᄫᔶᓣߎ⦄೼ϔϾᄤহЁˈᑊҹ䋳᭛ᄫᔶᓣߎ⦄೼঺ϔᄤহЁǄ✊ৢᕫࠄ
བϞ݀ᓣЁⱘ㒧䆎Ǆ՟བˈ೒ Ꮊջⱘᔦ㒧ㅫᄤǄ㒭ᅮᄤহC੠Cˈ㄀ϔℹ⹂ᅮ᭛ᄫ
L KnowMaterialˈᅗ೼CЁߎ⦄ˈ㗠ᅗⱘ䋳᭛ᄫKnowMaterial KnowMaterial೼CЁ
ߎ⦄Ǆ᠔ҹ㒧䆎ᰃϔᄤহˈ݊ᔶᓣЎ᭛ᄫC^L` PassExam੠C^L` Studyⱘ㘨ড়ǄВ঺
ϔϾ՟ᄤˈᑨ⫼ᔦ㒧㾘߭ࠄᄤহC AĢBĢCĢD੠C BĢEĢFᕫࠄ㒧ᵰЎᄤহAĢCĢ
DĢEĢFǄ
㸼 10-5 ᔦ㒧ㅫᄤ˄ੑ乬ᔶᓣ˅
Ǆ
㒭ᅮᄤহC1੠C2Ǆᔦ㒧ㅫᄤᵘ䗴ߎϔᄤহCՓC1ġC2ƅCǄ
1. 㒭ᅮ߱ྟᄤহC1੠C2ˈҢᄤহC1ЁᇏᡒϔϾ᭛ᄫLˈᑊϨLߎ⦄೼C2ЁǄ
2. 䗮䖛ড়ᑊC1੠C2Ёⱘ䰸њL੠L໪ⱘ᠔᳝᭛ᄫˈᔶ៤ᔦ㒧ᓣCǄ᳈㊒⹂ഄˈߎ⦄೼㒧ᵰCЁⱘ᭛ᄫ䲚ড়
Ў˖
C (C1 {L}) (C 2 {L})

݊ЁĤ㸼⼎䲚ড়ᑊˈĀˉā㸼⼎䲚ড়ᏂǄ
೒ 10-2 ᔦ㒧੠䗚ᔦ㒧ⱘ՟ᄤ
Ꮊ䖍Ўᑨ⫼ᔦ㒧㾘߭˄ⓨ㒢ⱘ˅Ң㒭ᅮᄤহC1੠C2Ё᥼⧚ߎᄤহCǄে䖍Ў݊䗚䖛⿟ⱘᑨ⫼˄ᔦ
㒇ⱘ˅ˈҢC੠C1Ё᥼䆎ߎC2Ǆ
ᕜᆍᯧ⫼ᔦ㒧ㅫᄤⱘ䗚䕀ᴹᔶ៤ϔᠻ㸠ᔦ㒇᥼⧚ⱘ䗚⎉㭈ㅫᄤOCCǄϔ㠀ᴹ䇈ˈ

䗚⎉㭈ㅫᄤᖙ乏೼㒭ᅮᔦ㒧ᓣC੠ϔ߱ྟᄤহCᯊ᥼ᇐߎ঺ϔ߱ྟᄤহCǄ㗗㰥ϔϾ՟ᄤˈ
㒭ᅮᔦ㒧ᓣC AĢBϨ߱ྟᄤহC BĢDǄབԩ᥼ᇐߎᄤহCҹՓCġCƅC˛佪‫⊼ˈܜ‬ᛣ⬅
ᔦ㒧ㅫᄤⱘᅮНˈӏᛣߎ⦄೼CЁԚϡ೼CЁⱘ᭛ᄫᖙ乏Ꮖ೼CЁߎ⦄Ǆ೼䖭Ͼ՟ᄤЁˈᅗ
㸼⼎Cᖙ乏ࣙ৿᭛ᄫAǄ݊⃵ˈ೼CЁߎ⦄Ԛϡ೼CЁⱘ᭛ᄫᖙЎᔦ㒧㾘߭⿏এњⱘ᭛ᄫˈ
಴ℸᅗⱘ䋳᭛ᄫᖙ乏೼C ЁǄ೼ℸ՟Ёˈᅗ㸼⼎Cᖙ乏ࣙ৿᭛ᄫDǄ಴ℸC AĢDǄ䇏
㗙ৃҹᕜᆍᯧഄ偠䆕ˈᑨ⫼ᔦ㒧㾘߭ࠄC੠C⹂ᅲѻ⫳њ᠔Ꮰᳯⱘᔦ㒧ᓣCǄ
⊼ᛣ೼Ϟ՟ЁC᳝঺ϔ⾡ৃ㛑ⱘ㾷Ǆ⹂ߛഄ䆆ˈCৃҹᰃ᳈⡍⅞ⱘᄤহAĢDĢBǄℸ
㾷Ϣ㄀ϔϾ㾷ⱘϡৠ೼ѢCЁࣙ৿њϔϾCЁߎ⦄ⱘ᭛ᄫǄҢЁৃᕫࠄⱘϔ㠀䆎⚍೼Ѣˈ䗚
ᔦ㒧ᰃϡ⹂ᅮⱘˈेৃ㛑᳝໮ϾᄤহCՓC੠Cѻ⫳ᔦ㒧ᓣCǄ೼݊Ё䖯㸠䗝ᢽⱘϔϾਃথ
ᓣᮍ⊩Ў‫أ‬ད᳈ⷁⱘᄤহˈ៪ㄝӋഄˈ‫؛‬ᅮCϢC≵᳝݅ৠⱘ᭛ᄫǄབᵰᓩܹ䖭⾡ᇍⷁᄤহ
ⱘ‫أ‬དˈᇍ䗚ᔦ㒧䖛⿟ⱘϔ㠀ᦣ䗄㾕㸼 Ǆ
㸼 10-6 䗚ᔦ㒧ㅫᄤ˄ੑ乬ᔶᓣ˅
Ǆ
㒭ᅮϸᄤহC੠C1ˈᅗ䅵ㅫߎC2ՓCġCƅCǄ
1. 㒭ᅮ߱ྟᄤহC1੠CˈᇏᡒϔϾ᭛ᄫLˈᅗߎ⦄೼ᄤহC1ЁԚϡߎ⦄೼CЁǄ
2. 䗮䖛ࣙ৿ϟ߫ⱘ᭛ᄫˈᔶ៤㄀ѠϾᄤহC2˖
C2 (C (C1 {L})) {L}
៥Ӏৃҹ෎Ѣབ䗚ᔦ㒧䖭ḋⱘ䗚⎉㭈ㅫᄤᓔথߎ㾘߭ᄺдㅫ⊩ᴹǄ⹂ߛഄ䆆ˈᄺдㅫ⊩
ৃՓ⫼䗚⎉㭈ᴹᵘ䗴ߎ‫؛‬䆒ˈℸ‫؛‬䆒Ϣ㚠᱃ⶹ䆚ϔ䍋⎉㭈䆁㒗᭄᥂Ǆϔ⾡ㄪ⬹ᰃՓ⫼ᑣ߫㽚
Ⲫㅫ⊩ˈᕾ⦃ഄҹ䖭⾡ᮍ⊩ᄺдHornᄤহ䲚Ǆ೼↣⃵ᕾ⦃Ёˈㅫ⊩䗝ᢽ≵᳝㹿ҹࠡᄺдࠄ
ⱘᄤহ㽚ⲪⱘϔϾ䆁㒗ḋ՟xifxi!Ǆ✊ৢᑨ⫼ᔦ㒧㾘߭ᴹ⫳៤⒵䎇 ( B h xi ) ƅfxi
ⱘ‫׭‬䗝‫؛‬䆒hiˈ݊ЁBЎ㚠᱃ⶹ䆚ࡴϞҹࠡᕾ⦃ЁᄺࠄⱘӏᛣᄤহǄ⊼ᛣ䖭ᰃϔϾḋ՟偅ࡼ
ⱘ᧰㋶ˈ಴Ў↣Ͼ‫׭‬䗝‫؛‬䆒ⱘᓎゟᰃЎњ㽚Ⲫϔ⡍ᅮḋ՟Ǆᔧ✊བᵰᄬ೼໮Ͼ‫׭‬䗝‫؛‬䆒ˈ䙷
М೼݊Ё䗝ᢽⱘㄪ⬹ᰃ䗝প೼݊Ҫḋ՟Ϟг᳔᳝催㊒ᑺⱘ‫؛‬䆒ǄCigol⿟ᑣՓ⫼њ㒧ড়䖭⾡
ᑣ߫㽚Ⲫㅫ⊩ⱘ䗚ᔦ㒧ˈҹℸϢ⫼᠋䖯㸠ѸѦҹ㦋ᕫ䆁㒗ḋ՟ᑊᓩᇐ݊೼ৃ㛑ⱘᔦ㒇᥼⧚ℹ
ⱘᎼ໻ぎ䯈Ёⱘ᧰㋶Ǆ✊㗠CigolՓ⫼њϔ䰊㸼⼎㗠ϡᰃੑ乬㸼⼎Ǆϟ䴶៥Ӏᦣ䗄Ў໘⧚ϔ
䰊㸼⼎᠔䳔ᇍᔦ㒇㾘߭ⱘᠽሩǄ
10.7.1 ж䱬ᖈ㔉
ᔦ㒧㾘߭ৃҹᕜᆍᯧഄᠽሩࠄϔ䰊㸼⼎ЁǄབੑ乬䘏䕥Ёϔḋˈᅗ䳔㽕䕧ܹϸϾᄤহˈ
䕧ߎ㄀ϝϾᄤহǄᅗϢੑ乬ᔦ㒧ⱘ݇䬂ϡৠ೼Ѣˈ䖭ϔ䖛⿟བҞ㽕෎Ѣড়ϔ˄unifying˅㕂
ᤶ᪡԰Ǆ
ᅮН㕂ᤶsubstitutionЎব䞣ࠄ乍ⱘӏᛣ᯴ᇘǄ՟བˈ㕂ᤶ© ^xBobyz`㸼⼎ব䞣 x

᳓ᤶЎ乍 Bobˈ㗠ব䞣 y ᳓ᤶЎ乍 zǄՓ⫼ヺো W©ҷ㸼ᑨ⫼ࠄϔ㕂ᤶ©ࠄᶤ㸼䖒ᓣ W ⱘ㒧
ᵰǄ՟བˈ㢹 L ᰃ᭛ᄫ FatherxBillˈϨ©ЎϞ䗄ⱘ㕂ᤶˈ߭ L© FatherBobBillǄ
བᵰL© L©ˈ߭⿄©Ўϸ᭛ᄫL੠Lⱘড়ϔ㕂ᤶ˄unifyingsubstitution˅Ǆ՟བˈ㢹
L FatherxyˈL FatherBillzˈϨ© ^xBillzy`ˈ䙷М©ᰃL੠Lⱘড়ϔ㕂ᤶˈ಴
ЎL© L© FatherBillyǄড়ϔ㕂ᤶⱘᛣНᰃ˖೼ᔦ㒧ⱘੑ乬ᔶᓣЁˈϸᄤহC੠Cⱘ
ᔦ㒧ᓣⱘ㦋ᕫᰃ䗮䖛⹂ᅮϔ೼CЁⱘᄤহLᑊϨL೼CЁǄ೼ϔ䰊ᔦ㒧Ёˈᅗ᥼ᑓЎҢᄤহ
CЁᇏᡒϔ᭛ᄫL੠೼CЁᇏᡒ᭛ᄫLˈՓᕫৃᡒࠄᇍѢL੠Lⱘᶤড়ϔ㕂ᤶ©˄ेˈՓL
© L©˅Ǆᔦ㒧㾘߭✊ৢᣝϟ䴶ⱘㄝᓣᓎゟᔦ㒧ᓣC˖
C (C1 {L1 })T (C 2 {L2 })T

ᔦ 㒧㾘 ߭ ⱘ ϔ 㠀 ᦣ 䗄㾕㸼 Ǆ Ў 䇈 ᯢ ᅗ ˈ ‫ ؛‬ᅮ C Whitex ĕ Swanx ঞ
C SwanFred Ǆ Ў ᑨ ⫼ ᔦ 㒧㾘 ߭ 佪 ‫ ܜ‬ᇚ C ㄝ Ӌ ഄ 㸼⼎ Ў ᄤ হ ⱘ ᔶ ᓣ C Whitex Ģ
SwanxǄ✊ৢৃᑨ⫼ᔦ㒧㾘߭Ǆ㄀ϔℹˈ‫ܜ‬ᡒࠄCЁⱘ᭛ᄫL Swanx੠CЁⱘ᭛ᄫ
L SwanFredǄབᵰ䗝ᢽড়ϔ㕂ᤶ© ^xFred`ˈ߭ϸϾᄤহ⒵䎇L© L© SwanFredǄ
಴ℸˈ㒧䆎CЎC^L`© WhiteFred੠C^L`© ˈेC WhiteFredǄ
㸼 10-7 ᔦ㒧㾘߭˄ϔ䰊ᔶᓣ˅
1. ᇏᡒC1Ёⱘ᭛ᄫL1ˈC2Ёⱘ᭛ᄫL2ˈҹঞ㕂ᤶ©ˈՓᕫL© L©Ǆ
2. 䗮䖛ࣙ৿C1©੠C©Ё䰸њL©੠L©ҹ໪ⱘ᭛ᄫˈᔶ៤ᔦ㒧ᓣCǄ᳈㊒⹂ഄ䆆ˈߎ⦄೼㒧䆎CЁⱘ
᭛ᄫ䲚ড়Ў˖
C (C1 {L1 })T (C 2 {L2 })T
10.7.2 䘼ᖈ㔉φж䱬᛻߫
៥Ӏৃҹ⫼ߚᵤ⊩᥼ᇐߎ䗚ᔦ㒧ㅫᄤˈᮍ⊩ᰃ䗮䖛ᇍᅮНᔦ㒧㾘߭ⱘᓣ 䖯㸠ҷ᭄

᪡԰Ǆ佪‫⊼ˈܜ‬ᛣᓣ Ёⱘড়ϔ㕂ᤶ©ৃ㹿ᚳϔഄߚ㾷Ў©੠©ˈ݊Ё© ©©ˈ©ࣙ
৿⍝ঞᄤহCЁব䞣ⱘ᠔᳝㕂ᤶˈ㗠©ࣙ৿⍝ঞCЁব䞣ⱘ᠔᳝㕂ᤶǄ䆹ߚ㾷ⱘড়⧚ᗻ೼
ѢC੠CᘏᰃᓔྟѢϡৠⱘব䞣ৡ˄಴ЎᅗӀᰃϡৠⱘܼ⿄䞣࣪䰜䗄˅ǄՓ⫼©ⱘ䖭⾡ߚ㾷ˈ
ৃᇚᓣ 䞡ᮄ㸼䖒Ў˖
C (C1 {L1 })T 1 (C 2 {L2 })T 2
䆄ԣ䖭䞠ⱘ‫ޣ‬োĀāҷ㸼䲚ড়ᏂǄ⦄೼བᵰ䰤ࠊ䗚ᔦ㒧ㅫᄤЎ᥼⧚ߎⱘCЁ≵᳝ϢC݅
ৠⱘ᭛ᄫ˄㸼⼎‫أ‬ད᳔ⷁⱘCᄤহ˅
ˈ䙷МৃᇚϞᓣ‫ݭ‬Ў˖
C (C1 {L1 })T 1 (C 2 {L2 })T 2
1
᳔ৢৃՓ⫼ᔦ㒧㾘߭ⱘᅮН L2 L1T 1T 2 ˈ㾷ߎCᴹᕫࠄ
䗚ᔦ㒧˖
1 1
C2 (C (C1 {L1 })T 1 )T 2 {L1T 1T 2 } ˄˅
ᓣ 㒭ߎњϔ䰊䘏䕥ⱘ䗚ᔦ㒧㾘߭Ǆབ೼ੑ乬ᔶᓣЁˈℸ䗚⎉㭈ㅫᄤᰃ䴲⹂ᅮᗻⱘǄ
⹂ߛഄ䆆ˈ೼ᑨ⫼ᅗⱘ䖛⿟Ёˈϔ㠀ৃᡒࠄᕙᔦ㒧ⱘᄤহC੠㕂ᤶ©੠©ⱘ໮⾡䗝ᢽǄ↣
ϔ㒘䗝ᢽ䛑ѻ⫳ϔϾϡৠⱘC㾷Ǆ
೒ ೒⼎њℸ䗚ᔦ㒧㾘߭ᑨ⫼೼ϔㅔऩ՟ᄤϞⱘ໮Ͼℹ偸Ǆ೼೒Ёˈ៥ӀᏠᳯḍ᥂
㒭 ᅮ ⱘ 䆁㒗 ᭄ ᥂ D GrandChild Bob Shannon ੠ 㚠 ᱃ ֵ ᙃ B ^FatherShannon Tom
FatherTomBob`ˈᄺдࠄⳂᷛ䇧䆡GrandChildyxⱘ㾘߭Ǆ㗗㰥೒ Ё䗚ᔦ㒧ᷥⱘ
᳔ϟ䴶ϔℹǄ䖭䞠ˈ៥Ӏ䆒㕂㒧䆎CЎ䆁㒗ḋ՟GrandChildBobShannonˈᑊϨҢ㚠᱃
ֵᙃЁ䗝ᢽᄤহC FatherShannonTomǄЎᑨ⫼䗚ᔦ㒧ㅫᄤˈᇍѢ᭛ᄫLা᳝ϔ⾡䗝ᢽˈ

⿄ЎFatherShannonTomǄ‫؛‬ᅮ៥Ӏ䗝ᢽ䗚㕂ᤶ© ^`Ϩ© ^Shannonx`Ǆ೼ℸᚙ‫މ‬ϟˈ
1 1
ᕫࠄⱘᄤহC Ўᄤহ (C (C1 {L1 })T 1 )T 2 (CT 1 )T 2 GrandChild ( Bob, x) ੠ᄤহ
1
{L1T 1T 2 } Father ( x, Tom) ⱘ 㘨 ড় Ǆ ಴ ℸ 㒧 ᵰ Ў ᄤ হ GrandChild ( Bob, x )
Father(x,Tom)ˈ៪ㄝӋⱘᄤহ GrandChild ( Bob, x) m Father ( x, Tom ) Ǆ⊼ᛣ䖭Ͼϔ㠀㾘
߭ϢCϔ䍋⎉㭈њ䆁㒗ḋ՟GrandChildBobShanonǄ
ҹⳌԐⱘᮍᓣˈ᥼⧚ᕫࠄⱘᄤহৃ԰Ў㄀ѠϾᔦ㒧ℹЁⱘ㒧䆎 Cˈབ೒ ᠔⼎Ǆ೼

䖭ϸℹЁⱘ↣ϔℹЁ䛑ৃ㛑᳝໮Ͼ䕧ߎˈ䖭প‫އ‬Ѣᇍ㕂ᤶⱘ䗝ᢽ˄㾕д乬 ˅Ǆ೼೒
ⱘ ՟ ᄤ Ё ˈ ⡍ ᅮ ⱘ 䗝 ᢽ ѻ ⫳ њ Ⳉ 㾝 Ϟ ৃ ⒵ 䎇 ⱘ ᳔ 㒜 ᄤ হ GrandChild ( y , x) m
Father ( x, z ) Father ( z , y ) Ǆ
೒ 10-3 ϔϾ໮ℹ䗚ᔦ㒧
݊ЁᏺᮍḚⱘᄤহЎ᥼⧚ℹⱘ㒧ᵰǄ೼↣ϔℹˈCᰃԡѢᑩ䚼ⱘᄤহˈC1 ᰃᎺ䖍ⱘᄤহˈC2
ᰃে䖍ᏺᮍḚⱘᄤহǄ೼䖭ϸϾ᥼⧚ℹЁˈ©䛑ᰃぎ㕂ᤶˈ㗠©㕂ᤶᰒ⼎೼C ϟᮍǄ⊼ᛣ᳔㒜
ⱘ㒧䆎˄᳔েϞ㾦ⱘᏺᮍḚᄤহ˅ᰃHornᄤহGrandChildyx mFatherxzFatherzyⱘ঺
ϔ⾡ᔶᓣǄ
10.7.3 䘼ᖈ㔉ቅ㔉
ὖᣀഄ䆆ˈ䗚ᔦ㒧ᦤկњϔ⾡ϔ㠀ⱘ䗨ᕘҹ㞾ࡼѻ⫳⒵䎇㑺ᴳ ( B h xi ) ƅfxiⱘ
‫؛‬䆒hǄ䖭ᰃ䗮䖛䗚䕀ᓣ 㒭ߎⱘᔦ㒧㾘߭ᕫࠄⱘǄҢℸᔦ㒧㾘߭Ё㾷ߎᄤহCˈᓣ
Ёⱘ䗚ᔦ㒧㾘߭ⱘᕜᆍᯧ᥼ᇐߎǄ
㒭ᅮϔ㒘ᓔྟᄤহˈৃ䗮䖛䞡໡ᑨ⫼ℸ䗚ᔦ㒧㾘߭⫳៤໮Ͼ‫؛‬䆒Ǆ⊼ᛣ䗚ᔦ㒧㾘߭‫݋‬᳝
ϔ Ӭ ⚍ ˈ ᅗ া ⫳ ៤ ⒵ 䎇 ( B h xi ) ƅ fxi ⱘ ‫ ؛‬䆒 Ǆ Ⳍ ড ˈ FOIL ⱘ ⫳ ៤ ‫ ⌟ ݡ‬䆩
˄generate-and-test˅᧰㋶೼↣ϔ᧰㋶ℹ⫳៤໮Ͼ‫؛‬䆒ˈࣙᣀϔѯϡ⒵䎇ℸ㑺ᴳⱘǄ✊ৢFOIL
䗮䖛㗗㰥᭄᥂Dᴹ೼䖭ѯ‫؛‬䆒Ё԰ߎ䗝ᢽǄ⬅Ѣ䖭ϔᏂᓖˈ៥Ӏৃᳳᳯ෎Ѣ䗚ᔦ㒧ⱘ᧰㋶᳈
᳝䩜ᇍᗻϨ᳈᳝ᬜǄ✊㗠ᅲ䰙᳾ᖙབℸǄϔϾॳ಴ᰃ䗚ᔦ㒧ㅫᄤ೼ӏᛣϔℹ⫳៤ᅗⱘ‫؛‬䆒ᯊˈ
া㛑㗗㰥ৃ⫼᭄᥂Ёⱘϔᇣ䚼ߚǄ㗠FOIL㗗㰥᠔᳝ⱘৃ⫼᭄᥂ˈ೼݊ᣝ䇁⊩⫳៤ⱘ‫؛‬䆒Ё
䖯㸠䗝ᢽǄՓ⫼䗚⎉㭈੠Փ⫼⫳៤‫⌟ݡ‬䆩ϸ⾡᧰㋶ㄪ⬹ⱘᏂ߿ҡᰃϔϾⷨおЏ乬ǄSrinivasan
ㄝ˄˅ᦤկњᇍ䖭ϸ⾡ᮍ⊩ⱘᅲ偠ᗻ↨䕗Ǆ
10.7.4 ⌑ौȽऻᇯૂ⏫㮪
ࠡϔ㡖ᣛߎњᔦ㒇੠䗚⎉㭈П䯈ⱘ㘨㋏Ǆ⬅Ѣҹࠡⴔ䞡Ѣ䆆䗄೼‫؛‬䆒᧰㋶Ёⱘϔ㠀ࠄ⡍
⅞ᑣˈ䙷М᳝ᖙ㽕ⷨお moregeneralthan ݇㋏੠䗚⎉㭈П䯈ⱘ㘨㋏ǄЎ䇈ᯢℸˈ㗗㰥བϟ
ⱘᅮН˖
x more-general-thanǄ㄀ 2 ゴЁⱘmore-general-than-or-equal-to݇㋏˄ıg˅ᅮНЎ˖
㒭ᅮϸᏗᇨؐߑ᭄hj(x)੠hk(x)ˈ៥Ӏ⿄hjıg hkᔧϨҙᔧ (x)hk ( x) o h j ( x) Ǆℸ
ıg݇㋏㹿⫼Ѣ䆌໮ᄺдㅫ⊩Ёҹᓩᇐ‫؛‬䆒ぎ䯈ⱘ᧰㋶Ǆ
x ©-ࣙᆍ˄©-subsumption˅Ǆ㗗㰥ϸϾᄤহCj੠CkˈᅗӀⱘᔶᓣ䛑ᰃHĢL1Ģ…Lnˈ
݊ЁHЎϔℷ᭛ᄫˈ㗠LiЎӏᛣ᭛ᄫǄ⿄ᄤহCj©-ࣙᆍᄤহCkˈᔧϨҙᔧᄬ೼ϔ
Ͼ㕂ᤶՓ C jT C k˄䖭䞠៥ӀᇚӏᛣᄤহCᦣ䗄Ў݊ᵤপᓣЁ৘᭛ᄫⱘ䲚ড়˅Ǆ
䆹ᅮН㾕Plotkin˄1970˅Ǆ
x ⎉㭈˄entailment˅㗗㰥ϸᄤহCj੠CkǄᄤহCj㹿⿄Ў⎉㭈ᄤহCk˄‫ݭ‬԰CjƅCk˅
ᔧϨҙᔧCkҢCjЁⓨ㒢⌒⫳Ǆ
䖭ϝϾᅮНП䯈᳝ҔМ‫ݙ‬೼㘨㋏˛佪‫ˈܜ‬ᇚıgⱘᅮН䞡ᮄ㸼⼎Ўϔ䰊ᔶᓣˈབ঺ϸϾ
ᅮНϔḋǄབᵰᇍᶤⳂᷛὖᗉcx㗗㰥ϔᏗᇨؐ‫؛‬䆒hxˈ݊Ёhx㸼⼎Ў᭛ᄫⱘড়পˈ䙷
Мৃ䞡ᮄ㸼⼎ℸ‫؛‬䆒Ўᄤহ˖
cxĕhx
䖭䞠៥Ӏ䙉ᕾ䗮ᐌⱘProlog㾷䞞ˈेx㢹ϡ㛑㹿䆕ᯢЎℷ՟ᯊˈ߭x㹿ߚ㉏Ўড՟Ǆ಴ℸˈ
ৃⳟߎࠡ䴶ᅮНⱘıgᑨ⫼ѢHornᄤহⱘࠡӊ˄៪㾘߭ԧ˅ǄHornᄤহ䱤৿ⱘৢӊЎⳂᷛὖᗉ
cxǄ
ıgᅮН੠©ࣙᆍᅮНП䯈ⱘ݇㋏ᰃҔМ˛⊼ᛣབᵰhıghˈ߭ᄤহCcxĕhx
ᰃ©ࣙᆍᄤহCcxĕhxǄ᳈䖯ϔℹˈेՓ೼ᄤহ᳝ϡৠⱘ༈䚼ᯊˈ©-ࣙᆍгৃ៤ゟǄ
՟བˈϟ䴶ⱘᚙᔶЁᄤহA©ࣙᆍᄤহB˖
AMotherxy ĕFatherxzġSpousezy

BMotherxLouiseĕFatherxBobġSpouseBobyġFemalex
಴Ўབᵰ䗝ᢽ© ^yLouisezBob`߭ AT B Ǆ䖭䞠ⱘ݇䬂ऎ߿೼Ѣıg䱤৿‫؛‬ᅮњϸ

Ͼᄤহⱘ༈䚼ᰃⳌৠⱘˈ㗠©-ࣙᆍৃ೼ᄤহ༈䚼ϡৠᯊ៤ゟǄ
᳔ৢˈ©ࣙᆍᰃ⎉㭈ⱘϔ⾡⡍⅞ᔶᓣǄेˈབᵰᄤহ A©ࣙᆍᄤহ Bˈ߭ AƅBǄ✊

㗠ˈ៥Ӏৃᡒࠄ䖭ḋⱘ A ੠ BˈՓ AƅB Ԛ A ᑊϡ©ࣙᆍ BǄ՟བϟ䴶ϸϾᄤহ˖
AElephantfatherBofx ĕElephantx
BElephantfatherBoffatherBofyĕElephanty
݊Ё fatherBofxЎϔߑ᭄ˈҷ㸼 x ⱘ⠊҆Ǆ⊼ᛣ㱑✊ B ৃ⬅ A ᕫࠄ䆕ᯢˈैϡᄬ೼㕂

ᤶ©Փ A©ࣙᆍ BǄ
བ䖭ѯ՟ᄤ᠔⼎ˈࠡ䴶ᇍ moregeneralthan ⱘᅮНᰃ©ࣙᆍⱘϔ⾡⡍⅞ᚙ‫ˈމ‬㗠©

ࣙᆍজᰃ⎉㭈ⱘ⡍⅞ᚙ‫މ‬Ǆ಴ℸˈ䗮䖛⊯࣪੠⡍࣪‫؛‬䆒ᴹ᧰㋶‫؛‬䆒ぎ䯈↨⫼ϔ㠀ⱘ䗚⎉㭈ㅫ
ᄤᴹ᧰㋶᳈Ўሔ䰤Ǆϡᑌⱘᰃˈ䗚⎉㭈䖭⾡᳔ϔ㠀ⱘᔶᓣৃѻ⫳᮴⊩໘⧚ⱘ᧰㋶Ǆ✊ৢЁ䯈
ⱘ©ࣙᆍⱘᅮНᦤկњԡѢ moregeneralthen ੠⎉㭈Ё䯈ⱘϔ⾡ὖᗉǄ
10.7.5 Progol
㱑✊ᇍѢ⫳៤‫׭‬䗝‫؛‬䆒ˈ䗚ᔦ㒧ᰃϔ⾡ᕜ਌ᓩҎⱘᮍ⊩Ǆ೼ᅲ䏉Ёᅗᕜᆍᯧᇐ㟈‫׭‬䗝‫؛‬
䆒ⱘ㒘ড়⟚⚌Ǆ঺ϔ⾡䗨ᕘᰃাՓ⫼䗚⎉㭈ᴹ⫳៤ϔϾ᳔⡍⅞‫؛‬䆒ˈᅗϢ㚠᱃ֵᙃϔ䍋⎉㭈
㾖ᆳⱘ᭄᥂Ǆ✊ৢˈ䖭Ͼ᳔⡍⅞‫؛‬䆒ৃ⫼Ѣ⹂ᅮ‫؛‬䆒ぎ䯈ⱘϔ㠀ࠄ⡍⅞᧰㋶䖍⬠ˈϢ FOIL
ЁՓ⫼ⱘ᧰㋶ϔḋˈԚ᳝ϔ᳈໮ⱘ䰤ࠊ˖া㗗㰥↨ℸ䖍⬠᳈ϔ㠀ⱘ‫؛‬䆒Ǆ䆹ᮍ⊩㹿⫼Ѣ Progol
㋏㒳ˈᅗⱘㅫ⊩ৃὖ䗄བϟ˖
ˊ⫼᠋ᣛᅮՓ⫼ϔϾফ䰤ⱘϔ䰊㸼⼎䇁㿔Ў‫؛‬䆒ぎ䯈ˤǄ䖭ѯ䰤ࠊ⫼Ā῵ᗕໄᯢ˄mode
declaration˅āᴹᦣ䗄ˈᅗ‫ܕ‬䆌⫼᠋ᣛᅮ㽕㗗㰥ⱘ䇧䆡੠ߑ᭄ヺোˈҹঞᅗӀⱘখ㗗㉏ൟ੠Ḑ
ᓣǄ
ˊProgolՓ⫼ᑣ߫㽚Ⲫ⊩ᴹҢˤЁᄺдϔ㒘㽚Ⲫ᭄᥂ⱘ㸼䖒ᓣǄᇍѢ↣Ͼ䖬≵㹿䖭ѯ
ᄺࠄⱘ㸼䖒ᓣ㽚Ⲫⱘḋ՟xifxi!ˈᅗ佪‫ܜ‬ᇏᡒ ˤ Ё᳔⡍⅞ⱘ‫؛‬䆒hi ˈՓ ( B h xi ) ƅ
fxiǄ᳈㊒⹂ഄ䆆ˈᅗ‫ܜ‬ᡒࠄ㛑䗮䖛ᑨ⫼k⃵ᔦ㒧㾘߭⎉㭈fxiⱘ‫؛‬䆒ˈ೼݊Ё䅵ㅫߎ᳔⡍
⅞ⱘ‫؛‬䆒ˈҢ㗠䖥ԐᕫࠄhiǄ
ˊ✊ৢProgol೼䖭Ͼ⬅᳔ϔ㠀‫؛‬䆒੠㄀ˎℹЁᕫࠄⱘ⡍⅞䖍⬠hi᠔⬠ᅮⱘ‫؛‬䆒ぎ䯈Ёᠻ
㸠њϔ㠀ࠄ⡍⅞᧰㋶Ǆ೼ℸ‫؛‬䆒䲚ড়Ёˈᅗᇏᡒ᳔᳝ᇣᦣ䗄䭓ᑺ˄⬅᭛ᄫⱘ᭄䞣ᑺ䞣˅ⱘ‫؛‬

䆒Ǆ䆹䚼ߚⱘ᧰㋶ᰃ⬅‫ڣ‬A 䙷ḋⱘਃথᓣ㾘߭ᓩᇐⱘˈᅗⱘׂ࠾᪡԰ৃ೼≵ׂ᳝࠾ᥝ᳔ⷁ‫؛‬
䆒ⱘ亢䰽ϟ䖯㸠Ǆ
㒚㡖ⱘ Progol ㅫ⊩㾕 Muggleton˄ˈ˅Ǆ
x ᑣ߫㽚Ⲫㅫ⊩ᄺдᵤপⱘ㾘߭䲚ˈᮍ⊩ᰃ‫ܜ‬ᄺдऩϾ㊒⹂ⱘ㾘߭ˈ✊ৢ⿏এ㹿
ℸ㾘߭㽚Ⲫⱘℷ՟ˈ‫ݡ‬೼࠽ԭḋ՟Ϟ䞡໡䖭ϔ䖛⿟ǄᅗᦤկњϔϾᄺд㾘߭䲚
ⱘ᳝ᬜⱘ䋾ှㅫ⊩ˈৃ԰Ў⬅乊৥ϟⱘ‫އ‬ㄪᷥᄺдㅫ⊩˄བ ID3˅ⱘ᳓ҷㅫ⊩Ǆ
‫އ‬ㄪᷥㅫ⊩ৃ㹿ⳟ԰ᑊ㸠㽚ⲪˈϢᑣ߫㽚ⲪⳌᇍᑨǄ
x ೼ᑣ߫㽚Ⲫㅫ⊩ЁˈᏆⷨおњ໮⾡ᮍ⊩ҹᄺдऩϾⱘ㾘߭Ǆ䖭ѯᮍ⊩ⱘϡৠ೼
ѢᅗӀ㗗ᶹ㾘߭ࠡӊぎ䯈ⱘㄪ⬹ϡৠǄϔϾᕜ⌕㸠ⱘǃ೼ CN2 ⿟ᑣЁՓ⫼ⱘᮍ
⊩ᰃᠻ㸠ϔ㠀ࠄ⡍⅞ⱘ᷅⢊᧰㋶ˈ⏤䖯ഄ⫳៤ᑊ⌟䆩᳈⡍⅞ⱘ㾘߭ˈⳈࠄᡒࠄ
ϔϾ䎇໳㊒⹂ⱘ㾘߭Ǆ݊Ҫⱘᮍ⊩Ң⡍⅞ࠄϔ㠀䖯㸠‫؛‬䆒᧰㋶ˈՓ⫼ḋ՟偅ࡼ
㗠ϡᰃ⫳៤ᑊ⌟䆩ˈᑊϨᑨ⫼њϡৠⱘ㒳䅵䞣ᑺⱘ㾘߭㊒ᑺᴹᣛᓩ᧰㋶Ǆ
x ϔ䰊㾘߭䲚˄ेࣙ৿ব䞣ⱘ㾘߭˅ᦤկњϔ⾡㸼ᕕ㛑࡯ᕜᔎⱘ㸼⼎Ǆ՟བˈ㓪
⿟䇁㿔 Prolog Փ⫼ϔ䰊 Horn ᄤহᑣ߫ᴹ㸼⼎ϔ㠀ⱘ⿟ᑣǄ಴ℸˈᄺдϔ䰊 Horn
ᄤহⱘ䯂乬гᐌ㹿⿄Ўᔦ㒇䘏䕥㓪⿟ⱘ䯂乬Ǆ
x ᄺдϔ䰊㾘߭䲚ⱘᮍ⊩ᰃᇚ CN2 Ёⱘᑣ߫㽚Ⲫㅫ⊩⬅ੑ乬ᔶᓣᠽሩࠄϔ䰊㸼
⼎Ǆ䆹ᮍ⊩೼ FOIL ⿟ᑣЁ՟⼎ˈᅗৃᄺдࣙᣀㅔऩ䗦ᔦ㾘߭䲚೼‫ⱘݙ‬ϔ䰊㾘߭
䲚Ǆ
x ᄺдϔ䰊㾘߭ⱘ঺ϔᮍ⊩෎ѢϔϾথ⦄˖ेᔦ㒇ᰃⓨ㒢ⱘ䗚䕀Ǆᤶ㿔Пˈᔦ㒇
ⱘ䯂乬ᰃᇏᡒϔϾ‫؛‬䆒 h ⒵䎇ϟ䴶ⱘ㑺ᴳǄ
(¢ xi , f ( xi )² D)( B h xi ) ƅfxi
݊ЁBᰃϔ㠀㚠᱃ֵᙃˈx1…xnᰃ䆁㒗᭄᥂DЁᅲ՟ⱘᦣ䗄ˈ㗠f(x1)…f(xn)Ў䆁㒗
ᅲ՟ⱘⳂᷛؐǄ
x ϔѯ⿟ᑣ䙉ᕾњᔦ㒇ᰃⓨ㒢ⱘ䗚䕀ⱘ㾖⚍ˈ䗮䖛䖤⫼❳ⶹⱘⓨ㒢᥼⧚ⱘ䗚᪡԰
ᴹ᧰㋶‫؛‬䆒Ǆ՟བ Cigol Փ⫼ⱘ䗚ᔦ㒧ᰃᔦ㒧ㅫᄤⱘ䗚䕀ˈ㗠ᔦ㒧ᰃ᱂䘡⫼Ѣᴎ
఼ᅮ⧚䆕ᯢⱘϔ⾡᥼⧚㾘߭ǄProgol 㒧ড়њ䗚⎉㭈ㄪ⬹੠ϔ㠀ࠄ⡍⅞ㄪ⬹ᴹ᧰
㋶‫؛‬䆒ぎ䯈Ǆ
ᄺд݇㋏ᦣ䗄ⱘᮽᳳᎹ԰ࣙᣀ Winston˄˅ⱘ㨫ৡⱘ⿟ᑣˈᅗᄺдབĀarchā䖭ḋ
ⱘὖᗉⱘ㔥㒰ᓣᦣ䗄ǄBanerjiⱘᎹ԰੠ Michalski ⱘ AQ ㅫ⊩㋏߫Ꮉ԰˄བ
MichalskiMichalskietal˅ᰃ᳔ᮽᇚ䘏䕥㸼⼎⫼Ѣᄺд䯂乬ⱘⷨおПϔǄPlotkin
˄˅ⱘ©ࣙᆍᅮН䕗ᮽഄᇍᔦ㒇੠ⓨ㒢П䯈ⱘ݇㋏䖯㸠њᔶᓣ࣪ǄVere˄˅гⷨ
おњᄺдⱘ䘏䕥㸼⼎䯂乬ˈϨ Buchanan˄˅ⱘ METADENDRAL ⿟ᑣৃᄺдࠄ݇㋏ᦣ
䗄ҹ㸼⼎ߚᄤ㒧ᵘЁৃ೼䋼䈅ҾЁ㹿ߚࡆⱘ䚼ߚǄ䆹⿟ᑣ៤ࡳഄথ⦄њϔѯ᳝⫼ⱘ㾘߭ˈᅗ
Ӏ೼࣪ᄺᄺᴃ乚ඳ㹿݀ᏗǄMitchell˄˅ⱘ‫׭‬䗝⍜䰸বൟぎ䯈ㅫ⊩㹿ᑨ⫼Ѣৠḋⱘ࣪ᄺ
㒧ᵘⱘ݇㋏ᦣ䗄Ǆ
䱣ⴔ ᑈҷᎹЁ Prolog 䇁㿔ⱘ᱂䘡࣪ˈⷨおҎਬᓔྟ⏅ܹⷨお Horn ᄤহ㸼⼎ⱘ݇㋏

ᦣ䗄Ǆ䕗ᮽⱘᄺд Horn ᄤহⱘᎹ԰ࣙᣀ Shapiroⱘ MIS ੠ Sammut Banerji˄˅
ⱘ MarvinǄ䖭䞠䅼䆎ⱘ Quinlan˄˅ⱘ FOIL ㅫ⊩ߎ⦄ৢˈᕜᖿ䱣Пѻ⫳њ໮Ͼᑨ⫼ϔ
䰊㾘߭ⱘϔ㠀ࠄ⡍⅞᧰㋶ⱘㅫ⊩ˈࣙᣀ MFOIL˄Dzeroski˅ ǃFOCLPazzanietalǃ
ǄFOCL ㅫ⊩೼
CLAUDIENDeRaedt Bruynooghe੠ MARKUS˄Grobelnik˅
㄀ゴЁᦣ䗄Ǆ
ᄺд Horn ᄤহⱘ঺ϔᴵⷨお䏃㒓ᰃ䗮䖛䗚⎉㭈ˈᰃ⬅ Muggleton Buntine˄˅ᦤ

ߎˈᅗⱘ෎⸔ᰃ Sammut Banerji˄˅੠ Muggleton˄˅Ё㉏Ԑⱘᛇ⊩Ǆℸ䏃㒓Ϟ
᳔䖥ⱘᎹ԰ⴔ䞡Ѣⷨおϡৠⱘ᧰㋶ㄪ⬹੠䰤ࠊ‫؛‬䆒ぎ䯈ҹՓᄺд䖛⿟᳈ᯧѢ໘⧚ⱘᮍ⊩Ǆ՟
བ Kietz Wrobel˄˅Փ⫼೼݊ RDT ⿟ᑣЁ㾘߭῵ᓣᴹ䰤ࠊᄺд䖛⿟Ёৃ㗗㰥ⱘ㸼䖒
ᓣⱘᔶᓣǄMuggleton Feng˄˅䅼䆎њᇚϔ䰊㸼⼎䰤ࠊЎ ijdeterminate ᭛ᄫǄCohen
˄˅䅼䆎њ GRENDEL ⿟ᑣˈᅗ᥹ফϔϾᰒᓣⱘ䇁㿔ᦣ䗄䕧ܹˈҹᦣ䗄ᄤহԧˈҢ㗠
‫ܕ‬䆌⫼᠋ᰒᓣഄ㑺ᴳ‫؛‬䆒ぎ䯈Ǆ
Lavrac Dzeroski˄˅ᦤկњᔦ㒇䘏䕥㓪⿟ⱘϔϾৃ䇏ᗻᕜᔎⱘᬭᴤǄ䖥ᳳ݊Ҫ᳝
⫼ⱘϧ乬᡹㗗੠᭛䲚ࣙᣀ˄Bergadano Gunetti˗Moriketal˗Muggleton
b˅ǄWrobel˄˅ⱘ㓐䗄ゴгᦤկњ䆹乚ඳⱘϔϾདᴤ᭭ǄBratko Muggleton
ὖ䗄њ ILP ೼ϔѯ䞡㽕䯂乬Ϟⱘ䖥ᳳᑨ⫼Ǆϔ㋏߫ⱘ ILP ᮍ䴶ⱘᑈᑺϧ乬䅼䆎Ӯгᦤկњ䖥
ᳳⷨお䆎᭛ⱘᕜདᴹ⑤˄՟བ DeRaedt˅Ǆ
д乬
㗗㰥ϔϾབCN 䙷ḋⱘᑣ߫㽚Ⲫㅫ⊩੠ϔϾབID 䙷ḋⱘᑊ㸠㽚Ⲫㅫ⊩ǄϸϾㅫ

⊩䛑㹿⫼ѢᄺдϔⳂᷛὖᗉˈᅗᅮН೼⬅nϾᏗᇨሲᗻড়প㸼⼎ⱘᅲ՟ϞǄབᵰID ᄺдࠄ
⏅ᑺЎdⱘᑇ㸵‫އ‬ㄪᷥˈᅗᇚࣙ৿ d Ͼϡৠⱘ‫އ‬ㄪ㒧⚍ˈ㗠Ϩ೼ᓎゟ݊䕧ߎ‫؛‬䆒ᯊ԰ߎ
d ⃵ϡৠ䗝ᢽǄབᵰ䆹ᷥ㹿䞡ᮄ㸼⼎Ўϔᵤপ㾘߭䲚ˈৃᔶ៤໮ᇥ㾘߭˛↣Ͼ㾘߭ᢹ᳝
໮ᇥࠡӊ˛ϔϾᑣ߫㽚Ⲫㅫ⊩Ўᄺдࠄৠḋⱘ㾘߭䲚䳔԰ߎ໮ᇥ⃵ϡৠⱘ䗝ᢽ˛བᵰ㒭ᅮ
Ⳍৠⱘ䆁㒗᭄᥂ˈાϔϾ㋏㒳Դ䅸Ў᳈ᆍᯧߎ⦄䖛ᑺᢳড়˛
ᬍ䖯㸼 ⱘ learnonerule ㅫ⊩ˈՓᅗ㛑ᄺдࠡӊЁࣙ৿ᅲ᭄ሲᗻ䯜ؐⱘ㾘߭

˄བ temprature!˅Ǆᣛߎᮄⱘㅫ⊩ৃҢ㸼 Ё԰ાѯׂᬍᕫࠄǄᦤ⼎˖㗗㰥೼‫އ‬ㄪᷥ
Ё䖭ᰃᗢḋᅠ៤ⱘǄ
ᬍ䖯㸼 ⱘ learnonerule ㅫ⊩ˈՓᅗ㛑ᄺдⱘ㾘߭ⱘࠡӊЁৃࣙ৿㉏ԐѢ

nationalityę^CanadianBrazilian`ⱘ㑺ᴳˈे⾏ᬷؐሲᗻৃপᶤᣛᅮ䲚ড়ЁӏᛣؐǄׂᬍৢ
ⱘ⿟ᑣᑨ᥶㋶ࣙ৿᠔᳝䖭ḋᄤ䲚ⱘ‫؛‬䆒ぎ䯈Ǆᣛߎᮄⱘㅫ⊩ৃҢ㸼 Ё԰ાѯׂᬍᕫࠄǄ
㗗㰥ᅲ⦄ Learnonerule ᧰㋶‫؛‬䆒ぎ䯈ᯊৃ䗝ⱘㄪ⬹ˈ⹂ߛഄ䆆ˈ㗗㰥ϟ߫᧰㋶

䖛⿟ሲᗻ˖
˄a˅⫳៤ᑊ⌟䆩 vs᭄᥂偅ࡼ
˄b˅ϔ㠀ࠄ⡍⅞ vs⡍⅞ࠄϔ㠀
˄c˅ᑣ߫㽚Ⲫ vsᑊ㸠㽚Ⲫ
䅼䆎㸼 ੠ Ёㅫ⊩Ё᠔‫خ‬䗝ᢽⱘད໘ǄᇍѢ᧰㋶ㄪ⬹Ёⱘ䖭ϝ⾡ሲᗻˈ䅼䆎䗝

ᢽ঺ϔᮍḜᯊⱘᕅડ˄ℷ䴶ⱘ੠䋳䴶ⱘ˅Ǆ
ᑨ⫼ੑ乬ᔶᓣⱘ䗚ᔦ㒧ࠄᄤহC AĢBˈC AĢBĢGǄ㒭ߎCⱘ㟇ᇥϸ⾡ৃ㛑㒧ᵰǄ

ᑨ⫼䗚ᔦ㒧ࠄᄤহC RBxĢPxA੠C SByĢRzxǄ㒭ߎCⱘ㟇ᇥ

ಯ⾡ৃ㛑㒧ᵰǄ䖭䞠A੠BЎᐌ䞣ˈx੠yЎব䞣Ǆ
㗗㰥೒ Ё᳔ϟ䴶ⱘ䗚ᔦ㒧ℹǄ㢹㒭ᅮ㕂ᤶ©੠©ⱘϡৠ䗝ᢽˈ᥼ᇐߎ㟇ᇥ

ϸ⾡ৃ㛑ѻ⫳ⱘϡৠ䕧ߎǄབᵰ⫼ᄤহFatherTomBob᳓ᤶњFatherShannonTomˈ᥼
ᇐߎℸ䗚ᔦ㒧ℹⱘϔϾ㒧ᵰǄ
㗗㰥ᴀゴЁᔦ㒇䯂乬ⱘᅮН˖
(¢ xi , f ( xi )² D)( B h xi ) ƅfxi
੠ࠡ䴶㄀ゴᇍᔦ㒇‫أ‬㕂ⱘᅮН˄ᓣ ˅П䯈ⱘ㘨㋏Ǆ݊Ёᔦ㒇‫أ‬㕂BbiasᅮНЎ㸼䖒

ᓣ
(xi X )( Bbias D xi ) ƅLxiD

݊ЁLxiDᰃᄺд఼೼Ң䆁㒗᭄᥂DϞᄺдৢ䌟ќᮄᅲ՟xiⱘߚ㉏ˈ㗠XЎᭈϾᅲ՟ぎ
䯈Ǆ⊼ᛣ㄀ϔϾ㸼䖒ᓣᰃЎњᦣ䗄៥ӀᏠᳯᄺд఼䕧ߎⱘ‫؛‬䆒ˈ㗠㄀ѠϾ㸼䖒ᓣᰃЎњᦣ䗄
ᄺд఼Ң䆁㒗᭄᥂Ё⊯࣪ⱘㄪ⬹Ǆ䆒䅵ϔᄺд఼ˈ݊ᔦ㒇‫أ‬㕂BbiasㄝѢ᠔ᦤկⱘ㚠᱃ⶹ䆚BǄ
ㅢ11ㄖ ࠼᷆ᆜҖ
⾎㓿㖁㔌઼ߣㆆṁ䘉ṧⲴᆖҐᯩ⌅䴰㾱аᇊᮠⴞⲴ䇝㓳ṧֻˈԕ䗮ࡠаᇊ㓗࡛Ⲵ⌋ॆ㋮
ᓖǄࡽ䶒ㄐ㢲䇘䇪Ⲵ⨶䇪⭼䲀઼ᇎ傼㔃᷌৽᱐ࠪҶ䘉аһᇎǄ࠶᷀ᆖҐ֯⭘‫ݸ‬傼⸕䇶઼╄㓾᧘
⨶ᶕᢙབྷ䇝㓳ṧֻᨀ‫ؑⲴ׋‬᚟ˈഐ↔ᆳнਇ਼ṧⲴ⭼䲀ᡰࡦ㓖Ǆᵜㄐ㘳㲁Ҷа⿽〠ѪสҾ䀓䟺
ⲴᆖҐ˄EBL˅Ⲵ࠶᷀ᆖҐᯩ⌅Ǆ൘สҾ䀓䟺ⲴᆖҐѝˈ‫ݸ‬傼⸕䇶⭘Ҿ࠶᷀˄ᡆ㘵䀓䟺˅㿲ሏ
ࡠⲴᆖҐṧֻᱟᘾṧ┑䏣ⴞḷᾲᘥⲴǄ❦ਾ䘉њ䀓䟺㻛⭘Ҿ४࠶䇝㓳ṧֻѝଚӋᱟ⴨‫Ⲵޣ‬⢩
ᖱˈଚӋᱟн⴨‫Ⲵޣ‬Ǆ䘉ṧṧֻቡਟสҾ䙫䗁᧘⨶䘋㹼⌋ॆˈ㘼нᱟสҾ㔏䇑᧘⨶ǄสҾ䀓䟺
ⲴᆖҐᐢ㻛ᡀ࣏ൠ⭘Ҿ൘਴⿽㿴ࡂ઼䈳ᓖԫ࣑ѝᆖҐᩌ㍒᧗ࡦ㿴ࡉǄᵜㄐ㘳㲁ᆖҐಘⲴ‫ݸ‬傼⸕
䇶↓⺞ᒦфᆼᮤᰦⲴสҾ䀓䟺ⲴᆖҐǄлаㄐ㘳㲁‫ݸ‬傼⸕䇶ਚᱟ䘁լ↓⺞ᰦˈሶᖂ㓣ᆖҐ઼࠶
᷀ᆖҐ㔃ਸ䎧ᶕǄ
11.1 ԁ㔃
ࡽ䶒ㄐ㢲ᐢ㘳㲁Ҷ਴⿽ᖂ㓣⌅ˈণ䙊䗷⺞ᇊ㜭ཏ㓿傼ൠ४࠶↓઼ֻ৽ֻⲴ⢩ᖱˈᶕӾ㿲ሏ
ࡠⲴ䇝㓳ṧֻѝ⌋ॆǄߣㆆṁᆖҐǃ⾎㓿㖁㔌ᆖҐǃᖂ㓣䙫䗁㕆〻ǃԕ৺䚇Ր㇇⌅ᱟԕ䘉⿽ᯩ
ᔿ᫽֌Ⲵᖂ㓣ᆖҐᯩ⌅Ǆ䘉Ӌᖂ㓣ᆖҐಘ൘ᇎ䐥ѝⲴањ‫ޣ‬䭞䲀ࡦ൘ҾˈᆳԜ൘ਟ⭘ᮠᦞн䏣
ᰦᙗ㜭䖳ᐞǄᇎ䱵кˈྲㅜ 7 ㄐᡰ䇘䇪Ⲵˈ⨶䇪࠶᷀ᱮ⽪Ӿ㔉ᇊᮠⴞⲴ䇝㓳ṧֻѝᆖҐ൘㋮ᓖ
кᆈ൘สᵜⲴкл⭼Ǆ
ᱟ੖㜭ᔰਁࠪ䘉ṧⲴᆖҐᯩ⌅ˈᆳԜнਇਟ⭘䇝㓳ᮠᦞⲴᮠ䟿ᡰᑖᶕⲴ䇝㓳㋮ᓖкⲴสᵜ
䲀ࡦᡰࡦ㓖˛ㆄṸᱟ㛟ᇊⲴˈਚ㾱ᡁԜ㜭䟽ᯠ㘳㲁алᆖҐ䰞仈ⲴᖒᡀǄа⿽࣎⌅ᱟ֯ᆖҐ㇇
⌅㜭᧕ਇᱮᔿⲴ‫ݸ‬傼⸕䇶ˈ࣐к䇝㓳ᮠᦞⲴа਼֌Ѫ䗃‫ޕ‬ǄสҾ䀓䟺ⲴᆖҐᱟ䘉ṧⲴа⿽ᯩ
⌅Ǆᆳ֯⭘‫ݸ‬傼⸕䇶ᶕ࠶᷀ᡆ䀓䟺⇿њ䇝㓳ṧֻˈԕ᧘⨶ࠪṧֻⲴଚӋ⢩ᖱоⴞḷ࠭ᮠ⴨‫ˈޣ‬
ଚӋн⴨‫ޣ‬Ǆ䘉Ӌ䀓䟺㜭֯ᆖҐಘ∄অ⤜‫׍‬䶐ᮠᦞ䘋㹼⌋ॆᴹᴤ儈Ⲵ㋮ᓖǄྲࡽаㄐᡰ㿱ࡠⲴ
䛓ṧˈᖂ㓣䙫䗁㌫㔏˄ྲ Cigol˅֯⭘‫ݸ‬傼㛼Ჟ⸕䇶ᶕᤷሬᆖҐǄ❦㘼ᆳԜ֯⭘㛼Ჟ⸕䇶᧘⨶
ࠪⲴ⢩ᖱᢙབྷҶ䗃‫ޕ‬ᇎֻⲴ᧿䘠ˈഐ↔໎࣐Ҷᖵᩌ㍒‫ٷ‬䇮オ䰤Ⲵ༽ᵲᓖǄ⴨৽ˈสҾ䀓䟺Ⲵᆖ
Ґ֯⭘‫ݸ‬傼⸕䇶ᶕ߿ሿᖵᩌ㍒‫ٷ‬䇮オ䰤Ⲵ༽ᵲᓖˈഐ↔߿ሿҶṧᵜ༽ᵲᓖᒦᨀ儈ҶᆖҐಘⲴ⌋
ॆ㋮ᓖǄ
Ѫ⨶䀓สҾ䀓䟺ⲴᆖҐⲴⴤ㿲᜿ѹˈ㘳㲁лഭ䱵䊑ỻⲴᆖҐԫ࣑Ǆ⺞࠷ൠ䇢ˈ‫ٷ‬ᇊᡁԜᐼ
ᵋ〻ᒿᆖҐ䇶࡛ỻተս㖞Ⲵ䟽㾱㊫࡛ˈ∄ྲⴞḷᾲᘥĀ唁ỻሶ൘є↕޵ђਾⲴỻⴈ⣦ᘱāǄമ
11-1 ᱮ⽪Ҷ↔ⴞḷᾲᘥⲴањ↓ֻǄᖃ❦ˈᖂ㓣䙫䗁ᯩ⌅ҏ㜭⭘ҾᆖҐ↔ⴞḷᾲᘥǄ❦㘼ˈ⭡
Ҿỻⴈ⴨ᖃ༽ᵲ˄ᴹ 32 њᆀˈਟԕ൘ 64 њᯩṬѝ˅ˈ㘼ф↔ᾲᘥᡰ᧿䘠Ⲵ⢩ᇊ⁑ᔿ⴨ᖃᗞ࿉
˄वਜ਼Ҷỻⴈкн਼ᆀⲴ⴨ሩս㖞˅ˈᡁԜ䴰㾱ᨀ‫׋‬ᡀॳкзⲴ㊫լҾമ 11-1 䘉ṧⲴ䇝㓳ṧ
ֻˈ᡽㜭ᵏᵋᖂ㓣ᆖҐࡠⲴ‫ٷ‬䇮㻛↓⺞ൠ⌋ॆǄ
മ 11-1 ⴞḷᾲᘥĀ唁ỻ൘є↕޵ђਾⲴỻⴈ⣦ᘱāⲴањ↓ֻǄ
⌘᜿ⲭ傜਼ᰦ᭫ࠫ唁⦻઼唁ਾǄ唁ỻᗵ享〫ࣘަ⦻ˈӾ㘼ⲭỻ㜭ਲ਼ᦹ唁ਾǄ
ᆖҐлỻԫ࣑Ⲵᴹ䏓ѻ༴൘ҾˈӪ㊫ਚ㾱ቁᮠⲴ䇝㓳ṧֻቡਟᆖҐࡠ䘉ṧⲴⴞḷᾲᘥǄᇎ
䱵кˈཊᮠӪ൘ⴻҶമ 11-1 䘉ṧањṧֻѻਾቡਟᨀࠪањⴞḷᾲᘥⲴа㡜‫ٷ‬䇮ˈྲĀ唁ਾ
઼唁⦻਼ᰦ㻛᭫ࠫⲴᛵߥāˈ㘼нՊᗇࡠ䈨ྲ䘉ṧⲴ‫ٷ‬䇮˄նҏ਼ṧᱟа㠤‫ٷ‬䇮˅˖Āഋњⲭ
‫ޥ‬䘈൘৏սⲴỻⴈ⣦ᘱāǄӪ㊫ᱟᘾṧӾӵӵањṧֻѝᡀ࣏ൠ⌋ॆⲴ઒˛
എㆄᱟӪ㊫䶎ᑨ‫׍‬䎆ҾԕԆԜ‫ޣ‬Ҿỻᆀਸ⌅〫ࣘⲴ‫ݸ‬傼⸕䇶ᶕ䀓䟺ᡆ࠶᷀䇝㓳ṧֻǄྲ᷌
䰞ѪӰѸമ 11-1 Ⲵ䇝㓳ṧֻᱟĀ唁ỻ൘є↕޵ђਾāⲴ↓ֻˈཊᮠӪՊ㔉ࠪ㊫լҾл䶒Ⲵ䀓
䟺˖ĀഐѪⲭ傜਼ᰦ᭫ࠫ唁⦻઼唁ਾˈ唁ᗵ享᩶㝡㻛ሶߋⲴຳ䙷ˈӾ㘼䇙ⲭਲ਼ᦹਾǄā䈕䀓䟺
Ⲵ䟽㾱ᙗ൘Ҿᆳᨀ‫׋‬Ҷᡰ䴰Ⲵؑ᚟ԕӾ䇝㓳ṧֻⲴ㓶㢲ѝਸ⨶⌋ॆࡠ↓⺞Ⲵа㡜‫ٷ‬䇮Ǆ↔䀓䟺
ѝᨀࡠⲴṧֻ⢩ᖱ˄ྲⲭ傜ǃ唁⦻ǃ唁ਾⲴս㖞˅ᱟоⴞḷᾲᘥ⴨‫ˈⲴޣ‬ᒦфᓄ䈕㻛वਜ਼൘а
㡜‫ٷ‬䇮ѝǄ⴨৽ˈ䀓䟺ѝ⋑ᴹᨀࡠⲴṧֻ⢩ᖱ˄ྲⲭỻⲴ‫Ⲵޥ‬⣦ᘱ˅ਟ㻛䇔Ѫᱟн⴨‫Ⲵޣ‬㓶
㢲Ǆ
൘↔лỻֻᆀѝˈᆖҐಘѪᔪ・‫ٷ‬䇮ˈᆳ䴰㾱Ⲵ‫ݸ‬傼⸕䇶ウㄏᱟӰѸ઒˛ᖸㆰঅˈᱟлỻ
Ⲵਸ⌅㿴ࡉ˖ণ傜ԕ৺ަԆᆀⲴਸ⌅〫ࣘ˗ሩᔸ㘵ᗵ享Ӕᴯ〫ᆀ˗ԕ৺㾱䎒ỻᗵ享᥹տሩᯩⲴ
⦻Ǆ⌘᜿ਚ㔉ᇊ䘉ṧⲴ‫ݸ‬傼⸕䇶ˈ൘৏ࡉкቡᴹਟ㜭ሩԫ᜿ỻⴈ⣦ᘱ䇑㇇ࠪᴰՈⲴỻᆀ〫ࣘǄ
❦㘼ˈᇎ䐥ѝ䘉ṧⲴ䇑㇇ਟ㜭ᶱѪ༽ᵲˈ㘼фণ֯ᡁԜӪ㊫൘ᦼᨑҶ↔ᆼᮤⲴлỻ⸕䇶ˈӽн
㜭䗮ࡠᴰՈⲴሩᔸǄഐ↔ˈ൘лỻ˄ԕ৺ަԆᩌ㍒ᇶ䳶Ⲵ䰞仈ˈྲ䈳ᓖ઼㿴ࡂ˅䘉ṧⲴӪ㊫ᆖ
Ґѝˈवਜ਼Ҷањᖸ䮯Ⲵਁ⧠‫ݸ‬傼⸕䇶Ⲵ䗷〻ˈᆳᱟ⭡ᡁԜ൘лỻᰦ䙷ࡠⲴ⢩ᇊṧֻᡰᕅሬ
ⲴǄ
ᵜㄐ᧿䘠Ҷ㜭㠚ࣘᔪ・઼ᆖҐ䘉ṧⲴ䀓䟺ⲴᆖҐ㇇⌅ǄᵜㄐⲴ࢙։䜘࠶ሶᴤ㋮⺞ൠᇊѹ࠶
᷀ᆖҐ䰞仈Ǆла㢲㔉ࠪҶањ⢩ᇊⲴสҾ䀓䟺ⲴᆖҐ㇇⌅ˈ〠Ѫ Prolog-EBGǄਾ㔝ࠐ㢲㘳
ḕҶ䘉⿽㇇⌅Ⲵа㡜⢩ᙗˈԕ৺ᆳоࡽ䶒ㄐ㢲ѝ䇘䇪Ⲵᖂ㓣ᆖҐ㇇⌅ѻ䰤Ⲵ㚄㌫Ǆᴰਾа㢲᧿
䘠Ҷᓄ⭘สҾ䀓䟺ⲴᆖҐԕᨀ儈བྷ⣦ᘱオ䰤ᩌ㍒Ⲵᙗ㜭ǄᵜㄐᡁԜ㘳㲁Ҷа⿽⢩↺ᛵߥˈণ⭏
ᡀ䀓䟺ᡰสҾⲴ‫ݸ‬傼⸕䇶ᱟᆼ‫ྲˈⲴ⺞↓ޘ‬൘лỻֻᆀѝӪ㊫ᴹ↓⺞⸕䇶ⲴᛵᖒǄㅜ 12 ㄐሶ
㘳㲁ᴤа㡜ⲴᆖҐᛵߥˈণ‫ݸ‬傼⸕䇶ਚᱟ䘁լ↓⺞ⲴᛵߥǄ
11.1.1 ᖈ㓩ૂ࠼᷆ᆜҖ䰤从
࠶઼᷀ᖂ㓣ᆖҐ䰞仈Ⲵ䟽㾱४࡛൘ҾˈᆳԜ䇮ᜣⲴᆖҐ䰞仈Ⲵᖒᔿн਼˖
x ൘ᖂ㓣ᆖҐѝˈᆖҐಘ㻛㔉Ҹањ‫ٷ‬䇮オ䰤Hˈᆳᗵ享Ӿѝ䘹ᤙањ䗃ࠪ‫ٷ‬䇮Ǆ䘈
ᴹањ䇝㓳ṧֻ䳶ਸD={<x1, f(x1)>, …<xn, f(xn)>}ˈަѝf(xi)ѪᇎֻxiⲴⴞḷ٬Ǆᆖ
ҐಘᡰᐼᵋⲴ䗃ࠪѪHѝо䘉Ӌ䇝㓳ṧֻа㠤Ⲵ‫ٷ‬䇮hǄ
x ൘࠶᷀ᆖҐѝˈᆖҐಘⲴ䗃‫ޕ‬वਜ਼оᖂ㓣ᆖҐ਼ṧⲴ‫ٷ‬䇮オ䰤 H ઼䇝㓳ṧֻ DǄ
ᆖҐಘ䘈ᴹਖа䗃‫˖ޕ‬ањ亶亶ฏ⨶䇪(domain theory)Bˈᆳ⭡ਟ⭘Ҿ䀓䟺䇝㓳ṧֻ
Ⲵ㛼Ჟ⸕䇶㓴ᡀǄᆖҐಘⲴᐼᵋⲴ䗃ࠪѪ H ѝⲴ‫ٷ‬䇮 hˈᆳᰒо䇝㓳ṧֻ D а
㠤ˈҏо亶ฏ⨶䇪 B а㠤Ǆ
Ѫ䈤᰾䘉а⛩ˈ൘лỻⲴֻᆀѝ⇿њᇎֻxiਟ᧿䘠а⢩ᇊỻⴈ⣦ᘱˈf(xi)Ⲵ٬൘xiᱟ唁ỻ൘
є↕޵ђਾⲴỻⴈ⣦ᘱᰦѪⵏˈ੖ࡉѪ‫ٷ‬ǄᡁԜਟྲㅜ 10 ㄐ䛓ṧᇊѹ‫ٷ‬䇮オ䰤HѪHornᆀਕ
䳶˄ণif-then 㿴ࡉ˅ˈަѝ㿴ࡉᡰ֯⭘Ⲵ䉃䇽㺘⽪ỻⴈк⢩ᇊᆀⲴս㖞ᡆ⴨ሩս㖞Ǆ亶ฏ⨶䇪
Bਟ⭡ᖒᔿॆⲴлỻ㿴ࡉ㓴ᡀˈ᧿䘠Ҷਸ⌅Ⲵ䎠ỻǃሩᔸ㘵䖞⍱㹼ỻǃԕ৺᥹տտሩᯩ⦻ᰦ㧧
㜌ㅹǄ
⌘᜿൘࠶᷀ᆖҐѝˈᆖҐಘᗵ享䗃ࠪа‫ٷ‬䇮ˈᰒо䇝㓳ᮠᦞа㠤ˈ৸о亶ฏ⨶䇪а㠤Ǆᖃ
B н⏥㮤 h Ⲵ੖ᇊᰦ˄ণ Bƅ/ h˅ˈᡁԜ〠 h о亶ฏ⨶䇪 B а㠤˄consistent˅Ǆ↔䱴࣐Ⲵа㠤
ᙗ㓖ᶏˈ߿ቁҶᖃᮠᦞн㜭অ⤜൘ H ѝߣᇊ h ᰦᆖҐಘ䶒ѤⲴ↗ѹᙗǄྲ᷌亶ฏ⨶䇪↓⺞ˈަ
ᴰਾ᭸᷌ቡᱟᨀ儈Ҷ䗃ࠪ‫ٷ‬䇮Ⲵ㋮ᓖǄ
⧠䈖㓶ӻ㓽алᵜㄐਾ䶒аⴤ⭘ࡠⲴ࠶᷀ᆖҐ䰞仈ⲴਖањֻᆀǄ㘳㲁аᇎֻオ䰤 Xˈަ
ѝ⇿њᇎֻ䜭ᱟаሩ⢙⨶ሩ䊑Ǆ⇿ሩ⢙⨶ሩ䊑⭡䉃䇽 Color, Volume, Owner, Material, Type ઼
Density ᧿䘠ˈ㘼єњሩ䊑ѻ䰤Ⲵ‫ޣ‬㌫⭘䉃䇽 On ᧿䘠Ǆ൘↔‫ٷ‬䇮オ䰤ѝˈᆖҐԫ࣑ᱟᆖҐⴞḷ
ᾲᘥĀєњ⢙⨶ሩ䊑ˈањਟ㻛ᆹ‫ޘ‬ൠਐ᭮൘ਖањкāˈ㺘⽪Ѫ䉃䇽 SafeToStack(x,y)ǄᆖҐ
↔ⴞḷᾲᘥᴹᇎ⭘Ⲵԧ٬ˈֻྲањᵪಘӪ㌫㔏㾱൘аᴹ䲀オ䰤ѝᆈ᭮н਼Ⲵ⢙⨶ሩ䊑Ǆ↔࠶
᷀ᆖҐⲴᆼᮤᇊѹ൘㺘 11-1 ѝ㔉ࠪǄ
㺘 11-1 ањ࠶᷀ᆖҐ䰞仈˖SafeToStack(x, y)
ᐢ⸕˖
ᇎֻオ䰤 X˖⇿њᇎֻ᧿䘠Ҷаሩሩ䊑ˈ᧿䘠Ѫ䉃䇽 Type, Color, Volume, Owner, Material, Density ઼
OnǄ
‫ ٷ‬䇮オ䰤 H˖ ⇿ њ ‫ٷ‬䇮 ᱟа 㓴 Horn ᆀਕ 㿴 ࡉ Ǆ⇿њ Horn ᆀ ਕⲴ ཤ 䜘 Ѫа њव ਜ਼ ⴞ ḷ䉃䇽
SafeToStack Ⲵ᮷ᆇǄHorn ᆀਕփѪ᮷ᆇⲴਸਆˈ䘉Ӌ᮷ᆇสҾ᧿䘠ᇎֻⲴ䉃䇽ˈԕ৺䉃䇽 LessThan,
Equal, GreaterThan ઼࠭ᮠ plus, minus ઼ timesǄֻྲл䶒Ⲵ Horn ᆀਕᱟ‫ٷ‬䇮オ䰤ѝⲴаઈ˖
SafeToStack(x, y)ĕVolume(x, vx)ġVolume(y, vy) ġLessThan(vx, vy)
ⴞḷᾲᘥ˖SafeToStack(x,y)
䇝㓳ṧֻ˖л䶒ᱮ⽪ҶањިරⲴ↓ֻ SafeToStack(Obj1, Obj2)˖
On(Obj1, Obj2) Owner(Obj1, Fred)
Type(Obj1, Box) Owner(Obj2, Louise)
Type(Obj2, Endtable) Density(Obj1, 0.3)
Color(Obj1, Red) Material(Obj1, Cardboard)
Color(Obj2, Blue) Material(Obj2, Wood)
Volume(Obj1, 2)
亶ฏ⨶䇪 B˖
SafeToStack(x, y)ĕFragile(y)
SafeToStack(x, y) ĕLighter(x, y)
Lighter(x, y) ĕWeight(x, wx) ġWeight(y, wy) ġLessThan(wx, wy)
Weight(x, w) ĕVolume(x, v) ġDensity(x, d)ġEqual(w, times(v, d))
Weight(x, 5) ĕType(x, Endtable)
Fragile(x) ĕMaterial(x, Glass)
…
≲䀓˖
H ѝањ‫ٷ‬䇮ˈо䇝㓳ṧ઼ֻ亶ฏ⨶䇪а㠤Ǆ
ྲ㺘 11-1 ᡰ⽪ˈᡁԜ䘹ᇊⲴ‫ٷ‬䇮オ䰤 H ѝ⇿њ‫ٷ‬䇮Ѫања䱦 if-then 㿴ࡉ䳶ˈᡆ〠 Horn

ᆀਕ˄ᵜㄐѝ䚥ᗚ㺘 10-3 ѝࡇࠪⲴа䱦 Horn ᆀਕⲴ䇠ਧ઼ᵟ䈝˅Ǆֻྲˈ㺘ѝᱮ⽪Ⲵ Horn
ᆀਕ‫ٷ‬䇮Ⲵֻᆀᯝ䀰˖ᖃ x Ⲵփ〟 Volume ሿҾ˄LessThan ˅y Ⲵփ〟 Volume ᰦ˄൘ Horn ᆀਕ
ѝਈ䟿 vx ઼ vy ࠶࡛㺘⽪ x ઼ y Ⲵփ〟٬˅ˈࡉሩ䊑 x ਟᆹ‫ึޘ‬ਐ˄SafeToStack˅൘ሩ䊑 y кǄ
⌘᜿ Horn ᆀਕ‫ٷ‬䇮ਟवਜ਼⭘Ҿ᧿䘠ᇎֻⲴԫ᜿䉃䇽ˈԕ৺ࠐњ䱴࣐Ⲵ䉃䇽઼࠭ᮠǄ㺘ѝ䘈ᱮ
⽪ҶањިරⲴ↓ֻ SafeToStack(obj1, obj2)Ǆ
Ѫ᰾⺞ൠ㺘䗮↔࠶᷀ᆖҐ䰞仈ˈ䘈ᗵ享ᨀ‫׋‬亶ฏ⨶䇪ˈԕ‫࠶ݵ‬䀓䟺ѪӰѸ㿲ሏࡠⲴ↓ֻ┑
䏣ⴞḷᾲᘥǄ൘ࡽ䶒Ⲵлỻֻᆀ䟼ˈ亶ฏ⨶䇪Ѫỻᆀ䎠⌅Ⲵ⸕䇶ˈӾѝᡁԜᔪ・ࠪѪӰѸ唁ỻ
ՊђਾⲴ䀓䟺Ǆ൘ᖃࡽֻᆀѝˈ亶ฏ⨶䇪ᗵ享ᖸᇩ᱃䀓䟺ѪӰѸањሩ䊑ਟ᭮൘ਖањѻкǄ
㺘ѝᱮ⽪Ⲵ亶ฏ⨶䇪वᤜᯝ䀰˖Āਟሶ x ᆹ‫ޘ‬ൠਐ᭮൘ y кˈྲ᷌ y нᱟ᱃⺾Ⲵ˄Fragile˅ā
ԕ৺Āሩ䊑 x ᱟ᱃⺾Ⲵ(Fragile)ˈᖃ x Ⲵᶀ䍘(Material)ᱟ⧫⪳˄Glass˅ᰦǄāྲᆖҐࡠⲴ‫ٷ‬䇮
аṧˈ亶ฏ⨶䇪⭡а㓴 Horn ᆀਕ᧿䘠ˈᆳ֯㌫㔏৏ࡉкਟԕ࣐‫ޕ‬ԫօᆖҐࡠⲴ‫ٷ‬䇮㠣ਾ㔝Ⲵ
亶ฏ⨶䇪ѝǄ⌘᜿亶ฏ⨶䇪वᤜྲ Lighter ઼ Fragile 䘉ṧⲴ䱴࣐䉃䇽ˈᆳԜн൘䇝㓳ṧֻⲴ᧿
䘠ѝˈնᱟ⭡ᴤ৏ᆀⲴᇎֻ኎ᙗྲ Material, Density ઼ Volume ֯⭘亶ฏ⨶䇪ѝަԆ㿴ࡉ᧘⨶ᗇ
ࠪǄᴰਾˈ⌘᜿㺘ѝᱮ⽪Ⲵ亶ฏ⨶䇪‫࠶ݵ‬䇱᰾䘉䟼ᱮ⽪Ⲵ↓ֻ┑䏣ⴞḷᾲᘥ SafeToStackǄ
11.2 ⭞ᇂ㗄Ⲻ亼ต⨼䇰ᆜҖφProlog-EBG
ྲࡽᡰ䘠ˈᵜㄐѫ㾱㘳㲁ⲴสҾ䀓䟺ⲴᆖҐᱟ൘亶ฏ⨶䇪ᖸᆼ㖾ⲴᛵߥлⲴˈণ亶ฏ⨶䇪
ᱟ↓⺞ⲴᒦфᆼᮤⲴǄањ亶ฏ⨶䇪㻛〠Ѫ↓⺞Ⲵˈᖃަѝ⇿њᯝ䀰䜭ᱟц⭼Ⲵⵏᇎ᧿䘠Ǆа
њ亶ฏ⨶䇪㻛〠ѪᆼᮤⲴ˄ሩᓄ㔉ᇊⲴⴞḷᾲᘥ઼ᇎֻオ䰤˅ˈᖃ亶ฏ⨶䇪㾶ⴆҶᇎֻオ䰤ѝ
ᡰᴹ↓ֻǄᦒ䀰ѻˈަᆼᮤᙗ䈤᰾⇿њ┑䏣ⴞḷᾲᘥⲴᇎֻ䜭ਟ⭡亶ฏ⨶䇪䇱᰾ަ┑䏣ᙗǄ⌘
᜿ࡽ䶒ሩᆼᮤᙗⲴᇊѹн㾱≲亶ฏ⨶䇪ਟ䇱᰾৽ֻн┑䏣ⴞḷᾲᘥǄ❦㘼ˈྲ᷌䚥ᗚ䙊ᑨ
Prolog ᜟֻˈн㜭䇱᰾Ⲵᯝ䀰ਟ䇔ᇊᱟ‫ٷ‬Ǆഐ↔䈕ᆼᮤᙗᇊѹਟवਜ਼‫ޘ‬䜘↓઼ֻ৽ֻǄ
䈫㘵൘↔ਟ㜭Պ䰞ˈሩҾᆖҐಘ‫ٷ‬ᇊᴹ䘉ṧⲴᆼ㖾亶ฏ⨶䇪ᱟ੖ਸ⨶˛㘼фˈᰒ❦ᆖҐಘ
ᴹҶањᆼ㖾Ⲵ亶ฏ⨶䇪ˈ䘈ᴹօᗵ㾱޽৫ᆖҐ˛ሩҾ↔䰞仈ਟ᤹ԕлє⛩എㆄ˖
x 俆‫ˈݸ‬ḀӋᛵᖒлᱟᴹਟ㜭ᨀ‫׋‬ᆼ㖾亶ฏ⨶䇪ⲴǄࡽ䶒ⲴлỻⲴ䰞仈ቡᱟ䘉ṧⲴ
ањֻᆀˈަѝỻᆀⲴਸ⌅䎠ᆀᖒᡀҶањᆼ㖾Ⲵ亶ฏ⨶䇪ˈ˄৏ࡉк˅ਟ⭘ᆳ
ᶕ᧘⨶ᴰՈⲴлỻㆆ⮕Ǆᴤ䘋а↕ˈ㲭❦ᖸᇩ᱃߉ࠪᶴᡀ亶ฏ⨶䇪Ⲵỻᆀਸ⌅↕
ᆀˈ㾱߉ࠪᴰՈлỻㆆ⮕ӽ❦ᖸ䳮Ǆ൘䘉⿽ᛵߥлˈᡁԜᴤᐼᵋሶ䘉ṧⲴ亶ฏ⨶
䇪ᨀ‫׋‬㔉ᆖҐಘˈᒦᐼᵋᆖҐಘᖒᡀⴞḷᾲᘥⲴᴹᑞࣙⲴ᧿䘠˄ྲ˖Āਟ㜭ђਾ
Ⲵỻተ⣦ᘱā˅Ǆᯩ⌅ᱟ䙊䗷ሩ⢩↺䇝㓳ṧֻ䘋㹼㘳ḕ઼⌋ॆǄ11.4 㢲᧿䘠Ҷ֯⭘
ᆼ㖾亶ฏ⨶䇪ⲴสҾ䀓䟺ⲴᆖҐᡀ࣏ൠᓄ⭘ࡠࠐњᩌ㍒ᇶ䳶Ⲵ䇑ࡂ઼Ոॆ䰞仈
ѝˈԕ㠚ࣘ᭩䘋ᙗ㜭Ǆ
x ㅜҼˈ൘䇨ཊᛵߥлн㜭ཏ‫ٷ‬ᇊᴹᆼ㖾Ⲵ亶ฏ⨶䇪Ǆ∄ྲᖸ䳮Ѫࡽ䶒䘉њ⴨ሩㆰ
অⲴ SafeToStack 䰞仈㔉ࠪᆼᮤ㘼↓⺞Ⲵ亶ฏ⨶䇪Ǆᴤᇎ䱵Ⲵᯩ⌅ᱟ‫ٷ‬ᇊᗵ享֯⭘
สҾнᆼ㖾亶ฏ⨶䇪Ⲵ䘁լਸ⨶Ⲵ䀓䟺ˈ㘼нᱟสҾᆼ㖾⸕䇶֌ࠪ⺞࠷䇱᰾Ǆᰐ
䇪ᘾṧˈᡁԜਟԕ䙊䗷㘳㲁⨶ᜣᛵߥлⲴᆼ㖾亶ฏ⨶䇪ˈᔰ࿻Ҷ䀓൘ᆖҐѝ֯⭘
䀓䟺ⲴⴞⲴǄㅜ 12 ㄐᡁԜሶ㘳㲁Ӿнᆼ㖾亶ฏ⨶䇪ѝᆖҐǄ
ᵜ㢲ኅ⽪Ҷањ〠Ѫ Prolog-EBG Ⲵ㇇⌅ (Kedar-Cabelli & McCarty 1987)ˈ֯⭘ᆳ֌Ѫࠐ⿽

สҾ䀓䟺ⲴᆖҐⲴԓ㺘ǄProlog-EBG ᱟаᒿࡇ㾶ⴆ㇇⌅˄㿱ㅜ 10 ㄐ˅Ǆᦒ䀰ѻˈᆳⲴ䗷〻ᱟ
ᆖҐঅњ Horn ᆀਕ㿴ࡉˈ〫৫↔㿴ࡉ㾶ⴆⲴ↓ֻˈ޽൘࢙։↓ֻк䟽༽䘉а䗷〻ˈⴤࡠ⋑ᴹ
ᵚ㾶ⴆⲴ↓ֻѪ→Ǆ㤕㔉ᇊаᆼᮤᒦ↓⺞Ⲵ亶ฏ⨶䇪ˈProlog-EBG ‫؍‬䇱䗃ࠪањ‫ٷ‬䇮˄㿴ࡉ
䳶˅ˈᆳᵜ䓛ᱟ↓⺞Ⲵᒦ㜭㾶ⴆ㿲ሏࡠⲴ↓ֻǄሩԫ᜿↓ֻ䳶ਸˈ⭡ Prolog-EBG 䗃ࠪⲴ‫ٷ‬䇮
वਜ਼а㓴ሩᓄҾ亶ฏ⨶䇪ⲴⴞḷᾲᘥⲴ䙫䗁‫࠶ݵ‬ᶑԦǄProlog-EBG ᱟ Mitchell et al.˄1986˅ӻ
㓽Ⲵ EBG ㇇⌅Ⲵ᭩䘋ˈᒦф㊫լҾ Dejong & Mooney˄1986˅᧿䘠Ⲵ EGGS ㇇⌅ǄProlog-
EBG ㇇⌅൘㺘 11-2 ѝࡇࠪǄ
11.2.1 䘆㺂⽰ׁ
Ѫ䈤᰾䈕㇇⌅ˈ޽⅑㘳㲁㺘 11-1 㔉ࠪⲴ䇝㓳ṧ઼ֻ亶ฏ⨶䇪Ǆ㺘 11-2 ࡇࠪⲴ Prolog-EBG

㇇⌅ᱟаᒿࡇ㾶ⴆ㇇⌅ˈᆳ⑀䘋ൠ㘳㲁䇝㓳ᮠᦞǄሩ⇿њᯠ↓ֻˈ㤕ᆳ䘈⋑㻛аᆖࡠⲴ Horn
ᆀਕ㾶ⴆˈ㇇⌅䙊䗷лࡇ↕僔⭏ᡀаᯠⲴ Horn ᆀਕ˖˄1˅䀓䟺ᯠⲴ↓ֻˈ˄2˅࠶᷀䈕䀓䟺
ԕ⺞ᇊаਸ䘲Ⲵ⌋ॆˈ˄3˅䙊䗷࣐‫ޕ‬аᯠⲴ Horn ᆀਕԕ㾶ⴆ䈕↓ֻԕ৺ަԆ⴨լᇎֻ᭩䘋ᖃ
ࡽ‫ٷ‬䇮Ǆл䶒ᡁԜ‫⅑׍‬㘳ḕ䘉йњ↕僔Ǆ
11.2.1.1 䀙䠀䇣㓹ṭׁ
༴⨶⇿њᯠṧֻⲴㅜа↕ᱟ᤹➗亶ฏ⨶䇪ᔪ・а䀓䟺ˈԕ䈤᰾䈕↓ֻྲօ┑䏣ⴞḷᾲᘥǄ
ᖃ亶ฏ⨶䇪↓⺞фᆼᮤᰦˈ↔䀓䟺㓴ᡀҶ䇝㓳ṧֻ┑䏣ⴞḷᾲᘥⲴањ䇱᰾˄proof˅Ǆྲ᷌
‫ݸ‬傼⸕䇶нᆼ㖾ˈ䀓䟺ѝⲴ䇠ਧᗵ享㻛ᢙኅԕ‫ݱ‬䇨䘁լⲴ৲ᮠˈ㘼нᱟᆼ㖾Ⲵ䇱᰾Ǆ
㺘 11-2 สҾ䀓䟺ⲴᆖҐ㇇⌅ Prolog-EBG
ሩ⇿њ䘈⋑㻛ᆖҐࡠⲴ Horn ᆀਕ䳶(LearnedRules)㾶ⴆⲴ↓ֻˈᔪ・ањᯠ Horn ᆀਕǄ䈕ᯠⲴ

Horn ᆀਕⲴࡋᔪᱟ䙊䗷˄1˅᤹亶ฏ⨶䇪䀓䟺䇝㓳ṧֻˈ˄2˅࠶᷀↔䀓䟺ԕ⺞ᇊṧֻⲴ⴨‫ޣ‬⢩ᖱˈ
˄3˅ᔪ・аᯠⲴ Horn ᆀਕˈᆳ൘䈕㓴⢩ᖱ┑䏣ᰦᗇࡠⴞḷᾲᘥǄ
Prolog-EBG(TargetConcept, TrainingExamples, DomainTheory)

LearnedRulesĕ{}
PosĕTrainingExamples ѝⲴ↓ֻ
ሩ Pos ѝ⋑ᴹ㻛 LearnedRules 㾶ⴆⲴ⇿њ PositiveExampleˈ‫ڊ‬ԕл᫽֌˖
1. 䀓䟺
z Explanation ĕ а њ ԕ DomainTheory 㺘⽪ Ⲵ 䀓䟺 ˄ 䇱 ᰾ ˅ ˈ 䈤 ᰾ Ѫ օ PositiveExample ┑ 䏣
TargetConcept
2. ࠶᷀
z SuffcientConditionsĕ᤹➗ Explanationˈ㜭ཏ‫┑࠶ݵ‬䏣 TargetConcept Ⲵ PositiveExample Ⲵᴰа㡜
⢩ᖱ䳶ਸ
3. ᭩䘋
z LearnedRulesĕLearnedRules+NewHornClauseˈަѝ NewHornClause ᖒᔿѪ˖
TargetConceptĕSufficientConditions
䘄എ LearnedRules
ሩ ᖃ ࡽ ṧ ֻ Ⲵ 䀓䟺㿱 മ 11-2 Ǆ ⌘ ᜿ ަ ѝ ᓅ 䜘 Ⲵ മ ᖒ ԓ 㺘 Ҷ 㺘 11-1 ѝ Ⲵ ↓ ֻ
SafeToStack(Obj1, Obj2)Ǆമѝк䜘Ѫሩ↔ṧֻᶴ䙐Ⲵ䀓䟺Ǆ⌘᜿↔䀓䟺˄ᡆ〠䇱᰾˅䈤᰾ഐѪ
Obj1 ∄ Obj2 ᴤ 䖫˄Lighter˅ˈᡰԕ Obj1 ਟԕᆹ‫ึޘ‬ਐ˄SafeToStack˅൘ Obj2 кǄᴤ䘋а
↕ˈ⸕䚃 Obj1 ᴤ䖫ᱟഐѪᆳⲴ䟽䟿˄Weight˅ਟԕ⭡ަᇶᓖ˄Density˅઼փ〟˄Volume˅᧘
ᗇˈ㘼ф Obj2 Ⲵ 䟽䟿˄Weight˅ਟӾ㥦ࠐ˄Endtable˅Ⲵ唈䇔Ⲵ䟽䟿˄Weight˅٬ᗇࠪǄ↔䀓
䟺สҾⲴ⢩ᇊ Horn ᆀਕ൘㺘 11-1 Ⲵ亶ฏ⨶䇪ѝᱮ⽪ࠪǄ⌘᜿↔䀓䟺ਚᨀࡠҶ Obj1 ઼ Obj2 Ⲵ
኎ᙗѝⲴаሿ䜘࠶˄ণሩᓄҾമѝ䱤ᖡ४ฏⲴ኎ᙗ˅Ǆ
㲭❦䘉䟼ሩҾ䇝㓳ṧ઼ֻ亶ฏ⨶䇪ਚᴹа⿽䀓䟺ˈа㡜ᛵߥлਟ㜭ᴹཊ⿽䀓䟺Ǆ䘉ṧˈ䘉
Ӌ䀓䟺ѝԫ᜿ⲴᡆᡰᴹⲴ䜭ਟ㻛֯⭘Ǆ⇿њ䀓䟺ਟሩ䇝㓳ṧֻᖒᡀн਼Ⲵ⌋ॆˈնᡰᴹ䀓䟺䜭
ሶ㻛㔉ᇊⲴ亶ฏ⨶䇪䇪䇱Ǆ൘ Prolog-EBG ѝˈ䀓䟺Ⲵ⭏ᡀ֯⭘Ҷྲ Prolog ѝⲴਾੁ䬮ᔿᩌ
㍒ǄProlog-EBG ྲ Prolog аṧˈ൘ᆳ᢮ࡠㅜањᴹ᭸䇱᰾ᰦ㓸→Ǆ
11.2.1.2 ࠼᷆䀙䠀
൘⌋ॆ䇝㓳ṧֻᰦ䶒ѤⲴ‫ޣ‬䭞䰞仈ᱟĀ൘ᖃࡽṧֻѝ䇨ཊ↓ྭѪⵏⲴ⢩ᖱѝˈଚањᱟ൘
а㡜ᛵߥлоⴞḷᾲᘥ⴨‫˛Ⲵޣ‬ā⭡ᆖҐಘᶴ䙐Ⲵ䀓䟺ሩ↔䰞仈֌ࠪҶⴤ᧕Ⲵഎㆄ˖↓ྭᱟ䛓
Ӌ൘䀓䟺ѝᨀ৺Ⲵ⢩ᖱǄֻྲˈമ 11-2 Ⲵ䀓䟺वਜ਼Ҷ Obj1 Ⲵ Densityˈն⋑ᴹᆳⲴ Owner ኎
ᙗǄഐ↔ˈSafeToStack(x,y)Ⲵ‫ٷ‬䇮ᓄवਜ਼ Density(x,0.3)ˈ㘼нवਜ਼ Owner(x,Fred)Ǆ䙊䗷᭦䳶മ
11-2 ѝ䀓䟺Ⲵਦ㔃⛩ѝᨀ৺Ⲵ⢩ᖱˈᒦሶ Obj1 ઼ Obj2 ᴯᦒѪ x ઼ yˈਟᖒᡀањ⭡亶ฏ⨶䇪
䇪䇱Ⲵа㡜㿴ࡉǄ
SafeToStac k ( x, y ) m Volume ( x,2) Density ( x,0.3) Type ( y , Endtable )
Explanation: 䀓䟺
Training Example: 䇝㓳ṧֻ
മ 11-2 䇝㓳ṧֻⲴ䀓䟺
л䜘Ⲵ㖁㔌ԕമᖒ㔈ࠪҶ㺘 11-1 ѝⲴ䇝㓳ṧֻ SafeToStack(Obj1,Obj2)Ǆമк䶒䜘࠶㔈ࠪҶ↔ṧ

ֻᘾṧ┑䏣ⴞḷᾲᘥ SafeToStack Ⲵ䀓䟺Ǆ䇝㓳ṧֻѝⲴ䱤ᖡ䜘࠶㺘⽪൘䀓䟺ѝ⭘ࡠⲴṧֻ኎ᙗǄަԆ
н⴨‫Ⲵޣ‬ṧֻ኎ᙗሶӾᖒᡀⲴ⌋ॆ‫ٷ‬䇮ѝ৫ᦹǄ
к 䶒 Ⲵ 㿴 ࡉ փ व ਜ਼ Ҷ 䇱 ᰾ ṁ ѝ ⇿ њ ਦ 㔃 ⛩ ˈ 䲔 Ҷ Ā Equal(0.6, times(2,0.3)) ā ઼
ĀLessThan(0.6, 5)āѻཆǄ৫ᦹ䘉єњᱟഐѪṩᦞᇊѹᆳԜᙫᱟ㻛┑䏣Ⲵˈо x ઼ y ᰐ‫ޣ‬Ǆ
䘎਼↔ᆖࡠⲴ㿴ࡉа䎧ˈ〻ᒿ䘈ਟԕᨀ‫ަ׋‬䇪䇱˖ሩ䇝㓳ṧֻⲴ䀓䟺ᖒᡀҶሩ↔㿴ࡉ↓⺞
ᙗⲴ䇱᰾Ǆ㲭❦↔䀓䟺ᱟѪҶ㾶ⴆ㿲ሏࡠⲴ䇝㓳ṧֻ㘼ᖒᡀˈ਼ṧⲴ䀓䟺ሶ䘲⭘Ҿԫօо↔а
㡜㿴ࡉ३䝽ⲴᇎֻǄ
к䶒Ⲵ㿴ࡉᶴᡀҶ↔䇝㓳ṧֻⲴањᖸᴹ᜿ѹⲴ⌋ॆˈഐѪᆳ৫䲔ҶṧֻⲴ䇨ཊоⴞḷᾲ
ᘥᰐ‫ޣ‬኎ᙗ˄ྲєњሩ䊑Ⲵ Color˅ˈ❦㘼䙊䗷ሩ䀓䟺ᴤԄ㓶ൠ࠶᷀ਟԕᗇࡠᴤа㡜Ⲵ㿴ࡉǄ
Prolog-EBG ਟ䇑㇇㜭⭡䀓䟺䇪䇱Ⲵᴰа㡜Ⲵ㿴ࡉˈᯩ⌅䙊䗷䇑㇇䀓䟺Ⲵᴰᕡࡽ‫˄ۿ‬weakest
preimage˅ˈᇊѹྲл˖
定义：㔃䇪 C ሩᓄҾ䇱᰾ P Ⲵᴰ ᴰᕡࡽ‫˄ۿ‬weakest preimage˅Ѫᴰа㡜Ⲵࡍ࿻ᯝ䀰䳶

ਸ Aˈ֯ᗇ A ᤹➗ P ⏥㮤 CǄ
ֻྲˈⴞḷᾲᘥ SafeToStack(x,y)ሩᓄ㺘 11-1 䀓䟺Ⲵᴰᕡࡽ‫⭡ۿ‬л䶒㿴ࡉⲴփ㔉ࠪǄ䘉ᱟ㜭

⭡മ 11-2 Ⲵ䀓䟺䇪䇱Ⲵᴰа㡜㿴ࡉ˖
SafeToStack(x, y)ĕVolume(x, vx)ġDensity(x,dx) ġ

Equal(wx, times(vx, dx)) ġLessThan(wx,5) ġ
Type(y, Endtable)
⌘᜿䘉њᴤа㡜Ⲵ㿴ࡉн㾱≲㔉ࠪ Volume ઼ Density Ⲵ⢩ᇊ٬ˈ㘼ࡽањ㿴ࡉ䴰㾱Ǆᆳਚ
ᱟሩ䘉↔䘉Ӌ኎ᙗⲴ٬䘋㹼ᴤа㡜Ⲵ㓖ᶏǄ
Prolog-EBG 䇑㇇ⴞḷᾲᘥⲴ‫ޣ‬Ҿ䀓䟺Ⲵᴰᕡࡽ‫Ⲵۿ‬䗷〻ˈ֯⭘Ⲵᱟа〠Ѫഎᖂ (regression)

Ⲵ䗷〻˄Waldinger 1977˅Ǆഎᖂ䗷〻䪸ሩⲴᱟ⭡ԫ᜿ Horn ᆀਕ䳶㺘⽪Ⲵ亶ฏ⨶䇪ǄᆳⲴᐕ֌
ᯩᔿᱟ൘䀓䟺ѝ৽༽ൠਾ䘰ˈ俆‫ݸ‬ሩᓄҾ䀓䟺ѝᴰᵛ䇱᰾↕䇑㇇ⴞḷᾲᘥⲴᴰᕡࡽ‫❦ˈۿ‬ਾሩ
ᓄҾަࡽа↕僔䇑㇇㔃᷌㺘䗮ᔿⲴᴰᕡࡽ‫↔׍ˈۿ‬㊫᧘Ǆ䈕䗷〻൘䙽শ䗷䀓䟺ѝᡰᴹ↕僔ਾ㓸
→ˈᗇࡠሩᓄҾ䀓䟺Ⲵਦ㢲⛩кⲴ᮷ᆇⲴⴞḷᾲᘥⲴᴰᕡࡽԦǄ
↔എᖂ䗷〻Ⲵ䘀㹼↕僔㿱മ 11-3 ᡰ⽪Ǆ൘↔മѝˈമ 11-2 ѝࠪ⧠䗷Ⲵ䀓䟺ԕḷ߶ᆇփ

˄䶎ᯌփ˅䟽⭫ࠪǄ㘼൘⇿а↕⭡എᖂ䗷〻ࡋᔪⲴ䗩㕈എᖂ㺘䗮ᔿԕᑖлࡂ㓯Ⲵᯌփᆇᱮ⽪Ǆ
↔䗷〻ᔰ࿻ҾṁⲴṩ䜘ˈަ䗩㕈㻛ࡍ࿻ॆѪа㡜ⴞḷᾲᘥ SafeToStack(x,y)Ǆㅜа↕ᱟ䇑㇇↔䗩
㕈㺘䗮ᔿሩᓄҾ䀓䟺ѝᴰᵛ˄ᴰк䶒Ⲵ˅᧘⨶㿴ࡉⲴᴰᕡࡽ‫ۿ‬Ǆ൘↔ᛵᖒл㿴ࡉѪ
SafeToStack(x, y) ĕLighter(x, y)ˈഐ↔ᗇࡠⲴᴰᕡࡽ‫ۿ‬Ѫ Lighter(x,y)Ǆ❦ਾˈ䙊䗷↔䀓䟺ѝла
Horn ᆀਕˈ䈕䗷〻㔗㔝ሩ↔ᯠ䗩㕈{Lighter(x, y)}䘋㹼എᖂˈᗇࡠഎᖂ㺘䗮ᔿ{Weight(x, wx),
LessThan(wx, wy), Weight(y, wy)}Ǆ↔ᔿ᜿ણ⵰ˈሩҾԫ᜿Ⲵ x ઼ yˈ㤕 x Ⲵ䟽䟿 wx བྷҾ y Ⲵ䟽
䟿 wy ˈ䀓䟺ᡀ・Ǆ↔䗩㕈Ⲵഎᖂԕ↔а↕↕Ⲵᯩᔿ䘰എࡠ䀓䟺Ⲵਦ㔃⛩ˈᴰ㓸ᗇࡠṁⲴਦ㔃
⛩кⲴа㓴⌋ॆ᮷ᆇǄ↔ᴰ㓸Ⲵ⌋ॆ᮷ᆇ䳶ਸˈྲമ 11-3 ᓅ䜘ᡰ⽪ˈᖒᡀҶᴰ㓸㿴ࡉⲴ㿴ࡉ
փǄ
മ 11-3 䇑㇇ SafeToStack(Obj1,Obj2)‫ޣ‬Ҿ䀓䟺Ⲵᴰᕡࡽ‫ۿ‬
ⴞḷᾲᘥӾ䀓䟺Ⲵṩ䜘˄㔃䇪˅ᔰ࿻എᖂˈл䱽ࡠਦ㔃⛩Ǆ൘⇿а↕˄⭡㲊㓯㺘⽪˅ˈᖃࡽ᮷ᆇ
䳶ਸ䗩㕈˄ᑖлࡂ㓯Ⲵᯌփ˅൘䀓䟺Ⲵањ㿴ࡉк㻛ਾੁഎᖂǄᖃ↔䗷〻ᆼᡀᰦˈ㔃᷌᮷ᆇਸਆᶴᡀ
ҶሩᓄҾ䀓䟺ⲴⴞḷᾲᘥⲴᴰᕡࡽ‫ۿ‬Ǆ↔ᴰᕡࡽ‫ۿ‬൘മⲴᓅ䜘ԕᯌփⲴ᮷ᆇᱮ⽪Ǆ
എᖂ䗷〻ⲴṨᗳᱟˈ൘⇿а↕䙊䗷亶ฏ⨶䇪Ⲵаᶑ Horn ᆀਕഎᖂᖃࡽ䗩㕈㺘䗮ᔿⲴ㇇

⌅Ǆ↔㇇⌅൘㺘 11-3 ѝ᧿䘠ᒦֻ⽪Ǆ㺘ѝⲴ㤳ֻሩᓄҾമ 11-3 ѝᴰᓅ䜘Ⲵഎᖂ↕Ǆྲ㺘ѝᱮ
⽪ⲴˈRegress ㇇⌅Ⲵ᫽֌䗷〻ᱟˈራ᢮ањ㖞ᦒ֯ Horn ᆀਕⲴཤо䗩㕈ѝⲴ⴨ᓄ᮷ᆇਸаˈ
⭘㿴ࡉփᴯᦒ䗩㕈ѝⲴ↔㺘䗮ᔿˈ޽ᓄ⭘ањਸа㖞ᦒࡠᮤњ䗩㕈Ǆ
⭡Prolog-EBG䗃ࠪⲴᴰ㓸Hornᆀਕᖒᔿྲл˖ᆀਕփ㻛ᇊѹѪк䘠䗷〻䇑㇇ࠪⲴᴰᕡࡽ
ԦǄᆀਕཤѪⴞḷᾲᘥᵜ䓛ˈԕ৺ᓄ⭘ࡠᆳкⲴ⇿аഎᖂ↕ѝⲴ⇿њ㖞ᦒ˄ྲ㺘 11-3 ѝⲴ㖞
ᦒ©hl˅Ǆᓄ⭘↔㖞ᦒᱟѪҶ൘ࡋᔪࠪⲴᆀਕཤ઼ᆀਕփѝ‫؍‬ᤱа㠤ਈ䟿਽ˈԕ৺ᖃ↔䀓䟺ਚ
ᓄ⭘ҾⴞḷᾲᘥⲴ⢩↺ᛵߥᰦ⢩ॆᆀਕཤǄྲࡽᤷࠪⲴˈሩҾᖃࡽⲴֻᆀˈᴰ㓸㿴ࡉѪ˖
SafeToStack(x, y)ĕVolume(x, vx)ġDensity(x,dx) ġ

Equal(wx, times(vx, dx)) ġLessThan(wx,5) ġ
Type(y, Endtable)
㺘 11-3 䙊䗷ањ Horn ᆀਕഎᖂа㓴᮷ᆇⲴ㇇⌅
⭡䗩㕈˄Frontier˅㔉ࠪⲴ᮷ᆇ䳶ਸ䙊䗷Rule㻛എᖂǄLiteralѪ↔䀓䟺ѝ⭡Rule᧘⨶ⲴFrontierᡀ
ઈǄ㖞ᦒ ©hi 㔉ࠪҶӾRuleⲴཤࡠ䀓䟺ѝሩᓄ᮷ᆇⲴਈ䟿㓖ᶏǄ↔㇇⌅俆‫ݸ‬䇑㇇ањ㜭֯RuleⲴཤо
LiteralਸаⲴ㖞ᦒ © hl ˈަᯩ⌅ᱟ֯ަо㖞ᦒ © hi а㠤Ǆ❦ਾ↔㖞ᦒ © hl 㻛ᓄ⭘Ҿᔪ・‫ޣ‬ҾRuleⲴ
FrontierⲴࡽ‫ۿ‬Ǆ㇇⌅ѝㅖਧĀˇā઼Āˉā㺘⽪䳶ਸᒦ઼䳶ਸᐞǄ䇠ਧ{z/y}㺘⽪⭘y㖞ᦒzǄ㺘ѝ䘈㔉
ࠪҶ࠶↕䘀㹼ⲴֻᆀǄ
Regress(Frontier, Rule, Literal, ©hi)

Frontier˖䙊䗷㿴ࡉ㻛എᖂⲴ᮷ᆇ䳶ਸ
Rule˖ањ Horn ᆀਕ
Literal˖൘ Frontier ѝⲴањ᮷ᆇˈᆳ⭡䀓䟺ѝⲴ Rule ᧘ᗇ
©hi˖֯RuleⲴཤо䀓䟺ѝⲴ⴨ᓄ᮷ᆇਸаⲴ㖞ᦒ
䘄എᶴᡀ Frontier Ⲵ‫ޣ‬Ҿ Rule Ⲵᴰᕡࡽ‫Ⲵۿ‬᮷ᆇ䳶ਸ
headĕRule Ⲵཤ
bodyĕRule Ⲵփ
©hlĕheadоLiteralⲴᴰа㡜ਸаˈ֯ᗇᆈ൘㖞ᦒ©li┑䏣˖
©li (©hl (head))= ©hi (head)
䘄എ©hl(Frontier-head+body)
⽪ֻ˄മ 11-3 ѝᴰл䶒Ⲵഎᖂ↕˅˖
Regress(Frontier, Rule, Literal, ©hi)ˈަѝ
Frontier = {Volume(x, vx), Density(x, dx), Equal(wx, times(vx, dx)), LessThan(wx, wy), Weight(y, wy)}
Rule = Weight(z, 5) ĕType(z, Endtable)
Literal = Weight(y, wy)
©hi = {z/Obj2}
headĕWeight(z, 5)
bodyĕType(z, Endtable)
©hlĕ{z/y, wy/5}ˈަѝ©li={y/Obj2}
䘄എ{Volume(x, vx), Density(x, dx), Equal(wx, times(vx, dx)), lessThan(wx, 5), Type(y, Endtable)}
11.2.1.3 ᭯䘑ᖉࢃ‫ٽ‬䇴
൘⇿а䱦⇥Ⲵᖃࡽ‫ٷ‬䇮⭡ᖃᰦᆖҐࡠⲴ Horn ᆀਕ䳶㓴ᡀǄ൘⇿а䱦⇥ˈᒿࡇ㾶ⴆ㇇⌅䘹

ਆањ䘈ᵚ㻛ᖃࡽ Horn ᆀਕ㾶ⴆⲴᯠ↓ֻˈ䀓䟺䈕↓ֻˈᒦ᤹➗к䶒Ⲵ䗷〻ᖒᡀᯠ㿴ࡉǄ⌘
᜿ᡁԜᐢᇊѹⲴ㇇⌅ѝਚᴹ↓ֻ㻛㾶ⴆˈ㘼фᆖҐࡠⲴ Horn ᆀਕ䳶ਚ亴⍻↓ֻǄሩҾањᯠ
ᇎֻˈྲ᷌ᖃࡽ㿴ࡉ亴⍻ަ↓ֻཡ䍕ˈࡉᆳ㻛࠶㊫Ѫ৽ֻǄ䘉ᱟо Prolog 䘉ṧⲴ Horn ᆀਕ᧘
⨶㌫㔏ѝḷ߶Ⲵཡ䍕੖ᇊᯩ⌅⴨੫ਸⲴǄ
11.3 ሯะӄ䀙䠀ⲺᆜҖⲺ䈪᱄
ྲᡁԜ൘кֻѝⴻࡠⲴ Prolog-EBG ሩঅњ䇝㓳ṧֻ䘋㹼䈖㓶࠶᷀ˈԕ⺞ᇊྲօᴰྭൠӾ

⢩↺ṧֻ⌋ॆࡠа㡜 Horn ᆀਕ‫ٷ‬䇮Ǆл䶒Ѫ↔㇇⌅Ⲵ㾱⛩˖
x Prolog-EBG н‫ۿ‬ᖂ㓣Ⲵᯩ⌅ˈᆳ䙊䗷䘀⭘‫ݸ‬傼⸕䇶࠶᷀অњṧֻԕӗ⭏ਸ⨶Ⲵ
˄justified˅а㡜‫ٷ‬䇮Ǆ
x ሩṧֻྲօ┑䏣ⴞḷᾲᘥⲴ䀓䟺ˈ⺞ᇊҶṧֻⲴଚӋ኎ᙗᱟ⴨‫˖Ⲵޣ‬ণ൘䀓䟺ѝ
ᨀ৺Ⲵ኎ᙗǄ
x ሩ䀓䟺Ⲵ䘋а↕࠶᷀ˈণഎᖂⴞḷᾲᘥԕ⺞ᇊަሩᓄ䀓䟺Ⲵᴰᕡࡽ‫ˈۿ‬ਟ᧘ሬࠪ
⴨‫ޣ‬⢩ᖱ٬Ⲵа㡜㓖ᶏǄ
x ⇿њᆖҐࡠⲴ Horn ᆀਕሩᓄҾ┑䏣ⴞḷᾲᘥⲴањ‫࠶ݵ‬ᶑԦǄᆖҐࡠⲴ Horn ᆀ
ਕ䳶㾶ⴆҶᆖҐಘ䙷ࡠⲴ↓ֻˈԕ৺ަԆо↔‫ޡ‬ӛ਼ṧ䀓䟺ⲴᇎֻǄ
x ᆖҐࡠⲴ Horn ᆀਕⲴ⌋ᙗሶ‫׍‬䎆Ҿ亶ฏ⨶䇪Ⲵᖒᔿԕ৺䇝㓳ṧֻ㻛㘳㲁ⲴᒿࡇǄ
x Prolog-EBG 䳀ਜ਼‫ٷ‬ᇊҶ亶ฏ⨶䇪ᱟ↓⺞фᆼᮤⲴˈྲ᷌亶ฏ⨶䇪н↓⺞ᡆнᆼ
ᮤˈᆖࡠⲴᾲᘥҏሶн↓⺞Ǆ
൘สҾ䀓䟺ⲴᆖҐѝᴹаӋ⴨‫Ⲵޣ‬㿲⛩ˈਟᴹࣙҾ⨶䀓ަ㜭઼࣋䲀ࡦ˖
x EBL ֌Ѫ⨶ ⨶䇪ᕅሬⲴṧֻ⌋ॆ˄theory-guided generalization of examples˅ǄEBL ֯
⭘㔉ᇊⲴ亶ฏ⨶䇪ԕӾṧֻѝਸ⨶ൠ⌋ॆˈ४࠶ࠪ⴨‫઼ޣ‬н⴨‫Ⲵޣ‬ṧֻ኎ᙗˈഐ
↔ਟԕ䚯‫⭘ݽ‬Ҿ㓟ᖂ㓣᧘⨶ѝⲴṧᵜ༽ᵲᓖ⭼䲀Ǆ䘉ᱟањ䳀ਜ਼൘к䶒᧿䘠Ⲵ
Prolog-BEG ㇇⌅ѝⲴ㿲⛩Ǆ
x EBL ֌Ѫṧ ṧֻᕅሬⲴ⨶䇪䟽ᔪ˄example-guided reformulation of theories˅ǄProlog-
EBG ㇇⌅㻛ⴻ֌ᱟа⿽䟽ᔪ亶ฏ⨶䇪ࡠа⿽ਟ᫽֌ᖒᔿⲴᯩ⌅Ǆ⺞࠷ൠ䇢ˈ䟽ᔪ
亶ฏ⨶䇪ᱟ䙊䗷ࡋᔪ䘉ṧⲴ㿴ࡉ˖˄a˅㜭Ӿ亶ฏ⨶䇪ѝ╄㓾⍮⭏ˈԕ৺˄b˅൘а
њ᧘⨶↕޵࠶㊫㿲ሏࡠⲴ䇝㓳ṧֻǄ䘉ṧˈᆖҐࡠⲴ㿴ࡉਟ㻛ⴻ֌ሶ亶ฏ⨶䇪䟽
ᔪѪа㓴⢩↺ᛵߥлⲴ㿴ࡉˈᆳ㜭൘ањ᧘⨶↕޵ሩⴞḷᾲᘥⲴᇎֻ࠶㊫Ǆ
x EBL ֌ѪĀ Āӵӵā䟽䘠ᆖҐಘᐢ㓿Ā⸕䚃āⲴ˄“just” restating what the learner
already “knows”˅Ǆ൘Ḁ⿽〻ᓖкˈ൘ SafeToStack ֻᆀѝⲴᆖҐಘᔰ࿻Ҿަⴞḷ
ᾲᘥⲴ‫ޘ‬䜘⸕䇶Ǆҏቡᱟ䈤ˈྲ᷌ᆳⲴࡍ࿻亶ฏ⨶䇪‫࠶ݵ‬䀓䟺Ҷԫօ䇝㓳ṧֻˈ
䛓Ѹᆳҏ㜭‫࠶ݵ‬亴⍻ަ࠶㊫Ǆ䛓ѸᆖҐⲴ᜿ѹ൘ଚ‫ݯ‬઒˛а⿽എㆄᱟˈ൘䇨ཊԫ
࣑ѝˈ৏ࡉкᐢ⸕Ⲵ઼ᇎ䐥кਟᴹ᭸䇑㇇Ⲵѻ䰤Ⲵ४࡛ᖸབྷˈഐ↔䘉⿽Ā⸕䇶䟽
ᔪāѪᆖҐⲴ䟽㾱ᖒᔿǄֻྲ൘лỻⲴֻᆀѝˈሩᔸⲴ㿴ࡉᶴᡀҶањᆼ㖾Ⲵ亶
ฏ⨶䇪ˈ৏ࡉк䏣ԕ䘋㹼ᆼ㖾ⲴሩᔸǄণ֯ྲ↔ˈӪԜӽ❦䴰㾱བྷ䟿Ⲵ㓿傼ᶕᆖ
ҐྲօᖸྭൠлỻǄ䘉↓ᱟ䘉ṧа⿽ᛵᖒˈ˄Ӫ㊫Ⲵ˅ᆖҐಘᐢ㓿⸕䚃Ҷᆼ㖾Ⲵ
亶ฏ⨶䇪ˈ㘼䘋а↕ᆖҐਚᱟĀㆰঅൠāሶ↔⸕䇶䟽ᔪѪਖа⿽ᖒᔿˈԕ⭘Ҿᴤ
ᴹ᭸Ⲵᤷሬ䘲ᖃⲴ㹼ѪǄᴹ਼ṧ኎ᙗⲴਖањֻᆀᱟᆖҐ⢋亯࣋ᆖ䈮〻˖สᵜⲴ
⢙⨶ᇊᖻᐢ㻛ㆰঅൠ䱸䘠ˈնᆖ⭏ӽᰗ䴰㾱൘ᆖᵏѝ㣡аབྷ䜘࠶ᰦ䰤ᆖҐ䘉а䈮
〻ˈԕᤕᴹᴤਟ᫽֌ᖒᔿⲴ⸕䇶ˈ❦ਾቡн䴰㾱൘ᴰਾⲴ㘳䈅ѝ⭘ᴰสᵜⲴᇊᖻ
ᶕ᧘ሬ⇿њ䰞仈Ⲵ䀓ǄProlog-EBG ᢗ㹼Ⲵቡᱟ䘉⿽ᖒᔿⲴ⸕䇶䟽ᔪˈᆳᆖҐࡠⲴ
㿴ࡉਟӾਟ㿲ሏⲴᇎֻ⢩ᖱ᱐ሴࡠ‫ޣ‬ҾⴞḷᾲᘥⲴ࠶㊫ˈᯩ⌅ᱟ֯ަоสᵜ亶ฏ
⨶䇪а㠤Ǆ֯⭘৏࿻Ⲵ亶ฏ⨶䇪ਟ㜭䴰㾱䇨ཊ᧘⨶↕઼ᖸਟ㿲Ⲵᩌ㍒᡽㜭ሩԫ᜿
ᇎֻ࠶㊫ˈ㘼ᆖҐࡠⲴ㿴ࡉਟ൘ањ᧘⨶↕޵࠶㊫㿲ሏࡠⲴᇎֻǄ
ഐ↔ˈ㓟㋩Ⲵ EBL 㠤࣋Ҿ䟽ᔪ亶ฏ⨶䇪ԕӗ⭏ਟঅ↕᧘⨶ࠪṧֻ࠶㊫Ⲵа㡜㿴ࡉǄ䘉⿽

⸕䇶䟽ᔪⲴ䗷〻ᴹᰦ㻛〠Ѫ⸕䇶≷㕆˄knowledge compilation˅ˈ㺘⽪䘉⿽䖜ᦒᱟѪҶ໎࣐᭸
⦷ˈ㘼н᭩ਈ㌫㔏⸕䇶Ⲵ↓⺞ᙗǄ
11.3.1 ਇ⧦᯦⢯ᖷ
Prolog-EBG ањᴹ䏓Ⲵ㜭࣋ᱟᖒᡀ൘䇝㓳ṧֻⲴ᧿䘠ѝ⋑ᴹᱮᔿࠪ⧠Ⲵᯠ⢩ᖱˈն䘉Ӌ
⢩ᖱᱟ൘᧿䘠䇝㓳ṧֻѝⲴа㡜㿴ࡉᰦᗵ䴰ⲴǄ䘉⿽㜭࣋൘ࡽа㢲Ⲵ࠶↕㇇⌅઼ᆖࡠⲴ㿴ࡉѝ
ֻ⽪Ǆ⺞࠷ൠ䈤ˈᆖࡠⲴ㿴ࡉᯝ䀰ሩ x Ⲵ Volume ઼ Density Ⲵᗵ㾱㓖ᶏѪަ҈〟ሿҾ 5Ǆᇎ䱵
кˈ䇝㓳ṧֻᒦнवਜ਼↔҈〟ԕ৺ᆳᓄਆⲴ٬Ⲵ᧿䘠Ǆ↔㓖ᶏᱟ⭡ᆖҐಘ㠚ࣘᖒᡀⲴǄ
⌘᜿↔ᆖҐࡠⲴĀ⢩ᖱā㊫լҾ⭡⾎㓿㖁㔌Ⲵ䳀㯿অ‫ݳ‬㺘⽪Ⲵ⢩ᖱ㊫රǄҏቡᱟ䈤ˈ䘉њ
⢩ᖱᱟਟ⭡ᐢᴹᇎֻ኎ᙗ䇑㇇ࠪⲴǃབྷ䟿▌൘Ⲵ⢩ᖱѻаǄ઼ਾੁՐ᫝㇇⌅аṧˈProlog-EBG
൘ަቍ䈅ᤏਸ䇝㓳ᮠᦞⲴ䗷〻ѝˈ㠚ࣘᖒᡀ䘉ṧⲴ⢩ᖱǄ❦㘼ˈн‫⾎ۿ‬㓿㖁㔌ѝ֯⭘㔏䇑䗷〻
Ӿཊњ䇝㓳ṧֻѝ᧘ሬࠪ䳀㯿অ‫ݳ‬⢩ᖱˈProlog-EBG ᓄ⭘Ҷањ࠶᷀䗷〻สҾঅњ䇝㓳ṧֻ
Ⲵ ࠶ ᷀ ᧘ ሬ ᯠ Ⲵ ⢩ ᖱ Ǆ к 䶒 Ⲵ ֻ ᆀ ѝ Prolog-EBG ⭘ ࠶ ᷀ Ⲵ ᯩ ⌅ ᧘ ሬ ࠪ ⢩ ᖱ
Volume · Density>5 ˈᆳᶕ㠚Ҿ⭘Ҿ䀓䟺অњ䇝㓳ṧֻⲴ亶ฏ⨶䇪Ⲵ⢩ᇊᇎֻॆǄֻྲˈ
ĀVolumn ઼ Density Ⲵ҈〟ᖸ䟽㾱ā䘉аᾲᘥᱟᶕ㠚Ҿᇊѹ Weight Ⲵ亶ฏ⨶䇪㿴ࡉǄ䈕҈〟ᗵ
享ሿҾ 5 Ⲵᾲᘥᶕ㠚Ҿਖཆєᶑ亶ฏ⨶䇪㿴ࡉˈᆳԜᯝ䀰 Obj1 ᗵ享∄㥦ࠐ˄EndTable˅ᴤ䖫
˄Lighter˅ˈԕ৺㥦ࠐ˄Endtable˅Ⲵ䟽䟿˄Weight˅ㅹҾ 5Ǆഐ↔ˈ↓ᱟ䘉Ӌ亶ฏ⨶䇪ѝⲴ৏
ᆀ亩Ⲵ⢩ᇊਸᡀ઼ᇎֻॆ᡽ሬ㠤Ҷ↔ᯠ⢩ᖱⲴᇊѹǄ
㠚ࣘᆖҐᴹ⭘⢩ᖱԕᢙབྷᇎֻ㺘⽪Ⲵ䰞仈ᱟᵪಘᆖҐⲴањ䟽㾱䰞仈Ǆ൘สҾ䀓䟺ⲴᆖҐ
ѝ࠶᷀᧘ሬᯠ⢩ᖱˈ઼൘⾎㓿㖁㔌Ⲵ䳀㯿অ‫ݳ‬ѝᖂ㓣᧘ሬᯠ⢩ᖱᨀ‫׋‬Ҷє⿽н਼Ⲵ䙄ᖴǄഐ
↔ˈᆳԜ‫׍‬䎆Ⲵؑ᚟ᶕⓀн਼˄ањᱟ൘䇨ཊṧֻкⲴ㔏䇑㿴ࡉˈਖањᱟ֯⭘亶ฏ⨶䇪Ⲵঅ
њṧֻ࠶᷀˅ˈᴹਟ㜭㔃ਸє⿽ᶕⓀ᧒㍒ࠪᯠⲴᯩ⌅Ǆ
11.3.2 ╊㔄ᆜҖ
㓟㋩Ⲵ Prolog-EBG ᱟањ╄㓾Ⲵ㘼нᱟᖂ㓣ⲴᆖҐ䗷〻Ǆҏቡᱟ䈤ˈ䙊䗷䇑㇇䀓䟺Ⲵᴰ

ᕡࡽ‫ˈۿ‬ᆳӗ⭏ањਟӾ亶ฏ⨶䇪 B ѝ╄㓾⍮⭏Ⲵ‫ٷ‬䇮 hˈ㘼ф㾶ⴆ䇝㓳ᮠᦞ DǄᴤ㋮⺞ൠ
䇢ˈProlog-EBG 䗃ࠪањ‫ٷ‬䇮 h ┑䏣л䶒Ⲵ㓖ᶏ˖
(¢ xi , f ( xi )² D)(h xi ) ƅf(xi) ˄11ˊ1˅
DġBƅh ˄11ˊ2˅
ަѝ䇝㓳ᮠᦞD⭡а㓴䇝㓳ṧֻ㓴ᡀˈxiѪㅜiњ䇝㓳ᇎֻˈf(xi)ѪᆳⲴⴞḷ٬˄fѪⴞḷ࠭
ᮠ˅Ǆ⌘᜿ㅜањ㓖ᶏਚᱟㆰঅൠሶᵪಘᆖҐⲴ䙊ᑨⲴ䴰≲ᖒᔿॆˈণ‫ٷ‬䇮h㜭ሩ䇝㓳ᮠᦞѝ
⇿њᇎֻxi↓⺞亴⍻ⴞḷ٬f(xi)*Ǆᖃ❦а㡜ᛵߥлᴹཊ⿽‫ٷ‬䇮┑䏣䘉а㓖ᶏǄㅜҼњ㓖ᶏ᧿䘠
ҶProlog-EBLѝ亶ฏ⨶䇪Ⲵ֌⭘˖䗃ࠪ‫ٷ‬䇮㻛䘋а↕㓖ᶏԕ֯ަ⍮⭏Ҿ亶ฏ⨶䇪઼ᮠᦞǄ䘉ㅜ
Ҽњ㓖ᶏ߿ቁҶᆖҐಘ൘ᗵ享䘹ᤙ‫ٷ‬䇮ᰦ䶒ѤⲴ↗ѹᙗǄഐ↔ˈ亶ฏ⨶䇪Ⲵ֌⭘ᱟ߿ቁ‫ٷ‬䇮オ
䰤Ⲵᴹ᭸㿴⁑ᒦ䱽վᆖҐⲴṧᵜ༽ᵲᓖǄ
֯⭘⴨լⲴ䇠ਧˈᡁԜਟ᧿䘠ࠪ Prolog-EBG ᡰ䴰Ⲵ亶ฏ⨶䇪Ⲵ⸕䇶㊫රǄ⺞࠷ൠ䇢ˈ

Prolog-EBG ‫ٷ‬ᇊ亶ฏ⨶䇪 B ⏥㮤䇝㓳ᮠᦞѝᇎֻⲴ࠶㊫Ǆ
(¢ xi , f ( xi )² D)( B xi ) ƅf(xi) ˄11ˊ3˅
䘉њሩ亶ฏ⨶䇪 B Ⲵ㓖ᶏ‫؍‬䇱Ҷሩ⇿њ↓ֻਟᶴ䙐ࠪ䀓䟺Ǆ
ሶ Prolog-EBG ᆖҐ䰞仈઼ᖂ㓣䙫䗁㕆〻˄ㅜ 10 ㄐ˅ⲴᆖҐ䰞仈֌а∄䖳ᖸᴹ᜿ѹǄ൘ㅜ

10 ㄐᡁԜ䇘䇪Ҷа㡜ॆⲴᖂ㓣ᆖҐԫ࣑ˈަѝሩᆖҐಘᨀ‫׋‬Ҷ㛼Ჟ⸕䇶 B´ǄᡁԜ֯⭘ B´㘼н
ᱟ B ᶕԓ㺘 ILP ᡰ֯⭘Ⲵ㛼Ჟ⸕䇶ˈഐѪᆳа㡜н┑䏣ᔿ 11.3 Ⲵ㓖ᶏǄILP ᱟањᖂ㓣ᆖҐ㌫
㔏ˈ㘼 Prolog-EBG ᱟ╄㓾ᆖҐ㌫㔏ǄILP ֯⭘ަ㛼Ჟ⸕䇶 B´ᶕᢙབྷᖵ㘳㲁Ⲵ‫ٷ‬䇮䳶ਸˈ㘼
Prolog-EBG ֯⭘ަ亶ฏ⨶䇪 B ᶕ߿ሿਟ᧕ਇ‫ٷ‬䇮Ⲵ䳶ਸǄྲᔿ 10.2 㺘⽪ⲴˈILP ㌫㔏䗃ࠪⲴ h
┑䏣л䶒Ⲵ㓖ᶏ˖
(¢ xi , f ( xi )² D)( B' h xi ) ƅf(xi)
⌘᜿↔㺘䗮ᔿоProlog-EBGሩhӗ⭏Ⲵ㓖ᶏ˄⭡ᔿ 11.2 ઼ 11.3 㔉ࠪ˅ѻ䰤Ⲵ㚄㌫Ǆ䘉њ൘

hкⲴILP㓖ᶏᱟᔿ 11.1 ѝ㓖ᶏⲴᕡॆᖒᔿǄILP㓖ᶏਚ㾱≲ ( B' h xi ) ƅf(xi)ˈ㘼Prolog-EBG
㾱≲ᴤѕṬⲴ (h xi ) ƅf(xi)Ǆ䘈㾱⌘᜿ILPѝ⋑ᴹሩᓄᔿ 11.2 ѝProlog-EBG㓖ᶏǄ
*
䘉䟼൘⏥㮤˄ƅ˅Ⲵᇊѹѝवਜ਼ҶPrologṧᔿⲴཡ䍕੖ᇊˈഐ↔ྲ᷌ṧֻн㜭㻛䇱᰾Ѫ↓ֻˈࡉᆳԜ㻛
⏥㮤Ѫ৽ֻǄ
11.3.3 ะӄ䀙䠀ⲺᆜҖⲺᖈ㓩‫څ‬㖤
എᗶㅜ 2 ㄐⲴਉ䘠ˈањᆖҐ㇇⌅Ⲵᖂ㓣‫ٿ‬㖞Ѫа㓴ᯝ䀰ˈᆳԜо䇝㓳ṧֻа䎧╄㓾⏥㮤
ᆖҐಘⲴਾ㔝亴⍻Ǆᖂ㓣‫ٿ‬㖞Ⲵ䟽㾱ᙗ൘Ҿᆳ࡫⭫ࠪᆖҐಘᱟᘾṧӾ㿲ሏࡠⲴ䇝㓳ṧֻ⌋ॆ
ⲴǄ
Prolog-EBG Ⲵᖂ㓣‫ٿ‬㖞ᱟӰѸ˛൘ Prolog-EBG ѝˈྲᔿ 11.2 ᡰ᧿䘠Ⲵˈ䗃ࠪⲴ‫ٷ‬䇮 h Ӿ

DġB ѝ╄㓾⍮⭏Ǆഐ↔亶ฏ⨶䇪 B Ѫа㓴ᯝ䀰ˈᆳԜо䇝㓳ṧֻа䎧⏥㮤䗃ࠪ‫ٷ‬䇮Ǆ⭡Ҿᆖ
ҐಘⲴ亴⍻Ӿ↔‫ٷ‬䇮 h ѝ⍮⭏ˈլѾ Prolog-EBG Ⲵᖂ㓣‫ٿ‬㖞ቡᱟ䗃‫ޕ‬ᆖҐಘѝⲴ亶ฏ⨶䇪
BǄᇎ䱵кਟԕ䘉ṧ䇔ᇊˈ䲔Ҷ䘈䴰㘳㲁ਖཆањ㓶㢲˖亶ฏ⨶䇪ਟ⏥㮤ཊњਟ䘹Ⲵ Horn ᆀਕ
䳶Ǆഐ↔ˈᖂ㓣‫ٿ‬㖞䘈䴰वਜ਼ Prolog-EBG ൘䘉Ӌਟ䘹Ⲵ Horn ᆀਕ䳶ѝ֌ࠪ䘹ᤙ䘉䜘࠶޵ᇩǄ
ྲк䶒ᡰ㿱ࡠⲴˈProlog-EBG ֯⭘ᒿࡇ㾶ⴆ㇇⌅нᯝᖒᡀ䱴࣐Ⲵ Horn ᆀਕⴤࡠᡰᴹⲴ↓ֻ㻛
㾶ⴆǄᴤ䘋а↕ˈ⇿њঅ⤜Ⲵ Horn ᆀਕᱟᖃࡽ䇝㓳ṧֻⲴ䀓䟺ᡰ䇨ਟⲴᴰа㡜ᆀਕ˄ণᴰᕡ
ࡽ‫˅ۿ‬Ǆഐ↔൘亶ฏ⨶䇪⏥㮤Ⲵ਴ Horn ᆀਕ䳶ѻѝˈᡁԜਟԕሶ Prolog-EBG Ⲵ‫ٿ‬㖞࡫⭫Ѫሩ
ᶱབྷа㡜ॆ Horn ᆀਕⲴሿ䳶ਸⲴ‫ྭٿ‬Ǆᇎ䱵к Prolog-EBG Ⲵ䍚ႚ㇇⌅ਚᱟѪራ᢮ᶱབྷа㡜ॆ
Horn ᆀਕⲴⵏ↓ᴰ⸝䳶ਸᡰ䴰Ⲵᖫᓅᩌ㍒㇇⌅Ⲵањ੟ਁᔿⲴ䘁լǄᰐ䇪ᘾṧˈProlog-EBG
ᖂ㓣‫ٿ‬㖞ӽਟ⭘䘉⿽ᯩᔿ䘁լ࡫⭫Ǆ
䘁լⲴ Prolog-EBG ᖂ㓣‫ٿ‬㖞˖亶ฏ⨶䇪 Bˈ࣐кሩᶱབྷа㡜ॆ Horn ᆀਕⲴሿ䳶ਸⲴ‫ྭٿ‬Ǆ
䘉䟼ᴰ䟽㾱Ⲵ㾱⛩൘ҾˈProlog-EBG Ⲵᖂ㓣‫ٿ‬㖞˄ণᆳӾ䇝㓳ᮠᦞѝ⌋ॆⲴㆆ⮕˅൘ᖸ
བྷ〻ᓖк⭡䗃‫Ⲵޕ‬亶ฏ⨶䇪⺞ᇊǄᆳоᡁԜᡰ䇘䇪䗷ⲴཊᮠᆖҐ㇇⌅ᆼ‫ޘ‬н਼ǄཊᮠᆖҐ㇇⌅
˄ྲ⾎㓿㖁㔌ˈߣㆆṁᆖҐ˅ѝᖂ㓣‫ٿ‬㖞ᱟᆖҐ㇇⌅Ⲵањപᇊ኎ᙗˈа㡜ᱟ⭡ަ‫ٷ‬䇮㺘⽪Ⲵ
䈝⌅ᡰ⺞ᇊⲴǄѪӰѸᢺᖂ㓣‫ٿ‬㖞֌Ѫањ䗃‫ޕ‬৲ᮠ㘼нᱟᆖҐಘⲴപᇊ኎ᙗॱ࠶䟽㾱˛䘉ᱟ
ഐѪˈྲᡁԜ൘ㅜ 2 ㄐ৺ަԆൠᯩ䇘䇪䗷Ⲵˈнᆈ൘ањ‫ޘ‬ተᴹ᭸Ⲵᖂ㓣‫ٿ‬㖞ˈ㘼фᰐ‫ٿ‬ᆖҐ
ᱟᰐ⭘ⲴǄഐ↔ԫօᔰਁ䙊⭘ᆖҐᯩ⌅Ⲵቍ䈅ˈ䜭㠣ቁՊ‫ݱ‬䇨ᖂ㓣‫ٿ‬㖞㜭ཏ䪸ሩᖵ䀓ߣⲴᆖҐ
䰞仈ᴹᡰн਼Ǆ൘ањᴤᇎ䐥ᙗⲴቲ⅑кˈ䇨ཊᆖҐԫ࣑ᴤᐼᵋ䗃‫ޕ‬亶ฏ⢩ᇊⲴ⸕䇶˄ྲ
SafeToStack ֻᆀѝⲴᴹ‫ ޣ‬Weight Ⲵ⸕䇶˅ԕᖡ૽ᆖҐಘӾ䇝㓳ᮠᦞѝ⌋ॆⲴᯩ⌅Ǆ⴨৽ˈ䙊
䗷䲀ࡦ‫ٷ‬䇮Ⲵ䈝⌅ᖒᔿ˄ྲ‫ߣ⸝ྭٿ‬ㆆṁ˅ᶕĀᇎ⧠āḀ䘲ᖃⲴ‫ٿ‬㖞ᙗࡉнཚ㠚❦Ǆᴰਾˈྲ
᷌㘳㲁ањᴤབྷⲴ䰞仈ˈањ㠚⋫ agent ྲօ䲿⵰ᰦ䰤᭩䘋ᆳⲴᆖҐ㜭࣋ˈ䛓Ѹᴰྭᱟᴹањ
ᆖҐ㇇⌅ˈᆳⲴ⌋ॆ㜭࣋ਟ൘ަ㧧ᗇࡠᴤཊⲴ亶ฏ⸕䇶ਾ໎ᕪǄ
11.3.4 ⸛䇼ቸ⅗ⲺᆜҖ
ྲᔿ 11.2 ᤷࠪⲴˈ⭡ Prolog-EBG 䗃ࠪⲴ‫ٷ‬䇮 h Ӿ亶ฏ⨶䇪 B ઼䇝㓳ᮠᦞ D ѝ╄㓾⍮⭏Ǆ

ᇎ䱵кˈ䙊䗷㘳ḕ Prolog-EBG ㇇⌅ˈᖸᇩ᱃ⴻࠪ h ⴤ᧕Ӿঅ⤜Ⲵ B ѝ⍮⭏ˈ㘼о D ᰐ‫ޣ‬ǄѪ
Ҷ⨶䀓䘉а⛩ˈᡁԜਟԕ‫ٷ‬ᜣᴹањ〠Ѫᶑⴞ᷊Ѯಘ˄Lemma-enumerator˅Ⲵ㇇⌅Ǆ䘉њ㇇⌅
สҾ亶ฏ⨶䇪 B ѝⲴᯝ䀰ㆰঅൠ᷊Ѯ㜭ᗇࡠⴞḷᾲᘥⲴᡰᴹ䇱᰾ṁǄሩ⇿њ䇱᰾ṁˈLemma-
enumerator ⭘о Prolog-EBG ⴨լⲴᯩ⌅䇑㇇ᴰᕡࡽ‫ۿ‬ᒦᶴ䙐ањ Horn ᆀਕǄ൘ Lemma-
enumerator ઼ Prolog-EBG ѻ䰤ᜏаⲴн਼ᱟˈLemma-enumerator ᘭ⮕䇝㓳ᮠᦞᒦ᷊Ѯࠪᡰᴹ
Ⲵ䇱᰾ṁǄ
⌘᜿ Lemma-enumerator 䗃ࠪⲴᱟ Prolog-EBG 䗃ࠪ Horn ᆀਕⲴ䎵䳶Ǆ⭡Ҿ䘉а⛩ˈӗ⭏

Ҷࠐњ䰞仈Ǆ俆‫ྲˈݸ‬᷌ᆳⲴ‫ٷ‬䇮অ⤜Ӿ亶ฏ⨶䇪ѝ⍮⭏ˈ䛓Ѹ Prolog-EBG ѝ䇝㓳ᮠᦞᴹӰ
Ѹ֌⭘˛ㆄṸ൘Ҿˈ䇝㓳ṧֻ֯ Prolog-EBG ‫⌘ޣ‬Ҿ⭏ᡀ㿴ࡉԕ㾶ⴆᇎ䱵ࠪ⧠ⲴṧֻⲴ࠶ᐳǄ
ֻྲˈ൘৏ᶕⲴлỻֻᆀѝˈᡰᴹਟ㜭Ⲵᶑⴞᮠᖸབྷˈ㘼൘䙊ᑨሩᔸѝࠪ⧠Ⲵỻⴈ⣦ᘱਚᱟ䈝
⌅кਟ㜭ࠪ⧠Ⲵỻⴈ⣦ᘱⲴаሿ䜘࠶Ǆഐ↔ˈ䙊䗷ਚ‫⌘ޣ‬ᇎ䱵кՊ䙷ࡠⲴ䇝㓳ṧֻˈ∄ቍ䈅᷊
ѮỻⴈⲴᡰᴹਟ㜭ᶑⴞˈ〻ᒿᴤਟ㜭ᗇࡠᴤሿⲴǃᴤ⴨‫Ⲵޣ‬㿴ࡉ䳶Ǆ
ӗ⭏ⲴㅜҼњ䰞仈ᱟˈProlog-EBG ᱟ੖㜭ᆖҐࡠањ䎵ࠪ䳀ਜ਼൘亶ฏ⨶䇪ѝⲴ⸕䇶Ⲵ‫ٷ‬
䇮˛ᦒ䀰ѻˈᆳᱟ੖㜭ᆖҐࡠањᇎֻⲴ࠶㊫ˈ䘉њᇎֻн㜭⭘৏࿻Ⲵ亶ฏ⨶䇪䘋㹼࠶㊫˄‫ٷ‬
ᇊᇊ⨶䇱᰾ಘᴹᰐ䲀Ⲵ䇑㇇䍴Ⓚ˅˛нᒨⲴᱟˈᆳн㜭‫ࡠڊ‬Ǆྲ᷌ Bƅhˈ䛓Ѹԫօ⭡ h ⏥㮤
Ⲵ࠶㊫ҏሶ⭡ B ⏥㮤Ǆ䘉ᱟ੖ᱟ࠶᷀ᆖҐᡆ╄㓾ᆖҐⲴപᴹ㕪䲧˛ᒦ䶎ྲ↔ˈྲлֻᡰ⽪Ǆ
ѪҶѮањ╄㓾ᆖҐⲴֻᆀˈަѝᆖҐࡠⲴ‫ٷ‬䇮 h ਟ⏥㮤ࠪ B н㜭⏥㮤Ⲵ㔃䇪ˈᡁԜᗵ享

ࡋᔪањ Bƅ/ h ն DġBƅh Ⲵֻᆀ˄എᗶᔿ 11.2 㔉ࠪⲴ㓖ᶏ˅Ǆਟԕ㘳㲁 B वਜ਼䘉ṧⲴᯝ
䀰˖Ā㤕 x ┑䏣ⴞḷᾲᘥˈ䛓Ѹ g(x)ҏ┑䏣Ǆāঅ⤜䘉њᯝ䀰н㜭⏥㮤ࠪԫօᇎֻⲴ࠶㊫Ǆ❦
㘼ˈаնᡁԜ㿲ሏࡠа↓ֻˈᆳ‫ݱ‬䇨╄㓾⌋ॆࡠަԆᵚ㿱ᇎֻǄֻྲˈ㘳㲁ᆖҐ PlayTennis Ⲵ
ⴞḷᾲᘥˈᆳ᧿䘠Ҷ Ross ᐼᵋᢃ㖁⨳ⲴᰕᆀǄ‫⇿ྲٷ‬њᰕᆀਚ㻛᧿䘠Ѫঅњ኎ᙗ Humidityˈ
ᒦф亶ฏ⨶䇪वਜ਼অњᯝ䀰Āྲ᷌ Ross ௌ⅒൘⒯ᓖ˄Humidity˅Ѫ x Ⲵᰕᆀᢃ㖁⨳ˈ䛓ѸԆҏ
ௌ⅒൘⒯ᓖሿҾ x Ⲵᰕᆀᢃ㖁⨳āˈਟ㻛ᖒᔿॆൠ᧿䘠Ѫ˖
(x ) IF((PlayTennis=Yes)ĕ(Humidity=x))
THEN ((PlayTennis=Yes) ĕ(Humidityİx))
⌘᜿↔亶ฏ⨶䇪нՊሩ PlayTennis ⲴᇎֻѝଚӋᱟ↓ֻˈଚӋᱟ䍏ֻ⏥㮤ࠪԫօ㔃䇪Ǆ❦

㘼ˈаնᆖҐಘ㿲ሏࡠањ↓ֻѝ Humidity=0.3ˈ亶ฏ⨶䇪䘎਼↔↓ֻа䎧⏥㮤ࡠл䶒Ⲵа㡜
‫ٷ‬䇮 h˖
(PlayTennis=Yes) ĕ(Humidityİ0.30)
ᾲᤜ䎧ᶕˈ↔ֻᆀ᧿䘠Ҷа⿽ᛵᖒަѝ Bƅ/ hˈն BġDƅhǄ䘉䟼ᆖࡠⲴ‫ٷ‬䇮⏥㮤Ⲵ亴⍻

н㜭㻛অ⤜Ⲵ亶ฏ⨶䇪⏥㮤Ǆᵟ䈝Ā⸕䇶ቲ⅑ⲴᆖҐā˄knowledge-level learning˅ᴹᰦ㻛⭘
Ҿ〠䘉⿽㊫රⲴᆖҐˈަѝᆖҐࡠⲴ‫ٷ‬䇮⏥㮤Ⲵ亴⍻䎵ࠪҶ㜭㻛亶ฏ⨶䇪⏥㮤Ⲵ㤳തǄ⭡ᯝ䀰
䳶ਸ Y ⏥㮤Ⲵᡰᴹ亴⍻Ⲵ䳶ਸᑨ〠Ѫ Y Ⲵ╄㓾䰝व˄deductive closure˅Ǆ䘉䟼Ⲵ‫ޣ‬䭞४࡛൘
Ҿˈ⸕䇶ቲ⅑ⲴᆖҐѝ B Ⲵ╄㓾䰝वᱟ Bˇh ╄㓾䰝वⲴⵏᆀ䳶Ǆ
⸕䇶ቲ⅑Ⲵ࠶᷀ᆖҐⲴਖањֻᆀᱟˈ㘳㲁а⿽㊫රⲴᯝ䀰ˈ䙊ᑨ〠Ѫ determinationˈᆳ
Ⲵ㓶㢲⹄ウ㿱 Russel˄1989˅ԕ৺ަԆаӋᐕ֌ǄDetermination ᯝ䀰ˈᇎֻⲴḀ኎ᙗᆼ‫ޘ‬ਆߣ
ҾḀӋ⢩ᇊ኎ᙗˈնнᗵᤷ᰾䘉⿽‫׍‬䎆ᙗⲴ⺞࠷ᙗ䍘Ǆֻྲˈ㘳㲁ᆖҐањⴞḷᾲᘥĀ䈤㪑㨴
⢉䈝ⲴӪāˈᒦф‫ٷ‬ᇊ亶ฏ⨶䇪Ѫঅњ determination ᯝ䀰ĀḀӪ䈤Ⲵ䈝䀰⭡ԆⲴഭ㉽ߣᇊǄā
ਚᴹ䘉ᶑ亶ฏ⨶䇪ˈн㜭ཏ⭘ᶕ࠶㊫↓઼ֻ৽ֻǄ❦㘼ˈྲ᷌ᡁԜ㿲ሏࡠĀJoeˈ23 ኱ˈᐖ᪷
ᆀˈᐤ㾯Ӫˈ䈤㪑㨴⢉䈝āˈ䛓ѸᡁԜቡਟԕ↔↓઼ֻ亶ฏ⨶䇪ѝᗇࡠ˖ĀᡰᴹⲴᐤ㾯Ӫ䜭䈤
㪑㨴⢉䈝āǄ
䘉Ӌֻᆀ䜭╄⽪Ҷ࠶᷀ᆖҐྲօӗ⭏н㜭⭡亶ฏ⨶䇪অ⤜⏥㮤Ⲵ‫ٷ‬䇮ǄަѝⲴ䗃ࠪ‫ٷ‬䇮 h
䜭┑䏣 DġBƅhˈնн┑䏣 BƅhǄ൘є⿽ᛵߥлˈᆖҐಘ䜭╄㓾˄deduce˅ࠪањਸ⨶Ⲵ‫ٷ‬
䇮ˈᆳᰒн㜭Ӿ亶ฏ⨶䇪ѝঅ⤜⍮⭏ˈҏн㜭Ӿ䇝㓳ᮠᦞѝঅ⤜⍮⭏Ǆ
11.4 ᩒ㍘᧝࡬⸛䇼Ⲻะӄ䀙䠀ⲺᆜҖ
ྲк䘠ᤷࠪⲴˈProlog-EBG ㇇⌅Ⲵᇎ䱵㜭࣋ਇ亶ฏ⨶䇪ᗵ享↓⺞фᆼᮤ䘉а㾱≲ᡰ䲀
ࡦǄ㜭ཏ┑䏣䘉а㾱≲ⲴᆖҐ䰞仈Ⲵањ䟽㾱㊫࡛Ѫ䙊䗷ᆖҐ֯༽ᵲⲴᩌ㍒〻ᒿ䙏ᓖ࣐ᘛǄᇎ
䱵кˈᓄ⭘สҾ䀓䟺ⲴᆖҐⲴᴰབྷ㿴⁑Ⲵቍ䈅ᐢ㓿ᔰ࿻䀓ߣᆖҐ᧗ࡦᩌ㍒Ⲵ䰞仈ˈᆳᴹᰦ৸㻛
〠ѪĀ࣐䙏ā˄speedup˅ᆖҐǄֻྲˈ‫ۿ‬ỻ㊫䘉ṧⲴሩᔸѝˈሩਸ⌅ᩌ㍒᫽֌Ⲵᇊѹԕ৺ᩌ
㍒ⴞḷⲴᇊѹᨀ‫׋‬ҶᆖҐᩌ㍒᧗ࡦ⸕䇶Ⲵањᆼᮤф↓⺞Ⲵ亶ฏ⨶䇪Ǆ
ྲօ⺞࠷ൠᇊѹᆖҐᩌ㍒᧗ࡦ䰞仈Ⲵᖒᔿԕ֯⭘สҾ䀓䟺ⲴᆖҐ˛㘳㲁ања㡜ᩌ㍒䰞
仈ˈަѝSѪਟ㜭ᩌ㍒⣦ᘱⲴ䳶ਸˈOѪਸ⌅ᩌ㍒㇇ᆀⲴ䳶ਸˈᆳሶа⿽ᩌ㍒⣦ᘱ䖜ᦒᡀਖа
⿽ᩌ㍒⣦ᘱˈ㘼фGѪ൘SкᇊѹⲴ䉃䇽ˈᆳ㺘⽪ଚ⿽⣦ᘱѪⴞⲴ⣦ᘱǄ䰞仈а㡜ᱟራ᢮а㌫
ࡇⲴ㇇ᆀˈᆳሶԫ᜿ࡍ࿻⣦ᘱsi䖜ॆѪḀᴰ㓸⣦ᘱsfˈ֯ⴞⲴ䉃䇽Gᗇࡠ┑䏣ǄᇊѹᆖҐ䰞仈ᖒ
ᔿⲴа⿽࣎⌅ᱟ䇙㌫㔏ሩOѝ⇿њ㇇ᆀᆖҐањ࠶・ⲴⴞḷᾲᘥǄ⺞࠷ൠ䇢ˈሩOѝ⇿њ㇇ᆀ
oˈᆳਟቍ䈅ᆖҐⴞḷᾲᘥĀ㜭⭘Oሬ㠤ⴞⲴ⣦ᘱⲴ⣦ᘱ䳶ਸāǄᖃ❦ウㄏ䘹ᤙଚањ֌Ѫᖵ
ᆖҐⲴⴞḷ⣦ᘱˈ‫׍‬䎆Ҿᗵ享֯⭘↔ᆖҐࡠⲴ⸕䇶Ⲵ䰞仈≲䀓ಘⲴ޵䜘㔃ᶴˈֻྲˈྲ᷌䰞仈
≲䀓ಘᱟањmeans-ends㿴ࡂ㌫㔏ˈᆳⲴᐕ֌䗷〻ᱟ䙊䗷ᔪ・઼䀓ߣᆀⴞⲴˈ䛓ѸᡁԜᐼᵋᆖ
ҐⲴⴞḷᾲᘥਟԕᱟĀA㊫රⲴᆀⴞⲴᗵ享൘B㊫රⲴᆀⴞⲴѻࡽ䀓ߣⲴ㿴ࡂ⣦ᘱ䳶ਸǄā
֯⭘สҾ䀓䟺ⲴᆖҐԕ᭩䘋ަᩌ㍒Ⲵањ㌫㔏ᱟProdigy˄Carbonell et al. 1990˅Ǆ Prodigy

ᱟањ亶ฏᰐ‫Ⲵޣ‬㿴ࡂ㌫㔏ˈᆳ᧕ਇԕ⣦ᘱオ䰤S઼㇇ᆀ䳶ਸOᇊѹⲴ䰞仈亶ฏǄ❦ਾᆳ䀓ߣ
䘉ṧᖒᔿⲴ䰞仈˖Āራ᢮ањ㇇ᆀᒿࡇ֯ࡍ࿻⣦ᘱsi䖜ᦒࡠ┑䏣ⴞⲴ䉃䇽GⲴ⣦ᘱǄā Prodigy
֯⭘ањmeans-ends㿴ࡂಘሶ䰞仈࠶䀓ѪᆀⴞⲴˈ䀓ߣ䘉ӋᆀⴞⲴˈ❦ਾਸᒦ䎧ᶕᡀѪᮤњ䰞
仈Ⲵ䀓Ǆ䘉ṧˈ൘ަᩌ㍒䰞仈䀓Ⲵ䗷〻ѝ Prodigy 䟽༽䶒Ѥ䘉ṧⲴ䰞仈˖Āла↕㾱䀓ߣⲴᱟ
ଚњᆀⴞⲴ˛āԕ৺ĀѪ䀓ߣ↔ᆀⴞⲴ㾱⭘ଚњ᫽֌˛Ǆā Minton˄1988˅᧿䘠ҶሶสҾ䀓
䟺ⲴᆖҐ䳶ᡀࡠ ProdigyⲴ䗷〻ˈᯩ⌅ᱟᇊѹа㓴䘲ਸҾ䘉⿽нᯝ䙷ࡠⲴ᧗ࡦߣㆆⲴⴞḷᾲ
ᘥǄֻྲˈањⴞḷᾲᘥᱟĀᆀⴞḷAᗵ享൘ᆀⴞḷBѻࡽ䀓ߣⲴ⣦ᘱ䳶ਸǄāሩ䘉њⴞḷᾲ
ᘥˈ⭡ProdigyᆖࡠⲴ㿴ࡉ൘ㆰঅⲴ⢙փึਐ䰞仈ѝⲴањֻᆀѪ˖
IF ᖵ䀓ߣⲴᆀⴞḷѻаѪ On(x,y)ˈᒦф
ᖵ䀓ߣⲴᆀⴞḷѻаѪ On(y,z)
THEN ൘ On(x,y)ѻࡽ䀓ߣ On(y,z)
Ѫ⨶䀓↔㿴ࡉˈ޽⅑㘳㲁മ 9-3 ѝ⽪ֻⲴㆰঅⲴඇ⣦⢙փึਐ䰞仈Ǆ൘മ⽪Ⲵ䰞仈ѝˈⴞ

Ⲵᱟሶ⢙ඇึਐᡀѪঅ䇽 universalǄProdigy ሶᢺ↔䰞仈࠶䀓Ѫࠐњ㾱䗮ࡠⲴᆀⴞⲴˈवᤜ
ON(U, N)ˈON(N, I)ㅹǄ⌘᜿к䶒Ⲵ㿴ࡉ३䝽ᆀⴞⲴ ON(U, N)઼ ON(N, I)ˈᒦфᔪ䇞൘䀓ߣᆀ
䰞仈 ON(U,N)ѻࡽ䀓ߣ ON(N, I)Ǆ↔㿴ࡉⲴ⨶⭡˄ԕ৺ Prodigy ⭘ҾᆖҐ↔㿴ࡉⲴ䀓䟺˅൘Ҿ
ྲ᷌ᡁԜԕ䘶ᒿ䀓ߣ䘉єњᆀⴞⲴˈሶՊ䙷ࡠߢケˈӾ㘼ᗵ享᫔⎸ ON(U, N)Ⲵ䀓ԕ䗮ࡠਖа
ᆀⴞⲴ ON(N, I)ǄProdigy ᆖҐ䗷〻俆‫ݸ‬䙷ࡠ䘉ṧањߢケˈ❦ਾ㠚ᡁ䀓䟺ߢケⲴ৏ഐˈᒦࡋ
ᔪањ㊫լҾԕкⲴ㿴ࡉǄަ᭸᷌൘Ҿ Prodigy ֯⭘‫ޣ‬Ҿਟ㜭ⲴᆀⴞⲴߢケⲴ亶ฏᰐ‫⸕Ⲵޣ‬
䇶ˈԕ৺‫ޣ‬Ҿ⢩ᇊ᫽֌Ⲵ亶ฏ⢩ᇊⲴ⸕䇶˄ྲᵪಘӪਚ㜭а⅑Ѯ䎧ањ⢙ඇ˅ˈԕᆖҐࡠᴹ⭘
Ⲵ亶ฏ⢩ᇊⲴ㿴ࡂ㿴ࡉˈྲк䶒ֻ⽪Ⲵ䛓њ㿴ࡉǄ
֯⭘สҾ䀓䟺ⲴᆖҐԕ㧧ਆ Prodigy Ⲵ᧗ࡦ⸕䇶ˈᐢ㓿൘н਼Ⲵ䰞仈亶ฏѝֻ⽪Ǆवᤜк

䶒ㆰঅⲴ⢙ඇึਐ䰞仈ˈԕ৺ަԆᴤ༽ᵲⲴ䈳ᓖ઼㿴ࡂ䰞仈 ǄMinton˄1988˅ᣕ੺Ҷ 3 њ亶
ฏѝⲴᇎ傼ˈަѝᆖҐࡠⲴ᧗ࡦ㿴ࡉᢺ䰞仈≲䀓Ⲵ᭸⦷ᨀ儈Ҷ 2 ࡠ 4 ‫ؽ‬Ǆᴤ䘋а↕ˈ䘉Ӌᆖࡠ
Ⲵ㿴ࡉⲴᙗ㜭൘䘉 3 њ䰞仈ѝо᡻߉㿴ࡉᴹਟ∄ᙗǄMinton ҏ᧿䘠ҶሩสᵜⲴสҾ䀓䟺ᆖҐⲴ
㤕ᒢᢙኅˈᆳԜᨀ儈ҶᆖҐ᧗ࡦ⸕䇶Ⲵ᭸⦷Ǆᯩ⌅वᤜㆰॆᆖҐࡠⲴ㿴ࡉԕ৺৫䲔䛓Ӌ᭦⳺ሿ
Ҿᔰ䬰Ⲵ㿴ࡉǄ
ਖањ㔃ਸҶḀ⿽ᖒᔿⲴสҾ䀓䟺ᆖҐⲴа㡜䰞仈≲䀓ṶᷦѪ Soar ㌫㔏˄Laird et al.

1986˗Newell 1990˅ǄSoar ᭟ᤱ㤳ത䖳ᇭⲴ䰞仈≲䀓ㆆ⮕ˈवਜ਼Ҷ Prodigy Ⲵ means-ends 㿴ࡂ
ㆆ⮕൘޵Ǆ❦㘼ˈ‫ ۿ‬Prodigy аṧˈSoar ѝⲴᆖҐᱟ䙊䗷䀓䟺ᖃࡽⲴ᧗ࡦㆆ⮕ѪӰѸሬ㠤վ
᭸Ǆᖃᆳ䙷ࡠањᩌ㍒䘹ᤙˈަѝ⋑ᴹањ⺞ᇊᰐ⯁ⲴㆄṸᰦ˄ྲла↕䈕ᓄ⭘ଚањ᫽֌
ㅖ˅ˈSoar ᙍ㘳䘉њᩌ㍒‫ܥ‬ተˈ֯⭘ྲ⭏ᡀ޽⍻䈅䘉⿽ᕡॆⲴᯩ⌅ᶕߣᇊ↓⺞Ⲵ㹼ࣘᯩੁǄ⭘
ᶕ䀓ߣ䘉⿽‫ܥ‬ተⲴ᧘⨶ਟ㻛⨶䀓Ѫሩሶᶕᘾṧ䀓ߣ㊫լ‫ܥ‬ተⲴ䀓䟺ǄSoar ֯⭘ਖа⿽н਼Ⲵส
Ҿ䀓䟺ᆖҐ〠Ѫ chunkingˈԕᣭਆࠪਟᓄ⭘⴨਼Ⲵ䀓䟺Ⲵа㡜ᶑԦǄSoar ᐢ㻛ᓄ⭘ҾཊᮠⲴ䰞
仈亶ฏˈᒦ㻛ᨀ䇞ѪӪ㊫ᆖҐ䗷〻ѝа⿽ᗳ⨶ᆖкਟ㹼Ⲵ⁑ර˄㿱 Newell 1990 ˅Ǆ
Prodigy ઼ Soar ╄⽪ҶสҾ䀓䟺ⲴᆖҐᯩ⌅ਟ㻛ᡀ࣏ᓄ⭘Ҿ൘н਼䰞仈亶ฏѝ㧧ਆᩌ㍒᧗

ࡦ⸕䇶Ǆ❦㘼ˈབྷཊᮠ੟ਁᔿᩌ㍒〻ᒿӽ❦֯⭘㊫լҾㅜ 1 ㄐ᧿䘠Ⲵᮠ٬䇴ՠ࠭ᮠˈ㘼нᱟ⭡
สҾ䀓䟺ⲴᆖҐ㧧ਆⲴ㿴ࡉǄ৏ഐᱟӰѸ˛ᇎ䱵кᴹаӋ䟽㾱Ⲵᇎ䐥䰞仈ᓄ⭘Ҷ EBL ᆖҐᩌ
㍒᧗ࡦǄ俆‫ˈݸ‬൘䇨ཊᛵߥлᗵ享ᆖҐⲴ᧗ࡦ㿴ࡉⲴᮠⴞ䶎ᑨབྷ˄ྲᮠॳњ㿴ࡉ˅Ǆᖃ㌫㔏ᆖ
Ґࡠ䎺ᶕ䎺ཊⲴ᧗ࡦ㿴ࡉԕ᭩䘋ᩌ㍒ˈ㾱㣡৫䎺ᶕ䎺བྷⲴᔰ䬰൘⇿↕ѝ३䝽䘉㓴㿴ࡉࡠᖃࡽᩌ
㍒⣦ᘱѝǄ⌘᜿䘉њ䰞仈ᒦнਚተ䲀ҾสҾ䀓䟺ⲴᆖҐˈᆳ൘⭘໎䮯Ⲵ㿴ࡉ䳶㺘⽪ަᆖࡠ⸕䇶
Ⲵԫ᜿㌫㔏ѝ䜭Պࠪ⧠Ǆᴹ᭸Ⲵ३䝽㿴ࡉ㇇⌅ਟ㕃઼䘉а䰞仈ˈնн㜭ᆼ‫⎸ޘ‬䲔ᆳǄMinton
˄1988˅䇘䇪Ҷ㓿傼ൠՠ䇑⇿њ㿴ࡉⲴ䇑㇇ᔰ䬰઼᭦⳺Ⲵㆆ⮕ˈਚ൘ՠ䇑Ⲵ᭦⳺䎵䗷ՠ䇑Ⲵᔰ
䬰ᰦ᡽ᆖҐ䘉Ӌ㿴ࡉˈᒦ൘ḀӋ㿴ࡉᴹ䍏᭸⭘ᰦࡐ䲔ᆳԜǄԆ᧿䘠Ҷྲօ֯⭘䘉⿽᭸⭘࠶᷀
˄utility analysis˅ᶕ⺞ᇊଚӋᓄ㻛ᆖҐଚӋ䈕ᘈ䇠ˈᖸབྷൠ໎ᕪҶ Prodigy ѝสҾ䀓䟺ᆖҐⲴ
ᴹ᭸ᙗǄֻྲˈ൘а㌫ࡇᵪಘӪⲴ⢙ඇึਐ䰞仈ѝˈProdigy 䙷ࡠҶ 328 њᵪՊਟᆖҐањᯠ
㿴ࡉˈնਚ࡙⭘Ҷަѝ 69 њˈᒦфᴰ㓸৫䲔Ҷվ᭸⭘Ⲵ㿴ࡉਾ࢙։ 19 њ㿴ࡉǄ Tambe et al.
˄1990˅઼ Doorenbos˄1993˅䇘䇪Ҷᘾṧ⺞ᇊ㿴ࡉѝ३䝽ᔰ䬰⢩࡛བྷⲴ㊫රˈᒦ䇘䇪Ҷሶ䘉
Ӌ㿴ࡉ䟽ᯠ㺘⽪Ѫᴤᴹ᭸Ⲵᖒᔿ઼Ոॆ㿴ࡉ३䝽Ⲵ㇇⌅Ǆ Doorenbos˄1993˅᧿䘠Ҷ䘉Ӌᯩ⌅
ᘾṧ֯ Soar ൘ањ䰞仈亶ฏѝᴹ᭸३䝽 100,000 ᶑ㿴ࡉˈ㘼нՊሩ⇿⣦ᘱ३䝽㿴ࡉⲴᔰ䬰ᴹབྷ
Ⲵ໎䮯Ǆ
ᓄ⭘ EBL ԕᆖҐᩌ㍒᧗ࡦਖањᇎ䐥кⲴ䰞仈൘Ҿˈཊᮠᛵߥлণ֯ሩᐼᵋⲴⴞḷᾲᘥ
ᔪ・䀓䟺ҏᴹ⴨ᖃབྷⲴ䇑㇇䟿Ǆֻྲˈ൘ỻ㊫䰞仈ѝᡁԜਟ㜭ᐼᵋᆖҐањⴞḷᾲᘥ˖Ā᫽֌
A ሬ㠤ᴰՈ䀓Ⲵ⣦ᘱǄāнᒨⲴᱟˈѪ䇱᰾ᡆ䀓䟺ѪӰѸ A ሬ㠤ᴰՈ䀓䴰㾱䀓䟺ަԆⲴ᫽֌Պ
ሬ㠤нྲ A Ⲵ䀓Ǆ䘉 а㡜䴰㾱ᩌ ㍒␡ᓖⲴᤷ ᮠ㓗Ⲵ䇑㇇䟿ǄChien˄1993 ˅઼ Tadepalli
˄1990˅᧒㍒ҶĀ⎸ᶱāᆖҐ઼Ā໎䟿āᆖҐⲴᯩ⌅ˈަѝ੟ਁᔿ㿴ࡉ㻛⭘Ҿӗ⭏䜘࠶Ⲵǃ䘁
լⲴǃն᱃䇑㇇Ⲵ䀓䟺Ǆо䀓䟺ᱟᆼ㖾ⲴᛵߥаṧˈаӋ㿴ࡉ㻛Ӿ䘉Ӌнᆼ㖾Ⲵ䀓䟺ѝᣭਆࠪ
ᶕǄᖃ❦䘉ӋᆖࡠⲴ㿴ࡉՊ⭡Ҿ䀓䟺Ⲵнᆼᮤᙗ㘼н↓⺞Ǆ㌫㔏䙊䗷ⴁ㿶൘ਾ㔝ᛵߥл㿴ࡉⲴ
ᙗ㜭ᶕ༴⨶↔䰞仈Ǆྲ᷌㿴ࡉਾᶕࠪ䭉ˈ䛓Ѹ৏࿻Ⲵ䀓䟺㻛໎䟿ൠᆼழԕ㾶ⴆᯠⲴᛵߥˈᒦф
Ӿ↔䀓䟺ѝᣭਆࠪᴤྭⲴ㿴ࡉǄ
ަԆᴹ䇨ཊ⹄ウᐕ֌᧒㍒ҶสҾ䀓䟺ⲴᆖҐⲴᓄ⭘ԕ᭩䘋สҾᩌ㍒Ⲵ䰞仈≲䀓ಘⲴ᭸⦷
˄ֻྲ Mitchell 1981; Silver 1983; Shavlik 1990; Mahadevan et al.1993; Gervasio & Dejong 1994;
Dejong 1994˅ǄBennett & Dejong ˄1996˅⹄ウҶสҾ䀓䟺ᆖҐ൘ᵪಘӪ㿴ࡂ㌫㔏Ⲵᓄ⭘ˈަ
㌫㔏ѝ᧿䘠ަц⭼઼㹼ѪⲴ亶ฏ⨶䇪ᱟнᆼ㖾ⲴǄDietterich & Flann ˄1995˅᧒㍒ҶสҾ䀓䟺
ᆖҐ઼໎ᕪᆖҐ˄㿱ㅜ 13 ㄐ˅Ⲵ䳶ᡀǄMitchell & Thrun˄1993˅᧿䘠ҶሶањสҾ䀓䟺Ⲵ⾎
㓿㖁㔌ᆖҐᯩ⌅˄㿱ㅜ 12 ㄐ䇘䇪Ⲵ EBNN ㇇⌅˅ᓄ⭘ࡠ໎ᕪᆖҐ䰞仈ѝǄ
ᵜㄐⲴ㾱⛩वᤜ
x 㓟㋩Ⲵᖂ㓣ᆖҐᯩ⌅ራ᢮ањ‫ٷ‬䇮ԕᤏਸ䇝㓳ᮠᦞˈо↔н਼ˈ㓟㋩Ⲵ࠶᷀ᆖҐ
ᯩ⌅ᩌራањ‫ٷ‬䇮ᤏਸᆖҐಘⲴ‫ݸ‬傼⸕䇶ᒦ㾶ⴆ䇝㓳ṧֻǄӪ㊫㓿ᑨ֯⭘‫ݸ‬傼⸕
䇶ᤷሬᯠ‫ٷ‬䇮ⲴᖒᡀǄᵜㄐ㘳ḕҶ㓟㋩Ⲵ࠶᷀ᆖҐᯩ⌅Ǆлаㄐӻ㓽ᖂ㓣-࠶᷀ᆖ
ҐⲴ㔃ਸǄ
x สҾ䀓䟺ⲴᆖҐᱟ࠶᷀ᆖҐⲴа⿽ᖒᔿˈަѝᆖҐಘ༴⨶⇿њᯠ䇝㓳ṧֻⲴᯩ⌅
ᱟ˄1˅᤹➗亶ฏ⨶䇪䀓䟺䈕ṧֻѝ㿲ሏࡠⲴⴞḷ٬ˈ˄2˅࠶᷀↔䀓䟺ԕ⺞ᇊ䀓
䟺ᡀ・Ⲵа㡜ᶑԦ˗˄3˅᭩䘋ަ‫ٷ‬䇮ԕਸᒦ䘉Ӌа㡜ᶑԦǄ
x Prolog-EBG ᱟањสҾ䀓䟺ⲴᆖҐ㇇⌅ˈᆳ֯⭘а䱦 Horn ᆀਕᶕ㺘⽪ަ亶ฏ⨶䇪
઼ᆖࡠⲴ‫ٷ‬䇮Ǆ൘ Prolog-EBG ѝˈ䀓䟺ণѪ Prolog 䇱᰾ˈ㘼Ӿ䀓䟺ѝᣭਆⲴ‫ٷ‬䇮
ᱟ↔䇱᰾Ⲵᴰᕡࡽ‫ۿ‬Ǆ֌Ѫ㔃᷌ˈ⭡ Prolog-EBG 䗃ࠪⲴ‫ٷ‬䇮Ӿަ亶ฏ⨶䇪ѝ╄㓾
⍮⭏Ǆ
x ྲ Prolog-EBG 䘉ṧⲴ࠶᷀ᆖҐᯩ⌅ᔪ・ᴹ⭘Ⲵѝ䰤⢩ᖱ֌Ѫ࠶᷀অ⤜䇝㓳ṧֻⲴ
ањ࢟᭸⭘Ǆ䘉⿽⭏ᡀ⢩ᖱⲴ࠶᷀䙄ᖴ㺕‫ݵ‬ҶྲਾੁՐ᫝䘉ṧⲴᖂ㓣ᯩ⌅ѝสҾ
㔏䇑ᯩ⌅Ⲵѝ䰤⢩ᖱ⭏ᡀ˄ྲ䳀㯿অ‫ݳ‬⢩ᖱ˅Ǆ
x 㲭❦ Prolog-EBG нՊӗ⭏㜭ᢙኅަ亶ฏ⨶䇪Ⲵ╄㓾䰝वⲴ‫ٷ‬䇮ˈަԆ╄㓾ᆖҐ䗷
〻ᴹ䘉њ㜭࣋Ǆֻྲˈањवਜ਼ determination ᯝ䀰˄ྲĀഭ㉽⺞ᇊ䈝䀰ā˅Ⲵ亶
ฏ⨶䇪ਟ㻛⭘Ҿо䇝㓳ᮠᦞа䎧╄㓾᧘⨶䎵ࠪ亶ฏ⨶䇪Ⲵ╄㓾䰝वⲴ‫ٷ‬䇮Ǆ
x ਟᓄ⭘↓⺞фᆼᮤⲴ亶ฏ⨶䇪Ⲵа㊫䟽㾱䰞仈Ѫབྷ⣦ᘱオ䰤ᩌ㍒Ⲵ䰞仈Ǆྲ
Prodigy ઼ Soar 䘉ṧⲴ㌫㔏ᐢֻ⽪ҶสҾ䀓䟺ⲴᆖҐᯩ⌅Ⲵ᭸⭘ˈᆳԜ㠚ࣘ㧧ਆᴹ
᭸Ⲵᩌ㍒㿴ࡉԕ࣐䙏ਾ㔝Ⲵ䰞仈≲䀓Ǆ
x 㲭❦สҾ䀓䟺ⲴᆖҐᯩ⌅ሩӪ㊫ᶕ䈤ᖸᴹ⭘ˈն㓟㋩Ⲵ╄㓾ᇎ⧠˄ྲ Prolog-
EBG˅ᴹа㕪⛩൘Ҿᆳ䗃ࠪⲴ‫ٷ‬䇮Ⲵ↓⺞ᙗਚ൘亶ฏ⨶䇪↓⺞ᰦ᡽㜭‫؍‬䇱Ǆ൘л
аㄐˈᡁԜ㘳ḕҶ㔃ਸᖂ㓣઼࠶᷀ᆖҐᯩ⌅Ⲵ䙄ᖴԕӾнᆼ㖾Ⲵ亶ฏ⨶䇪઼ᴹ䲀
䇝㓳ᮠᦞѝᴹ᭸ᆖҐǄ
࠶᷀ᆖҐᯩ⌅ⲴṩⓀਟ䘭ⓟࡠ Fikes et al.˄1972˅ᰙᵏⲴᐕ֌ˈᆳਟ䙊䗷ሩ ABSTRIPS ѝ

Ⲵ᫽֌ㅖⲴ࠶᷀ᆖҐᆿ᫽֌ㅖ (macro-operator)Ǆ䖳䘏аӋⲴᱟ Soloway(1977)Ⲵ⹄ウˈԆ൘ᆖ
Ґѝ֯⭘᰾⺞Ⲵ‫ݸ‬傼⸕䇶Ǆ㊫լҾᵜㄐ䇘䇪ⲴสҾ䀓䟺ⲴᆖҐᯩ⌅俆‫⧠ࠪݸ‬Ҿࠐњ 80 ᒤԓᰙ
ᵏ ᔰ ਁ Ⲵ ㌫㔏 ˈ व ᤜ DeJong(1981); Mitchell(1981); Winston et al.(1983); ઼ Silver(1983) Ǆ
DeJong & Mooney(1986)઼ Mitchell et al.(1986)ᨀ‫׋‬ҶሩสҾᴹ䀓䟺ᆖҐᯩ⌅Ⲵа㡜᧿䘠ˈ䘉Ӌ
ᕅਁҶ 80 ᒤԓᲊᵏሩ䘉њѫ仈Ⲵ⹄ウ✝▞Ǆ⭡‫׍‬䟼䈪ᯟབྷᆖᡰ‫Ⲵڊ‬а㌫ࡇสҾ䀓䟺ⲴᆖҐⲴ
⹄ウ⭡ DeJong(1993)᧿䘠ˈަѝवᤜ‫؞‬᭩䀓䟺Ⲵ㔃ᶴӾᗚ⧟Ⲵ઼ѤᰦⲴ䀓䟺ѝ↓⺞⌋ॆǄᴤཊ
ᴰ䘁Ⲵ⹄ウ⵰䟽ҾᢙኅสҾ䀓䟺Ⲵᯩ⌅ԕ֯⭘нᆼ㖾Ⲵ亶ฏ⨶䇪ˈԕ৺㔃ਸᖂ㓣ᆖҐ઼࠶᷀ᆖ
Ґ˄㿱ㅜ 12 ㄐ˅Ǆ‫ޣ‬ҾⴞⲴ઼‫ݸ‬傼⸕䇶൘Ӫ㊫઼ᵪಘᆖҐѝⲴ֌⭘ˈRam & Leake(1995)ᨀ‫׋‬
Ҷањ㔬ਸⲴਉ䘠ˈ㘼䘁ᵏสҾ䀓䟺ⲴᆖҐⲴᾲ㿸㿱 DeJong(1997)Ǆ
ᓄ⭘ᑖᴹᆼ㖾亶ฏ⨶䇪Ⲵᴰѕ㚳Ⲵቍ䈅ᱟ൘ᆖҐᩌ㍒᧗ࡦⲴ亶ฏˈᡆĀ࣐䙏āᆖҐǄ⭡
Laird et al.(1986)ᨀࠪⲴ Soar ㌫㔏઼ Carbonell et al.˄1990˅᧿䘠Ⲵ Prodigy ㌫㔏ᱟ֯⭘สҾ䀓
䟺ⲴᆖҐԕᆖҐ䰞仈≲䀓Ⲵєњᴰᡀ⟏Ⲵ㌫㔏Ǆ Rosenbloom & Laird(1986)䇘䇪Ҷ Soar ⲴᆖҐ
ᯩ⌅˄〠Ѫ chunking˅઼ަԆสҾ䀓䟺ᆖҐᯩ⌅ѻ䰤Ⲵ㍗ᇶ㚄㌫Ǆᴰ䘁 Dietterich & Flann
˄1995˅᧒㍒Ҷ㔃ਸสҾ䀓䟺ⲴᆖҐ઼໎ᕪᆖҐԕᆖҐᩌ㍒᧗ࡦⲴᯩ⌅Ǆ
㲭❦ᡁԜ䘉䟼Ⲵѫ㾱ⴞⲴᱟ⹄ウᵪಘᆖҐ㇇⌅ˈӽ䴰⌘᜿ࡠሩӪ㊫ᆖҐⲴᇎ傼ᙗ⹄ウѪӪ
㊫ᆖҐᱟสҾ䀓䟺䘉а⥌ᜣᨀ‫׋‬Ҷ᭟ᤱǄֻྲˈAhn et al.˄1987˅઼ Qin et al.˄1992˅ᾲ䘠Ҷ
᭟ᤱӪ㊫ᓄ⭘สҾ䀓䟺ᆖҐ䗷〻䘉а᧘ᜣⲴ䇱ᦞǄWisniewski & Medin˄1995˅᧿䘠ҶሩӪ㊫
ᆖҐⲴᇎ傼ᙗ⹄ウˈᆳᔪ䇞൘‫ݸ‬傼⸕䇶઼㿲ሏᮠᦞѻ䰤䘋㹼ѠᇼⲴ⴨ӂ֌⭘ԕᖡ૽ᆖҐ䗷〻Ǆ
Kotovsky & Baillargeon˄1994˅᧿䘠Ⲵᇎ傼䈤᰾ণ֯ 11 њᴸབྷⲴႤ‫ݯ‬൘ަᆖҐᰦҏᱟสҾަ
‫ݸ‬傼⸕䇶ⲴǄ
สҾ䀓䟺ⲴᆖҐѝᢗ㹼Ⲵ࠶᷀㊫լҾ Prolog 〻ᒿѝ֯⭘Ⲵࠐ㊫〻ᒿՈॆᯩ⌅ˈ∄ྲ䜘࠶䇴

ՠ˄partial evaluation˅Ǆvan Harmelen & Bundy ˄1988˅ᨀ‫׋‬Ҷሩ↔‫ޣ‬㌫Ⲵ䇘䇪Ǆ
Ґ仈
11.1 㘳㲁 ᆖ Ґ 䰞仈 Ѫ ᆖ Ґ ⴞ ḷ ᾲ ᘥ Ā ት տ ൘ ਼ а ᡯ ቻ ޵ Ⲵ є њ Ӫ ˈ ā 㺘⽪ Ѫ 䉃䇽
HouseMates(x,y)Ǆл䶒Ѫ↔ᾲᘥⲴањ↓ֻ˖
HouseMates(Joe, Sue)
Person(Joe) Person(Sue)
Sex(Joe, Male) Sex(Sue, Female)
HairColor(Joe, Black) HairColor(Sue, Brown)
Height(Joe, Short) Height(Sue, Short)
Nationality(Joe, US) Nationality(Sue, US)
Mother(Joe, Mary) Mother(Sue, Mary)
Age(Joe, 8) Age(Sue, 6)
л䶒Ⲵ亶ฏ⨶䇪ᴹࣙҾ㧧ਆ HouseMates ᾲᘥ˖
HouseMates(x, y)ĕInSameFamily(x, y)
HouseMates(x, y) ĕFraternityBrothers(x, y)
InSameFamily(x, y) ĕMarried(x, y)
InSameFamily(x, y) ĕYoungster(x) ġYoungster(y) ġSameMother(x,y)
SameMother(x,y) ĕMother(x,z) ġMother(y,z)
Youngster(x) ĕAget(x,a)ġLessThan(a,10)
ᓄ⭘ Prolog-EBG ㇇⌅ࡠ⌋ॆк䘠ᇎֻⲴԫ࣑ѝˈ֯⭘к䶒Ⲵ亶ฏ⨶䇪Ǆ⺞࠷ൠ䇢˖
˄a˅᡻ࣘᢗ㹼 Prolog-EBG ㇇⌅ᓄ⭘Ҿ↔䰞仈ˈҏቡᱟ䈤ˈ߉ࠪሩ↔ᇎֻ⭏ᡀⲴ䀓䟺ˈ߉

ࠪ↔䀓䟺ѝഎᖂⴞḷᾲᘥⲴ㔃᷌ˈԕ৺ᗇࡠⲴ Horn ᆀਕ㿴ࡉǄ
˄b˅‫ٷ‬ᇊⴞḷᾲᘥѪĀо Joe տ൘а䎧ⲴӪā㘼нᱟĀտ൘а䎧ⲴєњӪǄā⭘к䶒Ⲵ

⴨਼ⲴᖒᔿॆⲴᯩ⌅߉ࠪⴞḷᾲᘥǄ‫ٷ‬ᇊ䇝㓳ᇎ઼ֻ亶ฏ⨶䇪оԕࡽ⴨਼ˈProlog-EBG ሩ↔
ᯠⴞḷᾲᘥӗ⭏Ⲵ Horn ᆀਕᱟӰѸ˛
11.2 ྲ 11.3.1 㢲ᤷࠪⲴ, Prolog-EBG ਟᶴ䙐ࠪᒦ䶎ᇎֻⲴᱮᔿ⢩ᖱⲴᴹ⭘Ⲵᯠ⢩ᖱˈնᆳ

Ԝᱟ⭘ᱮᔿ⢩ᖱᇊѹⲴˈᒦфᴹࣙҾ᧿䘠ਸ䘲Ⲵ⌋ॆǄ䘉Ӌ⢩ᖱⲴ᧘ሬᱟ࠶᷀䇝㓳ṧֻ䀓䟺Ⲵ
ањ࢟᭸ᓄǄ᧘ሬᴹ⭘⢩ᖱⲴਖаᯩ⌅ᱟሩཊቲ⾎㓿㖁㔌֯⭘৽ੁՐ᫝㇇⌅ˈަѝᯠ⢩ᖱᱟส
Ҿབྷ䟿ṧֻⲴ㔏䇑኎ᙗ⭡䳀㯿অ‫ݳ‬ᆖҐࡠⲴǄ㜭੖᧘㦀а⿽ᯩ⌅ˈਟԕ㔃ਸ䘉Ӌ࠶᷀Ⲵ઼ᖂ㓣
Ⲵ䙄ᖴᶕ⭏ᡀᯠ⢩ᖱ˛˄⌘᜿˖䘉ᱟањᖵ䀓ߣⲴ⹄ウ䰞仈Ǆ˅
ㅢ12ㄖ ᖈ㓩ૂ࠼᷆ᆜҖⲺ㔉ਾ
㓟㋩Ⲵᖂ㓣ᆖҐᯩ⌅䙊䗷൘䇝㓳ṧֻѝራ᢮㓿傼ॆⲴ㿴ᖻᶕᖒᡀа㡜‫ٷ‬䇮Ǆ㓟㋩Ⲵ࠶᷀ᯩ
⌅֯⭘‫ݸ‬傼⸕䇶╄㓾᧘ሬа㡜‫ٷ‬䇮Ǆᵜㄐ㘳㲁ሶᖂ㓣઼࠶᷀Ⲵᵪࡦ㔃ਸ䎧ᶕⲴᯩ⌅ˈԕ㧧ᗇє
㘵ⲴՈ⛩˖ᴹ‫ݸ‬傼⸕䇶ᰦᴤ儈Ⲵ⌋ॆ㋮ᓖˈ઼‫׍‬䎆䇝㓳ᮠᦞ‫ݻ‬ᴽ‫ݸ‬傼⸕䇶Ⲵн䏣ǄᡰᗇࡠⲴ㔃
ਸⲴᯩ⌅∄㓟㋩Ⲵᖂ㓣ᆖҐᯩ⌅઼㓟㋩Ⲵ࠶᷀ᆖҐᯩ⌅ᙗ㜭䜭㾱儈Ǆᵜㄐ㘳㲁Ⲵᖂ㓣-࠶᷀ᆖ
Ґᯩ⌅਼ᰦสҾㅖਧ㺘⽪઼Ӫᐕ⾎㓿㖁㔌㺘⽪Ǆ
12.1 ࣞᵰ
൘ࡽࠐㄐᡁԜᐢ㓿㿱ࡠҶє⿽㊫රⲴᵪಘᆖҐ˖ᖂ㓣ᆖҐ઼࠶᷀ᆖҐǄᖂ㓣ᯩ⌅ྲߣㆆṁ
ᖂ㓣઼⾎㓿㖁㔌৽ੁՐ᫝ㅹˈᆳራ᢮ᤏਸ䇝㓳ᮠᦞⲴа㡜‫ٷ‬䇮Ǆ࠶᷀Ⲵᯩ⌅ྲ Prolog-EBGˈ
ᆳራ᢮ᤏਸ‫ݸ‬傼⸕䇶Ⲵа㡜‫ٷ‬䇮ˈ਼ᰦ֯ᆳ㾶ⴆ䇝㓳ᮠᦞǄ䘉є⿽ᆖҐ㤳රᡰสҾⲴሩᆖҐࡠ
Ⲵ‫ٷ‬䇮Ⲵ䇪䇱ᯩ⌅ᴹṩᵜⲴн਼ˈഐ↔ˈՈ㕪⛩ӂѪ㺕‫ݵ‬ǄሶᆳԜ㔃ਸ䎧ᶕᴹਟ㜭ᗇࡠᴤᕪᴹ
࣋ⲴᆖҐᯩ⌅Ǆ
㓟㋩Ⲵ࠶᷀ᆖҐᯩ⌅Ո⛩൘Ҿˈਟ⭘‫ݸ‬傼⸕䇶Ӿ䖳ቁⲴᮠᦞѝᴤ㋮⺞ൠ⌋ॆԕᕅሬᆖҐˈ
❦㘼ᖃ‫ݸ‬傼⸕䇶н↓⺞ᡆн䏣ᰦˈ䘉аᯩ⌅ਟ㜭Պ䘋‫↗ޕ‬䙄Ǆ㓟㋩Ⲵᖂ㓣ᯩ⌅ާᴹⲴՈ⛩ᱟн
䴰㾱ᱮᔿⲴ‫ݸ‬傼⸕䇶ˈᒦфѫ㾱สҾ䇝㓳ᮠᦞᆖҐࡠ㿴ᖻǄ❦㘼ˈ㤕䇝㓳ᮠᦞн䏣ᰦᆳ㜭Պཡ
䍕ˈᒦфՊ㻛ަѝ䳀ᔿⲴᖂ㓣‫ٿ‬㖞ᡰ䈟ሬˈ㘼ᖂ㓣‫ٿ‬㖞ᱟӾ㿲ሏᮠᦞѝ⌋ॆᡰᗵ䴰ⲴǄ㺘 12-1
ᾲ䘠Ҷє㘵Ⲵӂ㺕ⲴՈ⛩઼㕪䲧Ǆᵜㄐ㘳㲁Ⲵ䰞仈ᱟᘾṧሶҼ㘵㔃ਸᡀањঅ⤜Ⲵ㇇⌅ˈԕ㧧
ᗇᆳԜ਴㠚ⲴՈ⛩Ǆ
ᖂ㓣઼࠶᷀ᆖҐᯩ⌅ѻ䰤Ⲵн਼ਟӾᆳԜሩᆖҐࡠⲴ‫ٷ‬䇮䘋㹼Ⲵ䇪䇱˄justification˅Ⲵᙗ
䍘ѝⴻࠪǄ⭡㓟㋩Ⲵ࠶᷀ᆖҐ˄ྲ Prolog-EBG˅䗃ࠪⲴ‫ٷ‬䇮ᢗ㹼Ⲵᱟ䙫䗁˄logical˅䇪䇱˖䗃
ࠪⲴ‫ٷ‬䇮Ӿ亶ฏ⨶䇪઼䇝㓳ᮠᦞѝ╄㓾⍮⭏Ǆሩ㓟㋩Ⲵᖂ㓣ᆖҐᯩ⌅˄ྲ৽ੁՐ᫝˅䗃ࠪⲴ‫ٷ‬
䇮ᢗ㹼Ⲵᱟ㔏䇑䇪䇱˖䗃ࠪⲴ‫ٷ‬䇮Ӿ㔏䇑䇪ᦞѝ⍮⭏ˈᆳ䈤᰾䇝㓳ṧᵜ䏣ཏབྷӾ㘼ਟ㜭ԓ㺘ṧ
ֻⲴส߶࠶ᐳǄᖂ㓣Ⲵ㔏䇑䇪䇱൘ㅜ 7 ㄐ䇘䇪Ⲵ PAC ᆖҐѝᐢ㻛␵Რൠ䱀᰾Ǆ
ᰒ❦࠶᷀Ⲵᯩ⌅ᨀࠪ䙫䗁䇪䇱Ⲵ‫ٷ‬䇮ˈ㘼ᖂ㓣ᯩ⌅ᨀ‫׋‬㔏䇑䇪䇱Ⲵ‫ٷ‬䇮ˈᖸᇩ᱃ⴻࠪѪӰ
Ѹਟԕሶє㘵㔃ਸ䎧ᶕǄ䙫䗁Ⲵ䇪䇱Ⲵᕪᓖਚ⴨ᖃҾᆳԜᡰสҾⲴ‫ٷ‬ᇊᡆ‫ݸ‬傼⸕䇶Ǆྲ᷌‫ݸ‬傼
⸕䇶н↓⺞ᡆнਟᗇˈ䙫䗁䇪䇱ᱟнਟؑⲴфᰐ࣋ⲴǄ㔏䇑䇪䇱Ⲵᕪᓖ‫׍‬䎆ҾᆳԜสҾⲴᮠᦞ
઼㔏䇑‫ٷ‬ᇊǄᖃส߶࠶ᐳнਟؑᡆᮠᦞ〰㕪ᰦˈ㔏䇑䇪䇱ҏᱟнਟؑфᰐ࣋ⲴǄㆰ㘼䀰ѻˈє
⿽ᯩ⌅䪸ሩн਼Ⲵ㊫රⲴ䰞仈ᰦ᡽ᴹ᭸Ǆ䙊䗷є㘵Ⲵ㔃ਸˈᴹᵋᔰਁࠪᴤ䙊⭘ⲴᆖҐᯩ⌅ˈਟ
ԕ㾶ⴆ䖳ᒯⲴᆖҐԫ࣑Ǆ
㺘 12-1 㓟㋩Ⲵ࠶᷀ᆖҐ઼㓟㋩ᖂ㓣ᆖҐⲴ∄䖳
ᖂ㓣ᆖҐ ࠶᷀ᆖҐ
ⴞḷ ᤏਸᮠᦞⲴ‫ٷ‬䇮 ᤏਸ亶ฏ⨶䇪Ⲵ‫ٷ‬䇮
䇪䇱㔏䇑᧘⨶ ╄㓾᧘⨶
Ո⛩ 䴰㾱ᖸቁ‫ݸ‬傼⸕䇶 Ӿ〰㕪ⲴᮠᦞѝᆖҐ
㕪䲧〰㕪Ⲵᮠᦞˈн↓⺞Ⲵ‫ٿ‬㖞 нᆼ㖾Ⲵ亶ฏ⨶䇪
Inductive learning: ᖂ㓣ᆖҐ
Plentiful data: ѠᇼⲴᮠᦞ
No prior knowledge:ᰐ‫ݸ‬傼⸕䇶
Analytical learning: ࠶᷀ᆖҐ
Perfect prior knowledge:ᆼ㖾Ⲵ‫ݸ‬傼⸕䇶
Scarce data: 〰㕪Ⲵᮠᦞ
മ 12-1 ᆖҐԫ࣑Ⲵ࠶ᐳ㤳തǄ
൘ᴰᐖㄟˈ⋑ᴹਟ⭘Ⲵ‫ݸ‬傼⸕䇶ˈഐ↔䴰㾱㓟㋩Ⲵᖂ㓣ᆖҐᯩ⌅ԕ৺䖳儈Ⲵṧᵜ༽ᵲᓖǄ൘ᴰਣ
ㄟˈᴹᆼ㖾Ⲵ亶ฏ⨶䇪ˈਟԕ֯⭘ྲ Prolog-EBG 䘉ṧⲴ㓟㋩࠶᷀ᯩ⌅ǄᴤཊⲴᇎ䱵䰞仈սҾ䘉єњᶱ
ㄟѻ䰤Ǆ
മ 12-1 ᾲ䘠ҶᆖҐ䰞仈Ⲵ࠶ᐳ㤳തˈᆳ䲿⵰ਟ㧧ᗇⲴ‫ݸ‬傼⸕䇶઼䇝㓳ᮠᦞн਼㘼ਈॆǄ
൘ањᶱㄟˈᴹབྷ䟿Ⲵ䇝㓳ᮠᦞˈն⋑ᴹ‫ݸ‬傼⸕䇶Ǆ൘ਖаᶱㄟˈᴹᖸᕪⲴ‫ݸ‬傼⸕䇶ˈն䇝㓳
ᮠᦞᖸቁǄཊᮠᇎ䱵ᆖҐ䰞仈սҾ䘉єњᶱㄟѻ䰤Ǆֻྲˈ࠶᷀५⯇䇠ᖅⲴᮠᦞᓃԕᆖҐĀ⭘
⋫⯇᡻⇥ x ∄⋫⯇᡻⇥ y ᴤᴹ᭸Ⲵ⯵⯷āˈ䙊ᑨਟԕᔰ࿻Ҿ䘁լⲴ‫ݸ‬傼⸕䇶˄ྲ⯮⯵ѝ޵൘Ⲵ
ഐ᷌ᵪࡦⲴᇊᙗ⁑ර˅ˈ∄ྲ䇔ᇊᛓ㘵Ⲵփ⑙∄ԆⲴဃ਽ᴤ⴨‫ޣ‬Ǆ㊫լൠˈ൘࠶᷀ањ㛑⾘ᐲ
൪ᮠᦞᓃԕᆖҐⴞḷᾲᘥĀ㛑⾘٬൘ਾ 10 њᴸՊ㘫⮚Ⲵ‫ޜ‬ਨāѝˈྲ᷌ᐢᴹҶ㓿⍾ᆖⲴབྷᾲ
⸕䇶ˈਟԕᨀࠪ‫ޜ‬ਨⲴᙫ࡙⏖∄‫ޜ‬ਨḷᘇⲴ仌㢢ᴤ⴨‫ޣ‬Ǆ൘䘉є⿽䰞仈ѝˈᡁԜⲴ‫ݸ‬傼⸕䇶ᱟ
нᆼᮤⲴˈնᱮ❦ˈᆳᴹࣙҾ४࠶⴨‫઼ޣ‬н⴨‫Ⲵޣ‬⢩ᖱǄ
ᵜㄐ㘳㲁Ⲵ䰞仈ᱟ˖ĀᡁԜਟԕ䇮䇑ࠪᘾṧⲴ㇇⌅ˈ֯⭘䘁լⲴ‫ݸ‬傼⸕䇶㔃ਸਟ⭘ᮠᦞᶕ
ᖒᡀа㡜‫ٷ‬䇮˛ā⌘᜿ˈণ֯൘֯⭘㓟㋩Ⲵᖂ㓣ᆖҐ㇇⌅ᰦˈӽᴹᵪՊสҾ⢩ᇊᆖҐԫ࣑Ⲵ‫ݸ‬
傼⸕䇶ᶕ䘹ᤙ䇮䇑ᯩṸǄֻྲˈᖃᓄ⭘৽ੁՐ᫝ᶕ䀓ߣ䈝丣䇶࡛䘉ṧⲴ䰞仈ᰦˈ䇮䇑㘵ᗵ享䘹
ᤙ䗃‫઼ޕ‬䗃ࠪᮠᦞⲴ㕆⸱ᯩᔿǃ൘ởᓖл䱽ѝ㻛ᴰሿॆⲴ䈟ᐞ࠭ᮠǃ䳀㯿অ‫Ⲵݳ‬ᮠ䟿ǃ㖁㔌Ⲵ
ᤃᢁ㔃ᶴǃᆖҐ䙏⦷઼ߢ䟿ㅹǄ൘‫ڊ‬䘉Ӌ䘹ᤙᰦˈ䇮䇑㘵ਟԕሶ亶ฏ⢩ᇊⲴ⸕䇶፼‫ࡠޕ‬ᆖҐ㇇
⌅ѝǄ❦㘼㔃᷌ӽ❦ᱟ㓟㋩Ⲵᖂ㓣㇇⌅৽ੁՐ᫝Ⲵањᇎ⧠ˈ⭡䇮䇑㘵⢩↺ॆਾ䪸ሩ䈝丣䇶࡛
ԫ࣑ǄᡁԜⲴᝏ‫ޤ‬䏓Ⲵн൘Ҿ↔ˈ㘼൘Ҿањ㌫㔏㜭ሶ‫ݸ‬傼⸕䇶֌ѪᱮᔿⲴ䗃‫ޕ‬㔉ᆖҐಘˈ䇝
㓳ᮠᦞҏ਼ṧ֌Ѫᱮᔿ䗃‫ޕ‬Ǆ䘉ṧᆳԜӽѪ䙊⭘Ⲵ㇇⌅ˈն࡙⭘Ҷ亶ฏ⢩ᇊⲴ⸕䇶Ǆㆰ㾱ൠᾲ
ᤜалˈᡁԜᝏ‫ޤ‬䏓Ⲵᱟ֯⭘ᱮᔿ䗃‫Ⲵޕ‬亶ฏ⴨‫⸕ޣ‬䇶Ⲵ亶ฏᰐ‫ޣ‬㇇⌅Ǆ
ሩҾ㔃ਸҶᖂ㓣઼࠶᷀ᆖҐⲴн਼ᯩ⌅ˈᓄ֯⭘ӰѸṧⲴ߶ࡉᶕ∄䖳ᆳԜ઒˛⭡ҾᆖҐಘ
а㡜н㜭亴‫⸕ݸ‬䚃亶ฏ⨶䇪઼䇝㓳ᮠᦞⲴ䍘䟿ǄᡁԜᝏ‫ޤ‬䏓Ⲵᱟ㜭ሩമ 12-1 ᮤњ䰞仈㌫ࡇ䜭
ਟ᫽֌Ⲵа㡜ᯩ⌅Ǆ䘉ṧⲴᆖҐᯩ⌅ᓄާᴹԕлⲴ⢩↺኎ᙗ˖
x ྲ᷌⋑ᴹ亶ฏ⨶䇪ˈᆳ㠣ቁ㜭‫ۿ‬㓟㋩Ⲵᖂ㓣ᯩ⌅аṧᴹ᭸ᆖҐǄ
x ྲ᷌ᴹᆼ㖾Ⲵ亶ฏ⨶䇪ˈᆳ㠣ቁ㜭‫ۿ‬㓟㋩Ⲵ࠶᷀ᯩ⌅ṧᴹ᭸ᆖҐǄ
x ྲ᷌亶ฏ⨶䇪઼䇝㓳ᮠᦞ䜭нᆼ㖾ˈᆳᓄ㜭㔃ਸє㘵Ⲵ䮯༴ˈ∄অ㓟Ⲵᖂ㓣ᡆ࠶
᷀ᯩ⌅Ⲵᙗ㜭䜭㾱ྭǄ
x ᆳᓄ㜭༴⨶䇝㓳ᮠᦞѝᵚ⸕〻ᓖⲴᐞ䭉Ǆ
x ᆳᓄ㜭༴⨶亶ฏ⨶䇪ѝᵚ⸕〻ᓖⲴᐞ䭉Ǆ
⌘᜿䘉䟼ࡇࠪⲴᵏᵋⴞḷᖸ䳮䗮ࡠǄֻྲˈ༴⨶䇝㓳ᮠᦞѝⲴᐞ䭉ˈণ֯൘สҾ㔏䇑Ⲵᖂ
㓣ᯩ⌅ѝˈྲ᷌⋑ᴹḀӋ‫ݸ‬傼⸕䇶઼ሩᐞ䭉࠶ᐳⲴ‫ٷ‬ᇊˈ䘉ӽᱟ٬ᗇ⹄ウⲴ䰞仈Ǆ㔃ਸᖂ㓣઼
࠶᷀ᆖҐⲴᯩ⌅ᱟᖃࡽ⍫䏳Ⲵ⹄ウ亶ฏǄ㲭❦к䶒ࡇࠪⲴᱟᡁԜᐼᵋ㇇⌅㜭䗮ࡠⲴ㖾ྭᙗ䍘ˈ
ⴞࡽ⋑ᴹ㇇⌅㜭ԕᆼ‫ޘ‬а㡜ॆⲴᯩᔿ┑䏣ᡰᴹ䘉Ӌ㓖ᶏǄ
ла㢲ሩ㔃ਸᖂ㓣-࠶᷀ᆖҐⲴ䰞仈֌ࠪҶᴤ䈖㓶Ⲵ䇘䇪Ǆਾ䶒ࠐ㢲᧿䘠Ҷ 3 ⿽н਼Ⲵ䙄
ᖴˈ㔃ਸ䘁լⲴ‫ݸ‬傼⸕䇶઼ਟ⭘ᮠᦞᶕᤷሬᆖҐಘᩌ㍒ਸ䘲Ⲵ‫ٷ‬䇮Ǆ⇿⿽䙄ᖴ䜭ᐢ൘ཊњ䰞仈
亶ฏѝᱮ⽪ࠪᴹ䎵ࠪ㓟ᖂ㓣ᯩ⌅Ⲵᙗ㜭ǄѪᯩ‫∄ׯ‬䖳ˈᡁԜ֯⭘਼аֻᆀᶕ䈤᰾ 3 ⿽䙄ᖴǄ
12.2 ᆜҖⲺᖈ㓩-࠼᷆䙊ᖺ
12.2.1 ᆜҖ䰤从
ᾲ㘼䀰ѻˈᵜㄐ㘳㲁ⲴᆖҐ䰞仈Ѫ˖
ᐢ⸕˖
x ањ䇝㓳ṧֻ䳶ਸ Dˈਟ㜭वਜ਼ᐞ䭉
x ањ亶ฏ⨶䇪 Bˈਟ㜭वਜ਼ᐞ䭉
x ‫ى‬䘹‫ٷ‬䇮Ⲵオ䰤 H
≲䀓˖
x ањᴰྭൠᤏਸ䇝㓳ṧ઼ֻ亶ฏ⨶䇪Ⲵ‫ٷ‬䇮
Āᴰྭൠᤏਸ䇝㓳ṧ઼ֻ亶ฏ⨶䇪ā䘉ਕ䈍⺞࠷ਜ਼ѹᱟӰѸ˛ᡆ㘵䈤ˈᱟ੖Պ䘹ᤙањᤏ
ਸᮠᦞ〻ᓖ䖳ྭ㘼ᤏਸ⨶䇪䖳ᐞⲴ‫ٷ‬䇮ˈᡆ৽ѻ˛ѪҶᴤ㋮⺞䎧㿱ˈ䴰㾱ᇊѹሩᓄᮠᦞ઼ሩᓄ
Ҿ亶ฏ⨶䇪Ⲵ‫ٷ‬䇮䭉䈟⦷ᓖ䟿ˈ❦ਾ⭘䘉Ӌ䭉䈟⦷ᶕ㺘⽪䘉њ䰞仈Ǆഎᗶㅜ 5 ㄐѝerrorD(h)ᇊ
ѹѪDѝ㻛h䈟࠶㊫Ⲵṧֻᡰঐ∄ֻǄਟᇊѹh‫ޣ‬Ҿ亶ฏ⨶䇪BⲴ䭉䈟⦷errorB(h)ѪˈhоB൘࠶㊫
ањ䲿ᵪᣭਆᇎֻᰦна㠤Ⲵᾲ⦷Ǆ᧕лᶕቡਟቍ䈅⭘䘉Ӌ䭉䈟⦷Ⲵᖒᔿ࡫⭫ᡰᐼᵋⲴ䗃ࠪ‫ٷ‬
䇮ǄֻྲˈᡁԜਟԕ㾱≲‫ٷ‬䇮֯к䘠䭉䈟⦷ⲴḀ⿽㔬ਸᓖ䟿ᴰሿॆˈྲ˖
arg min k D errorD (h) k B errorB (h)

hH
㲭❦㋇ⴻ䎧ᶕ䘉ᖸਸ⨶ˈն䘈н␵ᾊᘾṧ⺞ᇊkD઼kBⲴ٬ˈԕᤷᇊᤏਸᮠᦞ઼ᤏਸ⨶䇪є
㘵Ⲵ⴨ሩ䟽㾱〻ᓖǄྲ᷌ᴹ䶎ᑨᐞⲴ⨶䇪ˈতᴹབྷ䟿ਟ䶐ᮠᦞˈᴰྭ֯errorD(h)Ⲵᵳ٬ᴤབྷǄ
ྲ᷌ᴹᖸᕪⲴ⨶䇪ˈ㘼ᮠᦞṧᵜᖸሿфᆈ൘བྷ䟿ಚ༠ˈᢺerrorB(h)Ⲵᵳ٬໎བྷՊᗇࡠᴰྭⲴ㔃
᷌Ǆᖃ❦ྲ᷌ᆖҐಘ亴‫ݸ‬н⸕䚃亶ฏ⨶䇪઼䇝㓳ᮠᦞⲴ䍘䟿ˈᆳቡн␵ᾊ䈕ᘾṧѪ䘉є䜘࠶䭉
䈟⦷࣐ᵳǄ
ᘾṧ⺞ᇊ‫ݸ‬傼⸕䇶઼ᮠᦞᵳ٬䘉а䰞仈ˈਖа⿽䀓ߣᯩ⌅ᱟ֯⭘䍍ਦᯟⲴ㿲⛩ᶕ㘳㲁Ǆഎ
ᗶалㅜ 6 ㄐˈ䍍ਦᯟᇊᖻ᧿䘠Ҷᘾṧ䇑㇇㔉ᇊ䇝㓳ᮠᦞ D ᰦ‫ٷ‬䇮 h Ⲵਾ傼ᾲ⦷ P(h|D)Ǆ⺞࠷
ൠ䇢ˈ䍍ਦᯟᇊᖻ䇑㇇↔ਾ傼ᾲ⦷ᱟสҾ㿲ሏࡠⲴᮠᦞ D ԕ৺‫ݸ‬傼⸕䇶Ⲵˈԕ P(h)ǃP(D)઼
P(D|h)Ⲵᖒᔿ㺘⽪Ǆഐ↔ᡁԜਟᢺ P(h)ǃP(D)઼ P(D|h)ⴻ֌ᱟḀ⿽ᖒᔿⲴ㛼Ჟ⸕䇶ᡆ亶ฏ⨶
䇪ˈ㘼фਟᢺ䍍ਦᯟ⨶䇪ⴻᡀа⿽Ѫ亶ฏ⨶䇪࣐ᵳⲴᯩ⌅ˈᆳо㿲ሏࡠⲴᮠᦞ D а䎧ˈ䍻Ҹ h
Ⲵਾ傼ᾲ⦷Ѫ P(h|D)Ǆ᤹➗䍍ਦᯟⲴ㿲⛩ˈᡰ䘹ᤙⲴ‫ٷ‬䇮ᓄѪਾ傼ᾲ⦷ѝᴰབྷⲴањˈᒦф䍍
ਦᯟ‫ޜ‬ᔿᨀ‫׋‬ҶѪ↔‫ݸ‬傼⸕䇶઼㿲ሏࡠᮠᦞⲴ䍑⥞࣐ᵳⲴ㢟ྭᯩ⌅ǄнᒨⲴᱟˈ䍍ਦᯟ‫ޜ‬ᔿ䳀
ਜ਼‫ٷ‬ᇊҶᤕᴹ‫ޣ‬Ҿ P(h)ǃP(D)઼ P(D|h)ᾲ⦷࠶ᐳⲴᆼ㖾Ⲵ⸕䇶Ǆᖃ䘉Ӌ䟿ਚᱟ䘁լᐢ⸕ᰦˈঅ
⤜䍍ਦᯟ‫ޜ‬ᔿ⋑ᴹ㿴ᇊྲօሶަо㿲ሏᮠᦞ㔃ਸ䎧ᶕǄ˄൘↔ᛵߥла⿽ᯩ⌅ᱟ‫ٷ‬ᇊᴹ P(h)ˈ
P(D)઼ P(D|h)ѻкⲴ‫ݸ‬傼ᾲ⦷࠶ᐳˈ❦㘼ਾ䇑㇇ਾ傼ᾲ⦷ P(h|D)Ⲵᵏᵋ٬Ǆ❦㘼䘉㾱≲ᴹ
P(h)ˈP(D)઼ P(D|h)ѻкⲴ‫ݸ‬傼࠶ᐳᯩ䶒Ⲵ䱴࣐⸕䇶ˈഐ↔ᒦ⋑ᴹⵏ↓䀓ߣ↔䰞仈Ǆ˅
ᖃ㘳㲁⢩ᇊ㇇⌅ᰦˈᡁԜՊ޽⅑㘳㲁Āᴰ֣āᤏਸ‫ٷ‬䇮઼ᮠᦞᱟӰѸਜ਼ѹǄ⧠൘ˈᡁԜਚ
ᱟㆰঅൠ〠ᆖҐ䰞仈ᱟѪҶ֯‫ٷ‬䇮൘ᮠᦞ઼亶ฏ⨶䇪к䭉䈟⦷ⲴḀ⿽㔬ਸᓖ䟿ᴰሿॆǄ
12.2.2 ‫ٽ‬䇴グ䰪ᩒ㍘
ྲօሶ亶ฏ⨶䇪઼䇝㓳ᮠᦞᴰྭൠ㔃ਸ䎧ᶕˈԕ䲀ࡦਟ᧕ਇ‫ٷ‬䇮Ⲵᩌ㍒˛䘉൘ᵪಘᆖҐѝ
ӽᱟᖵ⹄ウⲴ䰞仈Ǆᵜㄐ㘳ሏҶࠐ⿽ᐢᨀࠪⲴᯩ⌅ˈަѝ䇨ཊ㾱ሩᐢ䇘䇪䗷Ⲵᖂ㓣ᯩ⌅˄ྲ৽
ੁՐ᫝ˈFOIL˅䘋㹼ᢙኅǄ
ѪҶ䀓ਟ㜭䙄ᖴⲴ㤳തˈа⿽࣎⌅ᱟഎࡠࡽ䶒ሩᆖҐⲴⴻ⌅ˈণሶަⴻ֌ᱟа⿽ᩌ㍒ཊњ
ਟ䘹‫ٷ‬䇮オ䰤Ⲵԫ࣑ǄѪҶሶབྷཊᮠᆖҐԫ࣑࡫⭫Ѫᩌ㍒㇇⌅ˈ䴰㾱ᇊѹᖵᩌ㍒Ⲵ‫ٷ‬䇮オ䰤
Hˈᩌ㍒Ⲵᔰ࿻⛩ࡍ࿻‫ٷ‬䇮h0ˈᇊѹঅњᩌ㍒↕Ⲵᩌ㍒㇇ᆀ䳶ਸOˈԕ৺ᤷᇊᩌ㍒ⴞḷⲴⴞⲴ
ࡔᦞGǄᵜ䰤ѝ᧒㍒Ҷ 3 ⿽ᯩ⌅֯⭘‫ݸ‬傼⸕䇶ᶕ᭩ਈ㓟ᖂ㓣ᯩ⌅ѝᢗ㹼Ⲵᩌ㍒Ǆ
x ֯⭘‫ݸ‬傼⸕䇶᧘ሬࠪᩌ㍒䎧↕Ⲵࡍ࿻‫ٷ‬䇮Ǆ⭘䘉⿽ᯩ⌅ˈ亶ฏ⨶䇪B㻛⭘Ҿᔪ・а
њоBа㠤Ⲵࡍ࿻‫ٷ‬䇮h0Ǆ❦ਾԕ䘉њࡍ࿻‫ٷ‬䇮h0Ѫ䎧⛩ᓄ⭘ḷ߶ᖂ㓣ᯩ⌅Ǆֻ
ྲˈл䶒᧿䘠ⲴKBNN㌫㔏ᱟ᤹䘉⿽ᯩ⌅ᆖҐӪᐕ⾎㓿㖁㔌ⲴǄᆳ֯⭘‫ݸ‬傼⸕䇶ᶕ
䇮䇑ࡍ࿻㖁㔌Ⲵӂ㚄㔃ᶴ઼ᵳ٬ˈ䘉ṧˈ↔ࡍ࿻㖁㔌о㔉ᇊⲴ亶ฏ⨶䇪ᆼ‫ޘ‬а
㠤Ǆ❦ਾ↔ࡍ࿻㖁㔌‫ٷ‬䇮֯⭘৽ੁՐ᫝㇇⌅઼䇝㓳ᮠᦞ㻛ᖂ㓣ൠ㋮ॆǄӾањо
亶ฏ⨶䇪а㠤Ⲵ‫ٷ‬䇮ᔰ࿻ᩌ㍒ˈ֯ᗇᴰ㓸䗃ࠪ‫ٷ‬䇮ᴤᴹਟ㜭ᤏਸ↔⨶䇪Ǆ
x ֯⭘‫ݸ‬傼⸕䇶ᶕ᭩ਈ‫ٷ‬䇮オ䰤ᩌ㍒ⲴⴞḷǄ൘䘉⿽ᯩ⌅ѝˈⴞⲴࡔᦞ G 㻛‫؞‬᭩ˈ
ԕ㾱≲䗃ࠪ‫ٷ‬䇮ᤏਸ䇝㓳ṧֻⲴ਼ᰦҏᤏਸ亶ฏ⨶䇪Ǆֻྲˈл䶒᧿䘠Ⲵ EBNN
㌫㔏ԕ䘉⿽ᯩ⌅ᆖҐ⾎㓿㖁㔌Ǆ⾎㓿㖁㔌Ⲵᖂ㓣ᆖҐᢗ㹼ởᓖл䱽ᶕ֯㖁㔌൘䇝
㓳ᮠᦞкⲴ䈟ᐞᒣᯩᴰሿॆˈ㘼 EBNN ѝᢗ㹼ởᓖл䱽ᶕՈॆਖањࡔᦞǄ䘉њ
‫؞‬᭩Ⲵࡔᦞवਜ਼ањ䱴࣐亩ˈᆳ㺑䟿ҶᆖҐࡠⲴ㖁㔌⴨ሩҾ亶ฏ⨶䇪Ⲵ䈟ᐞǄ
x ֯⭘‫ݸ‬傼⸕䇶᭩ਈਟ⭘Ⲵᩌ㍒↕Ǆ൘↔ᯩ⌅ѝˈ亶ฏ⨶䇪‫؞‬᭩Ҷᩌ㍒㇇ᆀ䳶ਸ
OǄֻྲˈл䶒᧿䘠Ⲵ FOCL ㌫㔏ԕ䘉⿽ᯩ⌅ᆖҐ Horn ᆀਕ䳶ǄᆳสҾᖂ㓣㌫㔏
FOILǄFOIL ൘ਟ㜭Ⲵ Horn ᆀਕオ䰤кᢗ㹼䍚ႚᩌ㍒ˈ⇿↕䙊䗷࣐‫ޕ‬ањᯠ᮷ᆇ
ᶕ‫↓؞‬ᖃࡽ‫ٷ‬䇮ǄFOCL ൘‫ٷ↓؞‬䇮ѝ֯⭘亶ฏ⨶䇪ᶕᢙኅਟ⭘Ⲵ᮷ᆇ䳶ਸǄᆳ‫ݱ‬
䇨൘অњᩌ㍒↕ѝ࣐‫ޕ‬ཊњ᮷ᆇˈਚ㾱ᆳԜ㜭⭡亶ฏ⨶䇪‫؍‬䇱ަ↓⺞ᙗǄԕ䘉⿽
ᯩᔿˈFOCL ൘‫ٷ‬䇮オ䰤ѝ〫ࣘа↕⴨ᖃҾ֯⭘৏ᶕⲴ㇇⌅〫ࣘཊ↕Ǆ䘉ӋĀᆿ〫
ࣘā˄macro-moves˅ਟᶱབྷൠ᭩ਈᩌ㍒Ⲵᯩੁˈ䘉ṧᴰ㓸Ⲵоᮠᦞа㠤Ⲵ‫ٷ‬䇮о
ਚ֯⭘ᖂ㓣ᩌ㍒↕ᰦ᢮ࡠⲴ‫ٷ‬䇮н਼Ǆ
л䶒ࠐ㢲‫⅑׍‬ӻ㓽Ҷ䘉ࠐ⿽ᯩ⌅Ǆ
12.3 ֵ⭞‫ݾ‬僂⸛䇼ᗍࡦࡓခ‫ٽ‬䇴
а⿽֯⭘‫ݸ‬傼⸕䇶Ⲵᯩ⌅ᱟˈሶ‫ٷ‬䇮ࡍ࿻ॆѪᆼ㖾ᤏਸ亶ฏ⨶䇪ˈ❦ਾ᤹➗䴰㾱ᖂ㓣ൠ㋮
ॆ↔ࡍ࿻‫ٷ‬䇮ԕᤏਸ䇝㓳ᮠᦞǄ䘉⿽ᯩ⌅㻛⭘Ҿ KBANN˄Knowledge-Based Artificial Neural
NetworkˈสҾ⸕䇶ⲴӪᐕ⾎㓿㖁㔌˅㇇⌅ѝǄ൘ KBANN ѝˈ俆‫ݸ‬ᔪ・Ҷањࡍ࿻Ⲵ㖁㔌Ǆ
ሩ⇿њਟ㜭ᇎֻˈ㖁㔌䍻ҸᆳⲴ࠶㊫ㅹҾ亶ฏ⨶䇪䍻ҸⲴ࠶㊫Ǆ❦ਾᓄ⭘Ҷ৽ੁՐ᫝㇇⌅ᶕ䈳
ᮤࡍ࿻㖁㔌ˈ֯ަᤏਸ䇝㓳ṧֻǄ
ᖸᇩ᱃ⴻࠪˈ䈕ᢰᵟⲴࣘᵪ൘Ҿ˖ྲ᷌亶ฏ⨶䇪ᱟ↓⺞Ⲵˈࡍ࿻‫ٷ‬䇮ሶ↓⺞࠶㊫ᡰᴹ䇝㓳
ṧֻˈ㘼ᰐ䴰޽ሩަ‫↓؞‬Ǆ❦㘼ˈྲ᷌ࡍ࿻‫ٷ‬䇮н㜭ᆼ㖾ൠ࠶㊫䇝㓳ṧֻˈ䛓Ѹᆳ䴰㾱㻛ᖂ㓣
㋮ॆˈԕ᭩䘋ަ൘䇝㓳ṧֻкⲴᤏਸᓖǄഎᗶ൘㓟㋩ᖂ㓣Ⲵ৽ੁՐ᫝㇇⌅ѝˈᵳ٬а㡜㻛ࡍ࿻
ॆѪሿⲴ䲿ᵪ٬ǄKBANN 㛼ਾⲴⴤ㿲ਜ਼ѹ൘Ҿˈণ֯亶ฏ⨶䇪ᱟ䘁լ↓⺞Ⲵˈሶ㖁㔌ࡍ࿻ॆ
Ѫᤏਸ亶ฏ⨶䇪ˈ∄ࡍ࿻ॆѪ䲿ᵪᵳ٬ᴹᴤྭⲴ䘁լᔰㄟǄ䘉ᓄ䈕ՊᗇࡠᴹᴤྭⲴ⌋ॆ㋮ᓖⲴ
ᴰ㓸‫ٷ‬䇮Ǆ
䘉 ⿽ ֯ ⭘ 亶 ฏ ⨶ 䇪 ᶕ ࡍ ࿻ ॆ ‫ ٷ‬䇮 Ⲵ 䙄 ᖴ ᐢ 㓿㻛䇨 ཊ ⹄ ウ㘵 ᧒ ㍒䗷 Ǆ व ᤜ Shavlik &

Towell(1989)ˈTowell & Shavlik(1994), Fu (1989, 1993)઼ Pratt(1993a, 1993b)ǄᡁԜሶ ֯⭘
Shavlik & Towell˄1989˅᧿䘠Ⲵ KBANN ㇇⌅ᶕֻ⽪䘉а䙄ᖴǄ
12.3.1 KBANN ㇍⌋
KBANN 䘀⭘亶ฏ⨶䇪ᶕࡍ࿻ॆ‫ٷ‬䇮Ǆަѝ‫ٷ‬ᇊ亶ฏ⨶䇪⭘а㓴ભ仈ᖒᔿⲴ䶎䙂ᖂⲴ Horn

ᆀਕᶕ㺘⽪Ǆભ仈ᖒᔿ Horn ᆀਕ㺘⽪ᆳнवਜ਼ਈ䟿ǄKBANN Ⲵ䗃‫઼ޕ‬䗃ࠪྲл˖
ᐢ⸕˖
x а㓴䇝㓳ṧֻ
x ⭡䶎䙂ᖂભ仈ර Horn ᆀਕ㓴ᡀⲴ亶ฏ⨶䇪
≲䀓˖
x ањᤏਸ䇝㓳ṧֻⲴˈ㻛亶ฏ⨶䇪‫ٿ‬㖞ⲴӪᐕ⾎㓿㖁㔌
㺘 12-2KBANN ㇇⌅
亶ฏ⨶䇪㻛䖜ᦒѪㅹ᭸Ⲵ⾎㓿㖁㔌˄↕僔 1ˉ3˅ˈ❦ਾ⭘৽ੁՐ᫝㇇⌅ᖂ㓣㋮ॆ˄ㅜ 4 ↕˅ǄW

ᑨ䟿Ⲵ٬ਟਆѪ 0.4Ǆ
KBANN(Domain_Theory, Training_Examples)
Domain_Theory: 䶎䙂ᖂભ仈ර Horn ᆀਕ䳶ਸ
Training_Examples:ⴞḷ࠭ᮠⲴ<input, output>ሩⲴ䳶ਸ
࠶᷀↕˖ࡋᔪањㅹԧҾ亶ฏ⨶䇪Ⲵࡍ࿻㖁㔌
1. ሩ⇿њᇎֻ኎ᙗࡋᔪањ㖁㔌䗃‫ޕ‬
2. ሩ Domain_Theory Ⲵ⇿њ Horn ᆀਕˈྲлࡋᔪањ㖁㔌অ‫ݳ‬
䘎᧕↔অ‫Ⲵݳ‬䗃‫↔ࡠޕ‬ᆀਕⲴ‫ݸ‬㹼䇽⍻䈅Ⲵ኎ᙗ
ሩᆀਕⲴ⇿њ䶎䍏‫ݸ‬㹼䇽ˈ䍻Ҹᵳ٬ W 㔉ሩᓄⲴ sigmoid অ‫ݳ‬䗃‫ޕ‬
ሩᆀਕⲴ⇿њ䍏‫ݸ‬㹼䇽ˈ䍻Ҹᵳ٬-W 㔉ሩᓄⲴ sigmoid অ‫ݳ‬䗃‫ޕ‬
䇮㖞↔অ‫Ⲵݳ‬䰸٬w0Ѫ-(n-0.5)WˈަѝnѪᆀਕⲴ䶎䍏‫ݸ‬㹼䇽Ⲵᮠⴞ
3. ൘㖁㔌অ‫ݳ‬ѻ䰤໎࣐䱴࣐Ⲵ䘎᧕ˈ䘎᧕␡ᓖѪ i Ⲵ⇿њ㖁㔌অ‫␡ࡠݳ‬ᓖѪ i+1 Ⲵᡰᴹ㖁㔌অ‫Ⲵݳ‬
䗃‫ޕ‬ቲкǄ䍻Ҹ䘉Ӌ䱴࣐Ⲵ䘎᧕Ѫ᧕䘁 0 Ⲵ䲿ᵪᵳ٬Ǆ
ᖂ㓣↕˖㋮ॆ↔ࡍ࿻㖁㔌
4. ᓄ⭘৽ੁՐ᫝㇇⌅ᶕ䈳ᮤࡍ࿻㖁㔌ᵳ٬ԕᤏਸ Training_Examples
KBANN ㇇⌅वਜ਼єњ䱦⇥ˈ俆‫ݸ‬ᆳࡋᔪањᆼ㖾ᤏਸ亶ฏ⨶䇪ⲴӪᐕ⾎㓿㖁㔌ˈ❦ਾ֯
⭘৽ੁՐ᫝㇇⌅ᶕ㋮ॆࡍ࿻㖁㔌ԕᤏਸ䇝㓳ṧֻǄ㇇⌅Ⲵ㓶㢲ˈवᤜࡋᔪࡍ࿻㖁㔌Ⲵ㇇⌅൘㺘
12-2 ѝࡇࠪˈᒦሶ൘ 12.3.2 㢲䈤᰾Ǆ
㺘 12-3Cup ᆖҐԫ࣑
㺘ѝࡇࠪҶⴞḷᾲᘥ Cup Ⲵа㓴䘁լ亶ฏ⨶䇪઼а㓴䇝㓳ṧֻǄ
亶ฏ⨶䇪˖
CupĕStable, Liftable, OpenVessel
StableĕBottomIsFlat
LiftableĕGraspable, Light
GraspableĕHasHandle
OpenVesselĕHasConcavity, ConcavityPointsUp
䇝㓳ṧֻ˖
Cups Non-Cups
BottomIsFlat Ĝ Ĝ Ĝ Ĝ Ĝ Ĝ Ĝ Ĝ
ConcavityPointsUp Ĝ Ĝ Ĝ Ĝ Ĝ Ĝ Ĝ
Expensive Ĝ Ĝ Ĝ Ĝ
Fragile Ĝ Ĝ Ĝ Ĝ Ĝ Ĝ
HandleOnTop Ĝ Ĝ
HandleOnSide Ĝ Ĝ Ĝ
HasConcavity Ĝ Ĝ Ĝ Ĝ Ĝ Ĝ Ĝ Ĝ Ĝ
HasHandle Ĝ Ĝ Ĝ Ĝ Ĝ
Light Ĝ Ĝ Ĝ Ĝ Ĝ Ĝ Ĝ Ĝ
MadeOfCeramic Ĝ Ĝ Ĝ Ĝ
MadeOfPaper Ĝ Ĝ
MadeOfStyrofoam Ĝ Ĝ Ĝ Ĝ
12.3.2 жѠ⽰ׁ
Ѫֻ⽪ KBANN Ⲵ᫽֌ˈ㘳㲁㺘 12-3 ࡇࠪⲴањㆰঅⲴᆖҐ䰞仈ˈᆳਆ㠚Ҿ Towell &

Shavlik˄1989˅ᒦ⮕֌᭩ࣘǄ䘉䟼⇿њᇎֻԓ㺘а⢙⨶ሩ䊑Ǆ᧿䘠ҶᆳⲴ⢙⨶ᶀᯉǃᆳⲴ䖫䟽
ㅹㅹǄԫ࣑ᱟᆖҐᇊѹ൘䘉⢙⨶ሩ䊑кⲴⴞḷᾲᘥ CupǄ㺘 12-3 ᧿䘠Ҷ Cup ⴞḷᾲᘥⲴ䇝㓳
ṧ઼ֻ亶ฏ⨶䇪Ǆ⌘᜿亶ฏ⨶䇪ѝᇊѹ Cup Ѫањ StableǃLiftable ԕ৺ OpenVessel Ⲵሩ䊑Ǆ
亶ฏ⨶䇪䘈ᢺ䘉 3 њ኎ᙗᇊѹѪᴤสᵜⲴ኎ᙗˈণ᧿䘠Ҷ↔ᇎֻⲴ৏ᆀⲴǃਟ᫽֌Ⲵ኎ᙗǄ⌘
᜿亶ฏ⨶䇪ᒦнᱟо䇝㓳ṧֻᆼ‫ޘ‬а㠤ⲴǄֻྲˈ亶ฏ⨶䇪䭉䈟ൠ࠶㊫ㅜ 2 ઼ㅜ 3 њ䇝㓳ṧֻ
Ѫ৽ֻǄн䗷ˈ亶ฏ⨶䇪ᖒᡀҶⴞḷᾲᘥⲴᴹ᭸䘁լǄKBANN ֯⭘亶ฏ⨶䇪઼䇝㓳ṧֻа䎧
ᆖҐⴞḷᾲᘥˈਟԕ∄অ⤜֯⭘ަѝа⿽ᴤ㋮⺞Ǆ
൘KBANN㇇⌅Ⲵㅜа䱦⇥˄㇇⌅ѝⲴ 1-3 ↕˅ˈᶴᔪҶањо亶ฏ⨶䇪а㠤Ⲵࡍ࿻㖁

㔌ǄֻྲˈӾCupⲴ亶ฏ⨶䇪ѝᶴᔪⲴ㖁㔌᧿㔈Ҿമ 12-2 ѝǄа㡜䈤ᶕˈ㖁㔌Ⲵᶴᔪᱟ䙊䗷ሩ
亶ฏ⨶䇪ѝ⇿аHornᆀਕᔪ・ањsigmoidঅ‫ݳ‬ǄKBANN䚥Ӿᜟֻˈsigmoid䗃ࠪ٬བྷҾ 0.5 ᰦ
㻛䀓䟺ѪⵏˈሿҾ 0.5 ࡉѪ‫ٷ‬Ǆഐ↔⇿њঅ‫Ⲵݳ‬ᶴᔪᯩ⌅Ѫ˖ᖃሩᓄⲴHornᆀਕᆈ൘ᰦˈঅ‫ݳ‬
Ⲵ䗃ࠪቡབྷҾ 0.5Ǆሩ↔HornᆀਕⲴ⇿њ‫ݸ‬㹼䇽ˈቡᔪ・ަሩᓄⲴsigmoidঅ‫֌ݳ‬Ѫ䗃‫ޕ‬Ǆ❦ਾ
䇮㖞sigmoidঅ‫Ⲵݳ‬ᵳ٬ˈ֯ަ䇑㇇ᗇࠪަ䗃‫Ⲵޕ‬䙫䗁оǄ⺞࠷ൠ䇢ˈሩҾ⇿њሩᓄҾ䶎䍏‫ݸ‬
㹼䇽Ⲵ䗃‫ˈޕ‬ᵳ٬㻛䇮㖞ѪḀ↓ᑨ䟿WǄሩ⇿њሩᓄҾ䍏‫ݸ‬㹼䇽Ⲵ䗃‫ˈޕ‬ᵳ٬䇮Ѫ-WǄঅ‫Ⲵݳ‬
䰸٬ᵳw0䇮Ѫ-(n-0.5)WˈަѝnѪ䶎䍏‫ݸ‬㹼䇽ⲴᮠⴞǄᖃঅ‫ݳ‬䗃‫ޕ‬٬Ѫ 1 ᡆ 0 ᰦˈ䘉‫؍‬䇱Ҷᖃ
фӵᖃᡰᴹⲴᆀਕ‫ݸ‬㹼䇽┑䏣ᰦˈ䗃‫࣐Ⲵޕ‬ᵳ઼࣐кw0 Ѫ↓˄㘼ф↔sigmoidⲴ䗃ࠪབྷҾ
0.5˅Ǆ⌘᜿ሩҾsigmoidঅ‫ˈݳ‬ㅜҼቲ৺ԕਾⲴቲѝঅ‫ݳ‬䗃‫ޕ‬наᇊѪ 1 ᡆ 0ˈк䶒Ⲵભ仈ᰐ
⌅ᓄ⭘Ҿ↔Ǆ❦㘼ྲ᷌ѪW䘹ᤙ䏣ཏབྷⲴ٬ˈ↔KBANN㇇⌅ਟԕሩԫ᜿␡ᓖⲴ㖁㔌䘋㹼亶ฏ
⨶䇪㕆⸱ǄTowell & Shavlik˄1994˅൘ަཊᮠᇎ傼ѝ֯⭘W˙4.0Ǆ
⇿њ sigmoid অ‫ݳ‬䗃‫ޕ‬㻛䘎᧕ࡠ䘲ᖃⲴ㖁㔌䗃‫ޕ‬ᡆਖа sigmoid অ‫Ⲵݳ‬䗃ࠪˈԕ৽᱐亶ฏ

⨶䇪ѝሩᓄ኎ᙗⲴ‫׍‬䎆‫ޣ‬㌫മǄᴰਾа↕ˈ৸䱴࣐Ҷ䇨ཊ䗃‫⇿ࡠޕ‬њ䰸٬অ‫ˈݳ‬ᆳԜⲴᵳ٬䇮
㖞䘁լѪ 0Ǆ䘉Ӌ䱴࣐䘎᧕Ⲵ֌⭘ᱟ‫ݱ‬䇨㖁㔌㜭ᖂ㓣ᆖҐࡠⲴ޵ᇩਟ䎵ࠪ亶ฏ⨶䇪ѝᨀࠪⲴ‫׍‬
䎆‫ޣ‬㌫Ǆമ 12-2 ѝⲴ㋇ᇎ㓯㺘᰾ᵳ٬Ѫ W Ⲵঅ‫ݳ‬䗃‫ˈޕ‬㘼㓶㓯㺘᰾ࡍ࿻ᵳ٬㓖Ѫ 0 Ⲵ䘎᧕Ǆ
ᖸᇩ᱃傼䇱ሩҾ䏣ཏབྷⲴ W ٬ˈ↔㖁㔌䗃ࠪ٬ㅹҾ亶ฏ⨶䇪Ⲵ亴⍻Ǆ
മ 12-2 ањㅹԧҾ亶ฏ⨶䇪Ⲵ⾎㓿㖁㔌
䘉њ㖁㔌൘ KBANN ㇇⌅Ⲵㅜа䱦⇥ࡋᔪࠪᶕˈᆳӗ⭏Ⲵ䗃ࠪ࠶㊫ㅹҾ㔉ᇊⲴ亶ฏ⨶䇪ѝⲴᆀਕ

‫࠶Ⲵڊ‬㊫Ǆ㋇㓯㺘⽪ᵳ٬Ѫ W Ⲵ䘎᧕ˈሩᓄ亶ฏ⨶䇪ѝⲴᆀਕ‫ݸ‬㹼䇽Ǆ㓶㓯㺘⽪ᵳ٬䘁լѪ 0 Ⲵ䘎
᧕Ǆ
KBANN ⲴㅜҼ䱦⇥˄㺘 12-2 ѝ㇇⌅Ⲵㅜ 4 ↕˅֯⭘䇝㓳ṧ઼ֻ৽ੁՐ᫝㇇⌅ᶕ㋮ॆࡍ࿻

㖁㔌ᵳ٬Ǆᖃ❦ˈྲ᷌亶ฏ⨶䇪઼䇝㓳ṧֻнवਜ਼ᐞ䭉ˈࡍ࿻Ⲵ㖁㔌ቡᐢ㓿ᤏਸ䇝㓳ᮠᦞҶǄ
❦㘼൘ Cup ֻᆀѝˈ亶ฏ⨶䇪о䇝㓳ᮠᦞна㠤ˈᡰԕ↔↕僔Պ᭩ਈࡍ࿻㖁㔌Ⲵᵳ٬ǄᗇࡠⲴ
䇝㓳䗷Ⲵ㖁㔌ᱮ⽪൘മ 12-3 ѝˈ㋇ᇎ㓯㺘᰾ᴰབྷⲴ↓ᵳ٬ˈ㋇㲊㓯㺘᰾ᴰབྷ䍏ᵳ٬ˈ㓶㓯㺘
᰾ਟᘭ⮕Ⲵᵳ٬Ǆ㲭❦ࡍ࿻㖁㔌䈟࠶㊫Ҷ㺘 12-3 ѝࠐњ䇝㓳ṧֻˈնമ 12-3 ѝ㋮ॆҶⲴ㖁㔌
㜭ᆼ㖾ൠ࠶㊫ᡰᴹ䇝㓳ֻǄ
ᴹᗵ㾱∄䖳алᴰ㓸ᖂ㓣㋮ॆⲴ㖁㔌ᵳ٬઼亶ฏ⨶䇪ሬࠪⲴࡍ࿻ᵳ٬Ǆྲമ 12-3 ᡰ㿱ˈ

൘ᖂ㓣↕ѝਁ⧠Ҷ‫ޘ‬ᯠⲴ‫׍‬䎆‫ޣ‬㌫ˈवᤜ Liftable অ‫ݳ‬ሩ MadeofStyrofoam Ⲵ‫׍‬䎆‫ޣ‬㌫Ǆᗵ享⢒
䇠ˈ㲭❦ḷᴹ Liftable Ⲵঅ‫ݳ‬ᴰࡍ⭡ᆳⲴ Horn ᆀਕᇊѹˈնਾᶕ⭡৽ੁՐ᫝‫؞‬᭩Ⲵᵳ٬ᐢ㓿ᆼ
‫ޘ‬᭩ਈҶ↔䳀㯿অ‫Ⲵݳ‬᜿ѹǄ൘㖁㔌㻛䇝㓳䗷ਾˈ䈕অ‫ݳ‬ਟ㜭ᴹҶоࡍ࿻Ⲵ Liftable 䇠ਧᰐ‫ޣ‬
Ⲵ䶎ᑨн਼Ⲵ᜿ѹǄ
മ 12-3 ሩࡍ࿻㖁㔌ᖂ㓣㋮ॆਾⲴ㔃᷌
KBANN ֯⭘䇝㓳ṧֻᶕ‫؞‬᭩Ӿ亶ฏ⨶䇪ѝሬࠪⲴ㖁㔌ᵳ٬Ǆ⌘᜿ަѝᯠӗ⭏Ⲵ Liftable ሩ

MadeOfStyrofoam ઼ HandleOnTop Ⲵ‫׍‬䎆ᙗǄ
12.3.3 䈪᱄
ᾲᤜൠ䇢ˈKBANN ⭘࠶᷀ⲴᯩᔿࡋᔪҶㅹԧҾ㔉ᇊ亶ฏ⨶䇪Ⲵ㖁㔌ˈ❦ਾᖂ㓣ൠ㋮ॆ↔
ࡍ࿻‫ٷ‬䇮ԕᴤྭൠᤏਸ䇝㓳ᮠᦞǄ൘↔䗷〻ѝˈᆳѪҶ᭩ਈ亶ฏ⨶䇪઼䇝㓳ᮠᦞна㠤Ⲵᛵߥ
㘼‫؞‬᭩㖁㔌ᵳ٬Ǆ
о㓟ᖂ㓣Ⲵ৽ੁՐ᫝˄ᆳᔰ࿻Ҿ䲿ᵪⲴᵳ٬˅⴨∄ˈKBANN Ⲵྭ༴൘Ҿˈᆳ൘㔉ᇊ䘁լ
↓⺞亶ฏ⨶䇪ᰦˈ㜭ཏ∄৽ੁՐ᫝ᴹᴤ儈Ⲵ⌋ॆ㋮ᓖˈ⢩࡛ᱟ൘䇝㓳ᮠᦞ〰㕪ᰦǄ൘ࠐ⿽ᇎ䱵
㌫㔏ѝˈKBANN ઼ަԆࡍ࿻ॆ‫ٷ‬䇮Ⲵ䙄ᖴᐢᱮ⽪ࠪՈҾ㓟ᖂ㓣Ⲵ㌫㔏ǄֻྲˈTowell et al.
˄1990˅᧿䘠Ҷሶ KBANN ᓄ⭘Ҿ࠶ᆀ䚇Ր䰞仈ǄަѝⲴԫ࣑ᱟᆖҐ䇶࡛〠Ѫ◰ਁ४ฏ
˄promoter region˅Ⲵ DNA ⡷ᯝˈᆳᖡ૽สഐⲴ⍫ᙗǄ൘↔ᇎ傼ѝˈKBANN Ⲵ亶ฏ⨶䇪Ӿа
њ࠶᷀䚇Րᆖᇦ䛓䟼㧧ਆˈ㘼◰ਁ४ฏⲴ䇝㓳ṧֻѝवਜ਼ 53 њ↓઼ֻ 53 њ৽ֻǄᙗ㜭䇴ՠ֯
⭘ҶĀ⮉а⌅ā˄leave-one-out˅ˈ㌫㔏䘀㹼 106 ⅑Ǆ⇿⅑ᗚ⧟ѝ KBANN ⭘ 105 њṧֻ䇝
㓳ˈᒦ൘࢙։Ⲵ 1 њк⍻䈅Ǆ䘉 106 ⅑ᇎ傼Ⲵ㔃᷌㻛〟㚊䎧ᶕᨀ‫׋‬ሩⵏᇎ䭉䈟⦷Ⲵՠ䇑Ǆ
KBANN 䭉䈟⦷Ѫ 4ˋ106ˈ㘼ḷ߶Ⲵ৽ੁՐ᫝䭉䈟⦷Ѫ 8ˋ106ǄKBANN Ⲵањਈ⿽⭡ Fu
˄1993˅ᇎ⧠ˈᆳᣕ੺൘਼ṧᮠᦞкⲴ䭉䈟⦷Ѫ 2ˋ106Ǆഐ↔ˈ‫ݸ‬傼⸕䇶൘䘉Ӌᇎ傼ѝᖸབྷ
〻 ᓖ ൠ ߿ ሿ Ҷ 䭉䈟 ⦷ Ǆ ↔ ᇎ 傼 Ⲵ 䇝㓳 ᮠ ᦞ ਟ ԕ Ӿ з 㔤㖁 ൰
http://www.ics.uci.edu/~mlearn/MLReository.html кᗇࡠǄ
Fu˄1993˅઼ Towell et al.˄1990˅䜭ᣕ੺˖Ӿᴰ㓸䇝㓳䗷Ⲵ㖁㔌ѝᣭਆⲴ Horn ᆀਕˈਟ

ᨀ‫׋‬ањ㜭ᴤྭᤏਸ䇝㓳ᮠᦞⲴ亶ฏ⨶䇪Ǆ㲭❦ᴹᰦਟ㜭ӾᆖҐࡠⲴ㖁㔌ᵳ٬᱐ሴഎањ㋮ॆ
Ⲵ Horn ᆀਕ䳶ˈն൘а㡜ᛵᖒл䘉⿽֌⌅ᱟᴹ䰞仈ⲴǄഐѪḀӋᵳ٬䇮㖞⋑ᴹⴤ᧕ሩᓄⲴ
Horn ᆀਕǄCraven & Shavlik˄1994˅઼ Craven˄1996˅᧿䘠ҶਖཆⲴᯩ⌅ԕӾᆖҐ䗷Ⲵ㖁㔌
ѝᣭਆㅖਧ㿴ࡉǄ
Ѫ⨶䀓 KBANN Ⲵᇊѹˈᴹᗵ㾱㘳㲁ަѝⲴ‫ٷ‬䇮ᩌ㍒о㓟ᖂ㓣Ⲵ৽ੁՐ᫝㇇⌅ѝᴹӰѸ४

࡛Ǆ䘉є⿽㇇⌅ѝᢗ㹼Ⲵ‫ٷ‬䇮オ䰤ᩌ㍒൘മ 12-4 ѝ⽪᜿Ǆྲަѝᱮ⽪Ⲵˈ‫ޣ‬䭞४࡛൘Ҿᢗ㹼
ᵳ٬䈳㢲ᡰสҾⲴࡍ࿻‫ٷ‬䇮Ǆ൘ᴹཊњ‫ٷ‬䇮˄ᵳ٬ੁ䟿˅㜭ᤏਸᮠᦞⲴᛵߥл˄䘉⿽ᛵߥ൘䇝
㓳ᮠᦞ〰㕪ᰦᴤਟ㜭ࠪ⧠˅ˈKBANN ᴤᴹਟ㜭᭦ᮋࡠ䘉ṧⲴ‫ٷ‬䇮ˈᆳӾ䇝㓳ᮠᦞѝⲴ⌋ॆо
亶ฏ⨶䇪Ⲵ亴⍻ᴤ⴨լǄਖаᯩ䶒ˈ৽ੁՐ᫝᭦ᮋࡠⲴ⢩ᇊ‫ٷ‬䇮ᴤਟ㜭ᱟሿᵳ٬Ⲵ‫ٷ‬䇮ˈᆳབྷ
㠤ሩᓄҾ൘䇝㓳ṧֻ䰤ᒣ━ᨂ٬Ⲵ⌋ॆ‫ٿ‬㖞Ǆㆰ㾱ൠ䈤ˈKBANN ֯⭘ањ亶ฏ⢩ᇊⲴ⨶䇪ᶕ
‫ٿ‬㖞⌋ॆˈ㘼৽ੁՐ᫝㇇⌅֯⭘ањ亶ฏᰐ‫Ⲵޣ‬䈝⌅‫ٿ‬㖞˄‫ੁٿ‬ҾሿⲴᵳ٬˅Ǆ⌘᜿൘↔ᾲ䘠
ѝᡁԜᘭ⮕Ҷᩌ㍒ѝተ䜘ᶱሿ٬Ⲵᖡ૽Ǆ
Hypothesis Space˖ ‫ٷ‬䇮オ䰤
Hypotheses that fit training data equally well˖ԕ਼ṧ〻ᓖᤏਸ䇝㓳ᮠᦞⲴ‫ٷ‬䇮
Initial hypothesis for KBANN˖ KBANN Ⲵࡍ࿻‫ٷ‬䇮
Initial hypothesis for Backpropagation˖ ৽ੁՐ᫝Ⲵࡍ࿻‫ٷ‬䇮
മ 12-4KBANN ѝⲴ‫ٷ‬䇮オ䰤ᩌ㍒
KBANN ࡍ࿻ॆ㖁㔌֯ަᤏਸ亶ฏ⨶䇪ˈ㘼৽ੁՐ᫝ሶ㖁㔌ࡍ࿻ॆѪ䲿ᵪሿᵳ٬Ǆ❦ਾᆳԜ֯⭘
⴨਼Ⲵởᓖл䱽㿴ࡉ৽༽㋮ॆᵳ٬Ǆᖃ᢮ࡠཊњ㜭ᤏਸ䇝㓳ᮠᦞⲴ‫ٷ‬䇮ᰦ˄ྲ䱤ᖡ४ฏᡰ⽪˅ˈ
KBANN ઼৽ੁՐ᫝ਟ㜭᢮ࡠнаṧⲴ‫ٷ‬䇮ˈഐѪᆳԜⲴ䎧⛩н਼Ǆ
KBANN Ⲵተ䲀ᙗѻаѪˈᆳਚ㜭֯⭘ભ仈亶ฏ⨶䇪ˈণᰐਈ䟿Ⲵ Horn ᆀਕ䳶Ǆྲ᷌㔉

Ҹᖸн㋮⺞Ⲵ亶ฏ⨶䇪ˈKBANN ҏਟ㜭㻛䈟ሬˈӾ㘼ަ⌋ॆ㋮ᓖਈᗇվҾ৽ੁՐ᫝Ǆн䗷ˈ
KBANN ઼⴨‫ޣ‬㇇⌅⺞ᇎ൘㤕ᒢᇎ䱵䰞仈ѝᱮ⽪ࠪᴹࣙҾᆖҐǄ
KBANN ᱟ㔃ਸ࠶઼᷀ᖂ㓣ᆖҐⲴࡍ࿻ॆ‫ٷ‬䇮䙄ᖴѝⲴа⿽Ǆ䘉а䙄ᖴⲴަԆֻᆀवᤜ Fu
˄1993˅˗Gallant(1988)˗Bradshaw et al.˄1989˅˗Yang & Bhargava(1990)˗Lacher et al.
˄1991˅Ǆ䘉Ӌ䙄ᖴн਼ѻ༴൘Ҿᔪ・ࡍ࿻‫ٷ‬䇮Ⲵᇎ䱵֯⭘Ⲵᢰᵟǃᵳ٬䈳ᮤⲴ৽ੁՐ᫝Ⲵᓄ
⭘ǃԕ৺Ӿ㋮ॆҶⲴ㖁㔌ѝᣭਆㅖਧ᧿䘠Ⲵᯩ⌅ǄPratt˄1993aˈ1993b˅᧿䘠Ⲵањࡍ࿻ॆ‫ٷ‬
䇮䙄ᖴѝˈ‫ݸ‬傼⸕䇶ᱟ䙊䗷‫ࡽݸ‬ሩ⴨‫ޣ‬ԫ࣑ᆖҐࡠⲴ⾎㓿㖁㔌ᶕᨀ‫Ⲵ׋‬Ǆ䇝㓳䍍ਦᯟ㖞ؑ㖁Ⲵ
٬Ⲵᯩ⌅˄ྲ 6.11 㢲ѝ䇘䇪Ⲵ˅ҏਟ㻛ⴻ֌ᱟ⭘‫ݸ‬傼⸕䇶ᶕࡍ࿻ॆ‫ٷ‬䇮Ǆ䘉䟼‫ݸ‬傼⸕䇶ሩᓄ
Ҿа㓴ᶑԦ⤜・ᙗ‫ٷ‬ᇊˈᆳ⺞ᇊҶ䍍ਦᯟ㖁Ⲵമ㔃ᶴˈ❦ਾަᶑԦᾲ⦷㺘Ӿ䇝㓳ᮠᦞѝᖂ㓣ᗇ
ࡠǄ
12.4 ֵ⭞‫ݾ‬僂⸛䇼᭯਎ᩒ㍘ⴤḽ
к䶒Ⲵ䙄ᖴ⭡ањᆼ㖾ᤏਸ亶ฏ⨶䇪Ⲵ‫ٷ‬䇮ᔰ࿻ởᓖл䱽ᩌ㍒ˈ❦ਾ൘䴰㾱ᰦ᭩ਈ↔‫ٷ‬䇮
ԕᴰབྷ〻ᓖൠᤏਸ䇝㓳ᮠᦞǄ֯⭘‫ݸ‬傼⸕䇶Ⲵਖаᯩ⌅ᱟሶᆳਸᒦࡠởᓖл䱽ѝ䴰ᴰሿॆⲴ䈟
ᐞࡔᦞˈ䘉ṧ㖁㔌享ᤏਸⲴᱟ䇝㓳ᮠᦞ઼亶ฏ⨶䇪Ⲵ㓴ਸ࠭ᮠǄ⺞࠷ൠ䇢ˈᡁԜ㘳㲁Ⲵ‫ݸ‬傼⸕
䇶Ⲵᖒᔿᱟⴞḷ࠭ᮠⲴḀ⿽ᐢ⸕ⲴሬࠪᔿǄаӋ㊫රⲴ‫ݸ‬傼⸕䇶ਟԕᖸ㠚❦ൠ⭘↔ᖒᔿ㺘⽪Ǆ
ֻྲˈ൘䇝㓳ањ⾎㓿㖁㔌ԕ䇶࡛᡻߉ᆇㅖᰦˈᡁԜਟԕᤷᇊⴞḷ࠭ᮠⲴḀ⿽ሬᮠˈԕ㺘⽪䘉
⿽Ⲵ‫ݸ‬傼⸕䇶˖ĀᆇㅖⲴ⺞䇔⤜・Ҿമ‫Ⲵۿ‬ᗞሿᒣ઼〫᯻䖜Ǆā
л䶒᧿䘠Ⲵ TangentProp ㇇⌅䇝㓳⾎㓿㖁㔌ˈ֯ަ਼ᰦᤏਸ䇝㓳٬઼䇝㓳ሬᮠǄ12.4.4 㢲

䈤᰾Ҷᘾṧ⭘㊫լҾ 12.3 㢲֯⭘Ⲵ Cup ֻᆀѝⲴᯩ⌅Ӿ亶ฏ⨶䇪ѝ㧧ᗇ䘉Ӌ䇝㓳ሬᮠǄ⺞࠷
ൠ䇢ˈᆳ䇘䇪Ҷ EBNN ㇇⌅ᘾṧᶴ䙐অ⤜ṧֻⲴ䀓䟺ˈԕᣭਆࠪ䇝㓳ሬᮠᶕ‫ ׋‬TangentProp ֯
⭘ǄTangentProp ઼ EBNN ᐢ൘ཊњ亶ฏѝ㻛⽪㤳ࠪᴹՈҾ㓟ᖂ㓣ᯩ⌅Ⲵᙗ㜭ˈवᤜᆇㅖ઼⢙
փ䇶࡛ˈԕ৺ᵪಘӪᝏ⸕઼᧗ࡦԫ࣑ѝǄ
12.4.1 TangentProp ㇍⌋
TangentProp ˄Simard et al. 1992˅᧕ਇⲴ亶ฏ⸕䇶㻛㺘⽪ѪሩᓄҾަ䗃‫ޕ‬ਈᦒⲴⴞḷ࠭ᮠ

ⲴሬᮠǄ㘳㲁ањᆖҐԫ࣑ˈवਜ਼ањᇎֻオ䰤X઼ⴞḷ࠭ᮠfǄ㠣↔ᡁԜᡰ֌Ⲵ‫ٷ‬ᇊѝ⇿њ䇝
㓳ṧֻᖒᔿѪ<xi,f(xi)>ˈᆳ᧿䘠ҶḀᇎֻxi઼ަ䇝㓳٬f(xi)ǄTangentProp㇇⌅䘈‫ٷ‬ᇊᨀ‫׋‬Ҷⴞḷ
࠭ᮠⲴн਼Ⲵ䇝㓳ሬᮠ˄training derivative˅Ǆֻྲˈྲ᷌⇿њᇎֻxi᧿䘠Ѫањᇎᮠˈ䛓Ѹ⇿
wf ( x) wf ( x )
њ䇝㓳ṧֻᖒᔿਟ㜭Ѫ ¢ xi , f ( xi ), ² Ǆ䘉䟼㺘⽪ⴞḷ࠭ᮠ൘⛩x=xi кሩxⲴሬ
wx xi wx xi
ᮠǄ
ѪҶӾⴤ㿹к⨶䀓൘ᆖҐѝнӵᨀ‫׋‬䇝㓳٬ҏᨀ‫׋‬䇝㓳ሬᮠⲴྭ༴ˈ㘳㲁ањㆰঅⲴԫ
࣑ˈ൘മ 12-5 㺘⽪ǄަѝᖵᆖҐⲴⴞḷ࠭ᮠfᱮ⽪Ҿަѝᴰᐖ䗩ⲴമᖒˈᆳสҾᡰᱮ⽪Ⲵ 3 њ
䇝㓳ṧֻ<x1, f(x1)>ˈ<x2, f(x2)>઼<x3, f(x3)>ǄᴹҶ䘉 3 њṧֻˈ৽ੁՐ᫝㇇⌅ਟᵋᗇࡠањᒣ
━࠭ᮠ‫ٷ‬䇮ˈྲѝ䰤മᱮ⽪Ⲵ࠭ᮠgǄᴰਣ䗩Ⲵമᱮ⽪Ҷᨀ‫׋‬䇝㓳ሬᮠ˄ᡆᯌ⦷˅֌Ѫ⇿њ䇝
wf ( x)
㓳ṧֻⲴ䱴࣐ؑ᚟˄ྲ ¢ x1 , f ( x1 ), ² ˅Ⲵ᭸᷌Ǆ䙊䗷ᤏਸ䇝㓳٬f(xi)਼ᰦᤏਸ䘉Ӌሬᮠ
wx x1
wf ( x)
ˈᆖҐಘਟԕᴤྭൠӾ〰⮿䇝㓳ᮠᦞѝ↓⺞⌋ॆǄᾲᤜൠ䈤ˈवਜ਼䇝㓳ሬᮠⲴ᭸᷌ᱟ
wx xi
ѪҶ‫ݻ‬ᴽ৽ੁՐ᫝ѝⲴ䈝⌅ᖂ㓣‫ٿ‬㖞˄ᆳ‫ྭٿ‬਴⛩䰤Ⲵᒣ━ᨂ٬˅ˈሶަᴯᦒѪᡰᐼᵋⲴሬᮠ
Ⲵᱮᔿ䗃‫ؑޕ‬᚟Ǆ㔃᷌‫ٷ‬䇮hᱮ⽪൘ᴰਣ䗩Ⲵമѝˈᆳᨀ‫׋‬Ҷሩⵏᇎⴞḷ࠭ᮠfⲴᴤ㋮⺞ՠ䇑Ǆ
മ 12-5 ⭘ TangentProp ᤏਸ٬઼ሬᮠ
ԔfѪⴞḷ࠭ᮠˈᆳⲴйњṧֻ<x1, f(x1)>ˈ<x2, f(x2)>઼<x3, f(x3)>ᐢ⸕ǄสҾ䘉Ӌ⛩ˈᆖҐಘਟ㜭

⭏ᡀ‫ٷ‬䇮gǄྲ᷌ሬᮠҏᐢ⸕ˈᆖҐಘਟԕ⌋ॆࡠᴤ㋮⺞ⲴhǄ
൘к䘠ֻᆀѝˈᡁԜਚ㘳㲁Ҷㆰঅ㊫රⲴⴞḷ࠭ᮠሬᮠǄᇎ䱵кˈTangentPropਟ᧕ਇሩᓄ
Ҿ䗃‫ޕ‬xⲴ਴⿽ਈᦒⲴ䇝㓳ሬᮠǄֻྲˈ㘳㲁ᆖҐ䇶࡛᡻߉ᆇㅖⲴԫ࣑Ǆ⺞࠷ൠ䇢ˈ‫ٷ‬ᇊ䗃‫ޕ‬x
ሩᓄҾवਜ਼অњ᡻߉ᆇㅖⲴമ‫ˈۿ‬㘼ԫ࣑ᱟ↓⺞࠶㊫↔ᆇㅖǄ൘↔ԫ࣑ѝˈᡁԜਟ㜭ᐼᵋ੺䇹
ᆖҐಘĀⴞḷ࠭ᮠሩҾമ‫ۿ‬ѝᆇㅖⲴᗞሿ᯻䖜нਇᖡ૽āǄѪ䗃‫ݸ↔ޕ‬傼⸕䇶㔉ᆖҐಘˈᡁԜ
俆‫ݸ‬ᇊѹањਈᦒs(¢,x)ˈᆳᢺമ‫ۿ‬x᯻䖜¢ᓖǄ⧠൘ᡁԜਟᢺ᯻䖜нਈᙗⲴᯝ䀰ྲл㺘⽪˖
ሩ⇿њ䇝㓳ᇎֻxi ˈⴞḷ࠭ᮠሩᓄ↔ਈᦒⲴሬᮠѪ 0˄ণ᯻䖜䗃‫ޕ‬മ‫ۿ‬н᭩ਈⴞḷ࠭ᮠⲴ
٬˅Ǆᦒ䀰ѻˈᡁԜਟሩ⇿њ䇝㓳ᇎֻxiᯝ䀰л䶒Ⲵ䇝㓳ሬᮠ˖
wf ( s (D , xi ))
0
wD
ަѝfѪⴞḷ࠭ᮠˈ㘼s(¢,xi)Ѫᓄ⭘↔ਈᦒsࡠമ‫ۿ‬xiᗇࡠⲴമ‫ۿ‬Ǆ
䘉ṧⲴ䇝㓳ሬᮠ൘ TangentProp ѝᘾṧ㻛⭘Ҿ㓖ᶏ⾎㓿㖁㔌Ⲵᵳ٬˛൘ TangentProp ѝ䘉Ӌ

䇝㓳ሬᮠ㻛ਸᒦࡠởᓖл䱽ѝ享ᴰሿॆⲴ䈟ᐞ࠭ᮠѝˈഎᗶㅜ 4 ㄐѝ৽ੁՐ᫝㇇⌅ᢗ㹼ởᓖл
䱽䈅മ֯䈟ᐞᒣᯩ઼ᴰሿॆ˖
E ¦ ( f ( x ) fˆ ( x ))
i
i i
2
ަѝxiԓ㺘ㅜiњ䇝㓳ᇎֻˈf ԓ㺘ⵏᇎⴞḷ٬ˈ㘼 fˆ ԓ㺘ᆖҐࡠⲴ⾎㓿㖁㔌㺘⽪Ⲵ࠭ᮠǄ
൘TangentPropѝˈ䈟ᐞ࠭ᮠѝᯠ໎Ҷа亩ԕ᜙㖊䇝㓳ሬᮠ઼ᆖҐࡠⲴ⾎㓿㖁㔌࠭ᮠⲴᇎ䱵
ሬᮠ fˆ ѻ䰤Ⲵ࠶↗Ǆа㡜ൠˈTangentPropਟ᧕ਇཊњਈᦒ˄ֻྲˈᡁԜᐼᵋᯝ䀰᯻䖜нਈ
ᙗˈ਼ᰦᯝ䀰ᆇㅖ䇶࡛ѝⲴᒣ〫нਈᙗ˅ˈ⇿њਈᦒᖒᔿᗵ享Ѫsj(¢,x)ˈަѝ¢Ѫ䘎㔝৲ᮠˈ
㘼sj ਟᗞˈ㘼фsj(0,x)=x˄ֻྲሩҾ 0 ᓖⲴ᯻䖜ˈ࠭ᮠণѪᚂㅹ࠭ᮠ˅Ǆሩ⇿њ䘉ṧⲴਈᦒ
sj(¢,x)ˈTangentProp㘳㲁ᤷᇊⲴ䇝㓳ሬᮠ઼ᆖҐࡠⲴ⾎㓿㖁㔌Ⲵᇎ䱵ሬᮠ䰤Ⲵ䈟ᐞᒣᯩǄ‫؞‬᭩
ਾⲴ䈟ᐞ࠭ᮠѪ˖
ª § wf ( s (D , x )) wfˆ ( s (D , x )) ·
2
º
¦i ««( f ( xi ) fˆ ( xi )) 2 P ¦j ¨¨ wD ¸ » ˄12ˊ1˅
j i j i
E
wD ¸ »
¬ © ¹ D 0¼
ަѝѪ⭘ᡧᨀ‫Ⲵ׋‬ᑨ䟿ˈԕ⺞ᇊᤏਸ䇝㓳ᮠᦞ઼ᤏਸ䇝㓳ሬᮠѻ䰤Ⲵ⴨ሩ䟽㾱ᙗǄ⌘᜿
E ᇊѹѝㅜа亩Ѫ৏ᶕⲴ䇝㓳ᮠᦞ਼㖁㔌ѻ䰤Ⲵ䈟ᐞᒣᯩˈ㘼ㅜҼ亩Ѫ䇝㓳ሬᮠ਼㖁㔌ѻ䰤Ⲵ
䈟ᐞᒣᯩǄ
Simard et al. ˄1992˅㔉ࠪҶ֯↔ᢙኅⲴ䈟ᐞ࠭ᮠᴰሿॆⲴởᓖл䱽㿴ࡉǄᆳਟ⭡㊫լҾ

ㅜ 4 ㄐѝ৽ੁՐ᫝㿴ࡉѝⲴᯩ⌅䘋㹼᧘ሬǄ
12.4.2 ⽰ׁ
Simard ㅹᨀ‫׋‬Ҷ TangentProp Ⲵ⌋ॆ㋮ᓖ਼㓟ᖂ㓣৽ੁՐ᫝ѻ䰤Ⲵ∄䖳㔃᷌ˈ䪸ሩⲴ䰞仈

Ѫ᡻߉ᆇㅖ䇶࡛Ǆᴤ⺞࠷ൠ䇢ˈ䘉䟼Ⲵԫ࣑ᱟѪঅњᮠᆇ 0 ࡠ 9 Ⲵമ‫֌ۿ‬ḷ䇠Ǆ൘ањᇎ傼
ѝˈTangentProp ઼৽ੁՐ᫝䜭⭘н਼བྷሿⲴ䇝㓳䳶䘋㹼䇝㓳ˈ❦ਾสҾᆳԜ൘⤜・Ⲵ 160 њ
ṧֻⲴ⍻䈅䳶к䇴ՠᙗ㜭Ǆ㔉Ҹ TangentProp Ⲵ‫ݸ‬傼⸕䇶Ѫ˖ᮠᆇⲴ࠶㊫нഐമ‫≤Ⲵۿ‬ᒣ઼඲
ⴤᒣ〫㘼᭩ਈ˄ণ↔ⴞḷ࠭ᮠሩᓄҾ䘉ӋਈᦒⲴሬᮠѪ 0˅Ǆ㔃᷌ᱮ⽪൘㺘 12-4 ѝˈ䇱᰾Ҷ
TangentProp ֯⭘‫ݸ‬傼⸕䇶Ⲵ⌋ॆ㋮ᓖ⺞ᇎ儈Ҿ㓟৽ੁՐ᫝㇇⌅Ǆ
㺘 12-4TangentProp ઼৽ੁՐ᫝Ⲵ⌋ॆ㋮ᓖˈ䪸ሩ᡻߉ᮠᆇ䇶࡛䰞仈
TangentProp ⌋ॆ㋮ᓖᴤ儈ˈഐѪᆳᴹ‫ݸ‬傼⸕䇶˖ᮠᆇⲴ⺞ᇊᴹᒣ〫нਈᙗǄ䘉Ӌ㔃᷌ᶕ㠚Ҿ
Simard et al.˄1992˅Ǆ
䇝㓳䳶 ൘⍻䈅䳶кⲴ䭉䈟⦷Ⲯ࠶∄
བྷሿ TangentProp ৽ੁՐ᫝
10 34 48
20 17 33
40 7 18
80 4 10
160 0 3
320 0 0
12.4.3 䈪᱄
ᾲᤜൠ䈤ˈTangentProp ֯⭘Ⲵ‫ݸ‬傼⸕䇶ᖒᔿѪⴞḷ࠭ᮠሩᓄަ䗃‫ޕ‬ਈᦒⲴᡰᐼᵋⲴሬ
ᮠǄᆳ䙊䗷֯ањⴞḷ࠭ᮠᴰሿॆᶕ㔃ਸ‫ݸ‬傼⸕䇶઼㿲ሏࡠⲴ䇝㓳ᮠᦞˈⴞḷ࠭ᮠ਼ᰦᓖ䟿Ҷ
㖁㔌ሩᓄ䇝㓳ṧֻ٬Ⲵ䈟ᐞ˄ᤏਸᮠᦞ˅ˈ઼㖁㔌ሩᓄҾሬᮠⲴ䈟ᐞ˄ᤏਸ‫ݸ‬傼⸕䇶˅ǄⲴ
٬ߣᇊҶ㖁㔌൘ᮤњ䈟ᐞѝᤏਸ䘉є䜘࠶Ⲵ〻ᓖǄ㇇⌅Ⲵ㹼Ѫሩ٬᭿ᝏˈᆳᱟ⭡䇮䇑㘵䘹ᤙ
ⲴǄ
㲭❦ TangentProp ᡀ࣏ൠ㔃ਸҶ‫ݸ‬傼⸕䇶઼䇝㓳ᮠᮠᦞԕᤷሬ⾎㓿㖁㔌ᆖҐˈնᆳሩҾ‫ݸ‬
傼⸕䇶ѝⲴ䭉䈟励ἂᙗнᕪǄᖃ‫ݸ‬傼⸕䇶н↓⺞ᰦˈণ䗃‫ࡠޕ‬ᆖҐಘⲴ䇝㓳ሬᮠн㜭↓⺞৽᱐
ⵏᇎⴞḷ࠭ᮠⲴሬᮠᰦˈ㇇⌅ሶ䈅മᤏਸн↓⺞ⲴሬᮠˈӾ㘼ሬ㠤⌋ॆ㋮ᓖнྲᆼ‫ޘ‬ᘭ⮕‫ݸ‬傼
⸕䇶֯⭘㓟৽ੁՐ᫝㇇⌅Ⲵ㋮ᓖǄྲ᷌ᡁԜ亴‫⸕ݸ‬䚃䇝㓳ሬᮠѝ䭉䈟ࠪ⧠〻ᓖˈᡁԜਟ⭘䘉а
ؑ᚟䘹ᤙᑨ䟿ˈԕ⺞ᇊᤏਸ䇝㓳٬઼ᤏਸ䇝㓳ሬᮠⲴ⴨ሩ䟽㾱〻ᓖǄ❦㘼ˈ䘉аؑ᚟нཚਟ
㜭亴‫⸕ݸ‬䚃Ǆ൘ла㢲ᡁԜ䇘䇪Ҷ EBNN ㇇⌅ˈᆳਟ㠚ࣘṩᦞ example-by-example Ⲵส⹰䘹ᤙ
Ⲵ٬ˈԕ䀓ߣн↓⺞Ⲵ‫ݸ‬傼⸕䇶Ⲵ䰞仈Ǆ
ᴹᗵ㾱∄䖳ал TangentPropǃKBANN ઼৽ੁՐ᫝ᢗ㹼Ⲵ‫ٷ‬䇮オ䰤˄ᵳ٬オ䰤˅Ⲵᩌ㍒ᯩ

⌅ǄTangentProp 㔃ਸ‫ݸ‬傼⸕䇶ˈ䙊䗷᭩ਈ⭡ởᓖл䱽ᴰሿॆⲴⴞⲴ(objective)࠭ᮠᶕᖡ૽‫ٷ‬䇮
ᩌ㍒Ǆᆳ⴨ᖃҾ᭩ਈҶ‫ٷ‬䇮オ䰤ᩌ㍒Ⲵⴞḷˈྲമ 12-6 ᡰ⽪Ǆྲ৽ੁՐ᫝㇇⌅аṧ˄նо
KBANN н਼˅ˈTangentProp ᔰ࿻Ҿ䲿ᵪሿᵳ٬Ⲵࡍ࿻㖁㔌Ǆ❦㘼ˈᆳⲴởᓖ䇝㓳⌅ࡉӗ⭏Ⲵ
ᵳ٬ᴤᯠо৽ੁՐ᫝Ⲵн਼ˈӾ㘼ᗇࡠн਼Ⲵᴰ㓸‫ٷ‬䇮Ǆྲമѝᡰ⽪ˈ֯ TangentProp ⲴⴞⲴ
࠭ᮠᴰሿॆⲴ‫ٷ‬䇮䳶ਸн਼Ҿ֯৽ੁՐ᫝ⲴⴞⲴ࠭ᮠᴰሿॆⲴ‫ٷ‬䇮䳶ਸˈ䟽㾱Ⲵᱟˈྲ᷌䇝㓳
ṧ઼ֻ‫ݸ‬傼⸕䇶䜭↓⺞ˈᒦфⴞḷ࠭ᮠਟ⭘ ANN ㋮⺞㺘⽪ˈ䛓Ѹ┑䏣 TangentProp ⴞḷⲴᵳ
ੁ䟿䳶ਸሶѪ┑䏣৽ੁՐ᫝ⴞḷⲴᵳੁ䟿䳶ਸⲴᆀ䳶Ǆ䘉єњᴰ㓸‫ٷ‬䇮Ⲵ䳶ਸⲴᐞ࡛ѪаӋн
↓⺞Ⲵ‫ٷ‬䇮ˈᆳԜՊ㻛৽ੁՐ᫝㘳㲁ˈնՊഐѪ‫ݸ‬傼⸕䇶㘼㻛 TangentProp ࢄ䲔ᦹǄ
Hypothesis Space˖ ‫ٷ‬䇮オ䰤
Hypotheses that maximize fit to data and prior knowledge: ሩᮠᦞ઼‫ݸ‬傼⸕䇶䜭ᴹᴰབྷᤏ
ਸᓖⲴ‫ٷ‬䇮
Hypotheses that maximize fit to data ሩᮠᦞᤏਸᓖᴰབྷⲴ‫ٷ‬䇮
TangentProp Search TangentProp ᩌ㍒
Backpropagation Search ৽ੁՐ᫝ᩌ㍒
മ 12-6TangentProp ѝⲴ‫ٷ‬䇮オ䰤ᩌ㍒
TangentProp ሶ㖁㔌ࡍ࿻ॆѪ䲿ᵪሿᵳ٬ˈྲ৽ੁՐ᫝ѝаṧǄ❦㘼ˈᆳ֯⭘н਼Ⲵ䈟ᐞ࠭ᮠᶕ
ᕅሬởᓖл䱽ᩌ㍒ǄTangentProp ѝ֯⭘Ⲵ䈟ᐞवᤜҶ亴⍻䇝㓳٬Ⲵ䈟ᐞˈҏवᤜ亴⍻⭡‫ݸ‬傼⸕䇶ᨀ‫׋‬
Ⲵ䇝㓳ሬᮠⲴ䈟ᐞǄ
⌘᜿ˈሩⴞḷ࠭ᮠⲴ䇝㓳ሬᮠᤏਸⲴਖа⿽ᯩ⌅ᱟˈㆰঅൠሶ㿲ሏࡠⲴ䇝㓳ṧֻ䱴䘁Ⲵ䱴
࣐䇝㓳ṧֻ㔬ਸ䎧ᶕˈ֯⭘ᐢ⸕Ⲵ䇝㓳ሬᮠᶕՠ䇑䘉Ӌ䱴䘁ⲴᇎֻⲴ䇝㓳٬Ǆֻྲˈ൘к䶒Ⲵ
ᆇㅖ䇶࡛ԫ࣑ѝˈਟԕਆањ䇝㓳മ‫ˈۿ‬ሩަቁ䟿Ⲵᒣ〫ˈ❦ਾᯝ䀰ᒣ〫ਾⲴമ‫ۿ‬о৏ᶕⲴṧ
ֻ኎Ҿ਼а㊫Ǆਟԕᵏᵋ֯⭘৽ੁՐ᫝઼䘉Ӌ㔬ਸⲴṧֻˈ㜭ᗇࡠ⴨լҾ TangentProp ѝ֯⭘
৏࿻ṧ઼ֻሬᮠᡰᗇࡠⲴ㔃᷌ǄSimard et al.˄1992˅֌Ⲵᇎ傼ᱮ⽪є⿽ᛵߥлᴹ䘁լ⴨ㅹⲴ⌋
ॆ䭉䈟⦷ˈն TangentProp 㜭ᴤѪᴹ᭸ൠ᭦ᮋǄᴹ᜿ᙍⲴᱟㅜ 4 ㄐᨀࡠⲴᆖҐ傮傦⊭䖖Ⲵ
ALVINN ㌫㔏ˈ֯⭘Ҷᖸ⴨լⲴ䙄ᖴԕ㔬ਸ䱴࣐䇝㓳ṧֻǄᆳ֯⭘ᴹ‫ྲޣ‬օṩᦞ䮌ཤമ‫≤Ⲵۿ‬
ᒣᒣ〫ᶕ᭩ਈ傮傦ᯩੁⲴ‫ݸ‬傼⸕䇶ˈᶕࡋᔪཊњ㔬ਸⲴ䇝㓳ṧֻԕᢙ‫⇿ݵ‬њ㿲ሏࡠⲴ䇝㓳ṧ
ֻǄ
12.4.4 EBNN ㇍⌋
EBNN(Explanation-Based neural network)ণสҾ䀓䟺Ⲵ⾎㓿㖁㔌˄㿱 Mitchell & Thrun

1993a; Thrun 1996˅ˈ䘉⿽㇇⌅ԕє⿽ᯩᔿ᭩䘋Ҷ TangentProp ㇇⌅Ǆ俆‫ˈݸ‬ᆳн‫׍‬䶐⭘ᡧᨀ
‫׋‬䇝㓳ሬᮠˈ㘼ᱟሩ⇿њ䇝㓳ṧֻ㠚㹼䇑㇇↔䇝㓳ሬᮠǄ䇑㇇ᯩ⌅ᱟ䙊䗷⭘а྇㔉ᇊⲴ亶ฏ⨶
䇪ᶕ䀓䟺⇿њ䇝㓳ṧֻǄަ⅑ˈEBANN ⎹৺Ҷྲօ⺞ᇊᆖҐ䗷〻ѝᖂ㓣઼࠶᷀䜘࠶⴨ሩ䟽㾱
〻ᓖⲴ䰞仈˄ণྲօ䘹ᤙᔿ 12.1 ѝ৲ᮠⲴ٬˅ǄⲴ٬ᱟሩ⇿њ䇝㓳ṧֻ⤜・䘹ᤙⲴˈᆳ
สҾањ੟ਁᔿ㿴ࡉˈ㘳㲁亶ฏ⨶䇪㜭੖㋮⺞亴⍻⢩ᇊṧֻⲴ䇝㓳٬Ǆഐ↔ሩҾ䛓Ӌ㜭⭡亶ฏ
⨶䇪↓⺞䀓䟺Ⲵ䇝㓳ṧֻˈᆖҐⲴ࠶᷀ᡀ࠶㻛ᕪॆ˗㘼ሩн㜭↓⺞䀓䟺Ⲵṧֻˈ࠶᷀ᡀ࠶㻛ᕡ
ॆǄ
EBNNⲴ䗃‫ޕ‬वᤜ˄1˅ᖒᔿѪ<xi,f(xi)>Ⲵа㓴䇝㓳ṧֻˈнवਜ਼䇝㓳ሬᮠ˗˄2˅а㓴亶
ฏ⨶䇪ˈ㊫լҾสҾ䀓䟺ⲴᆖҐ˄ㅜ 11 ㄐ˅઼KBANNѝ֯⭘Ⲵˈնᆳ㺘⽪Ѫа㓴亴‫ݸ‬䇝㓳䗷
Ⲵ⾎㓿㖁㔌ˈ㘼нᱟHornᆀਕǄEBNNⲴ䗃ࠪᱟањ㜭䙬䘁ⴞḷ࠭ᮠfⲴᯠⲴ⾎㓿㖁㔌Ǆ↔ᆖҐ
ࡠⲴ㖁㔌㜭ཏᤏਸ䇝㓳ṧֻ<xi,f(xi)>ˈԕ৺Ӿ亶ฏ⨶䇪ѝᣭਆⲴfⲴ䇝㓳ሬᮠǄሩ䇝㓳ṧֻ
<xi,f(xi)>ⲴᤏਸᶴᡀҶᆖҐⲴᖂ㓣ᡀ࠶ˈ㘼ሩ亶ฏ⨶䇪ѝᣭਆⲴ䇝㓳ሬᮠⲴᤏਸᶴᡀҶᆖҐⲴ
࠶᷀ᡀ࠶Ǆ
Ѫ䈤᰾ EBNN ѝ֯⭘Ⲵ亶ฏ⨶䇪ˈ㘳㲁മ 12-7ǄമⲴк䶒䜘࠶ᱮ⽪Ⲵᱟⴞḷ࠭ᮠ Cup Ⲵ

EBNN 亶ฏ⨶䇪ˈ⇿аᯩඇ㺘⽪亶ฏ⨶䇪ѝањ⾎㓿㖁㔌Ǆ⌘᜿൘↔ֻѝˈ㺘 12-3 Ⲵㅖਧ亶
ฏ⨶䇪ѝ⇿њ Horn ᆀਕᴹаሩᓄⲴ㖁㔌ǄֻྲˈḷѪ Graspable Ⲵ㖁㔌䗃‫ޕ‬Ѫаᇎֻ᧿䘠ˈ
䗃ࠪѪ৽᱐ሩ䊑ᱟ੖ Graspable Ⲵ٬˄EBNN ިරᛵߥл⭘ 0.8 㺘⽪ⵏભ仈ˈ⭘ 0.2 㺘⽪‫ٷ‬ભ
仈˅Ǆ↔㖁㔌㊫լҾ㺘 12-3 ѝ㔉ࠪⲴ Graspable Ⲵ Horn ᆀਕǄḀӋ㖁㔌ԕަԆ㖁㔌Ⲵ䗃ࠪ֌
Ѫ䗃‫ˈྲֻ˄ޕ‬ᴰਣ䗩ḷѪ Cup Ⲵ㖁Ⲵ䗃‫ޕ‬Ѫ StableǃLiftable ઼ OpenVessel 㖁㔌Ⲵ䗃ࠪǄ˅
ഐ↔ˈ㓴ᡀ亶ฏ⨶䇪Ⲵ䘉Ӌ㖁㔌ਟԕ䬮᧕䎧ᶕˈሩ⇿њ䗃‫ޕ‬Ṹֻ᧘⨶ࠪⴞḷ࠭ᮠˈྲ Horn ᆀ
ਕѻ䰤Ⲵ䬮᧕аṧǄа㡜ൠˈ䘉Ӌ亶ฏ⨶䇪㖁㔌ਟ⭡Ḁཆ䜘Ⓚᨀ‫׋‬㔉ᆖҐಘˈᡆ㘵ҏਟᱟ਼а
㌫㔏ԕࡽᆖҐⲴ㔃᷌ǄEBNN ֯⭘䘉Ӌ亶ฏ⨶䇪ᶕᆖҐᯠⲴⴞḷ࠭ᮠǄᆳ൘↔䗷〻ѝн᭩ਈ亶
ฏ⨶䇪Ǆ
EBNNⲴⴞⲴᱟᆖҐањ᧿䘠ⴞḷ࠭ᮠⲴᯠ⾎㓿㖁㔌ǄᡁԜሶ↔ᯠ㖁㔌〠Ѫⴞḷ㖁㔌
˄target network˅Ǆ൘മ 12-7 Ⲵֻᆀѝˈⴞḷ㖁㔌Cuptargetᱮ⽪൘മⲴᓅ䜘ˈᆳⲴ䗃‫ޕ‬Ѫԫ᜿
Ⲵᇎֻ᧿䘠ˈ䗃ࠪѪ㺘⽪↔ሩ䊑ᱟ੖ѪCupⲴ٬Ǆ
EBNN䙊䗷ᢗ㹼ࡽа㢲᧿䘠ⲴTangentProp㇇⌅ᶕᆖҐⴞḷ㖁㔌ˈഎᗶалˈTangentProp䇝
㓳㖁㔌 ԕ ᤏ ਸ 䇝㓳 ٬ ઼ 䇝㓳 ሬ ᮠ Ǆ EBNN ᢺ ᆳ ᧕ ᭦ ࡠ Ⲵ 䗃 ‫ ޕ‬䇝㓳 ٬ <xi,f(xi)> Ր 䙂㔉
TangentPropǄ↔ཆˈEBNN䘈ᢺᆳӾ亶ฏ⨶䇪ѝ䇑㇇ࠪⲴሬᮠᨀ‫׋‬㔉TangentPropǄѪ⨶䀓
EBNNᱟྲօ䇑㇇䘉Ӌ䇝㓳ሬᮠⲴˈ޽⅑㘳㲁മ 12-7Ǆമкᯩᱮ⽪Ҷሩа⢩ᇊ䇝㓳ᇎֻxiˈ亶
ฏ⨶䇪֌ࠪⲴⴞḷ࠭ᮠ٬亴⍻ǄEBNNሩᓄҾ䗃‫ޕ‬ᇎֻⲴ⇿ањ⢩ᖱ䇑㇇↔亴⍻ⲴሬᮠǄֻ
ྲˈ൘മѝˈᇎֻxi ᧿䘠Ѫࠐњ⢩ᖱྲMadeOfStyrofoam=0.2˄ণѪ‫ˈ˅ٷ‬㘼亶ฏ⨶䇪亴⍻Ѫ
Cup˙0.8˄ণⵏ˅ǄEBNNሩᓄҾ⇿њᇎֻ⢩ᖱ䇑㇇↔亴⍻Ⲵ‫ٿ‬ሬˈᗇࡠл䶒Ⲵ‫ٿ‬ሬ䳶ਸ˖
ª wCup wCup wCup º

« wBottomIsFlat , wConcavityPointsUp ,..., wMadeOfStyrofoam »
¬ ¼x xi
Explanation of training example in terms of domain theory: ṩᦞ亶ฏ⨶䇪ᗇࡠⲴ䇝㓳ṧֻ
Ⲵ䀓䟺
Target network: ⴞḷ㖁㔌
മ 12-7 ൘ EBNN ѝа䇝㓳ṧֻⲴ䀓䟺
↔䀓䟺⭡亶ฏ⨶䇪㖁㔌˄к䜘˅֌ࠪⲴⴞḷ࠭ᮠ٬亴⍻ᶴᡀǄ䇝㓳ሬᮠӾ↔䀓䟺ѝᣭਆࠪᶕˈԕ
䇝㓳࠶⿫Ⲵⴞḷ㖁㔌Ǆ⇿њ⸙ᖒඇ㺘⽪ањঅ⤜Ⲵཊቲ⾎㓿㖁㔌Ǆ
䘉㓴ሬᮠᱟ亶ฏ⨶䇪亴⍻࠭ᮠሩ䗃‫ޕ‬ᇎֻⲴởᓖǄлḷ㺘⽪䘉Ӌሬᮠ൘x=xiк䇑㇇Ǆ൘ᴤ
а㡜Ⲵᛵߥлˈⴞḷ࠭ᮠᴹཊњ䗃ࠪঅ‫ˈݳ‬ởᓖሩ⇿њ䗃ࠪ䘋㹼䇑㇇Ǆ䘉њởᓖ⸙䱥㻛〠Ѫⴞ
ḷ࠭ᮠⲴ䳵ਟ∄㹼ࡇᔿ˄Jacobian˅Ǆ
wCup
Ѫ⨶䀓䘉Ӌ䇝㓳ሬᮠ൘ᑞࣙᆖҐⴞḷ㖁㔌Ⲵ䟽㾱ᙗˈ㘳㲁ሬᮠ Ǆྲ᷌亶ฏ⨶
wExpensive
wCup
䇪㕆⸱Ⲵ⸕䇶ѝExpensive⢩ᖱоⴞḷ࠭ᮠCupᰐ‫ˈޣ‬䛓ѸӾ↔䀓䟺ѝᣭਆⲴሬᮠ
wExpensive
Ⲵ٬Ѫ 0ǄѪ 0 Ⲵሬᮠ㺘⽪䘉ṧⲴᯝ䀰ˈ⢩ᖱExpensiveкⲴ᭩ਈሩCup٬Ⲵ亴⍻⋑ᴹᖡ૽Ǆਖ
аᯩ䶒ˈᖸབྷⲴ↓ሬᮠᡆ䍏ሬᮠ㺘⽪ᯝ䀰˖↔⢩ᖱоⴞḷ٬䶎ᑨ⴨‫ޣ‬Ǆഐ↔ˈӾ亶ฏ⨶䇪䀓䟺
ѝᣭਆⲴሬᮠᨀ‫׋‬Ҷ४࠶⴨‫઼ޣ‬н⴨‫ޣ‬⢩ᖱⲴ䟽㾱ؑ᚟Ǆᖃ䘉ӋᣭਆࠪⲴሬᮠ㻛ᨀ‫׋‬Ѫ
TangentPropⲴ䇝㓳ሬᮠԕᆖҐⴞḷ㖁㔌Cuptarget ˈᆳԜᨀ‫׋‬Ҷᤷሬ⌋ॆ䗷〻Ⲵᴹ⭘Ⲵ‫ٿ‬㖞
˄bias˅Ǆ䙊ᑨ⾎㓿㖁㔌ѝⲴ䈝⌅ᖂ㓣‫ٿ‬㖞൘䘉䟼㻛ᴯᦒѪӾ亶ฏ⨶䇪ѝᗇࡠⲴሬᮠᡰӗ⭏Ⲵ
‫ٿ‬㖞Ǆ
к䶒ᡁԜ᧿䘠Ҷ亶ฏ⨶䇪亴⍻ྲօ㻛⭘Ҿ⭏ᡀа㓴䇝㓳ሬᮠǄ㋮⺞ൠ䇢ˈᆼᮤⲴEBNN㇇
⌅ྲл˖㔉ᇊ䇝㓳ṧ઼ֻ亶ฏ⨶䇪ˈEBNN俆‫ࡋݸ‬ᔪањᯠⲴ‫ޘ‬䘎᧕ࡽ侸㖁㔌ԕ㺘⽪↔ⴞḷ࠭
ᮠǄ䈕ⴞḷ㖁㔌㻛ࡍ࿻ॆѪ䲿ᵪሿᵳ٬ˈྲ൘৽ੁՐ᫝ѝ䛓ṧǄ❦ਾˈEBNNሩ⇿њ䇝㓳ṧֻ
<xi,f(xi)>⺞ᇊ⴨ᓄⲴ䇝㓳ሬᮠˈԕє↕僔ᇎ⧠Ǆㅜа↕⭘亶ฏ⨶䇪ᶕ亴⍻ᇎֻxi Ⲵⴞḷ࠭ᮠ
٬ǄԔA(xi)ԓ㺘↔亶ฏ⨶䇪ሩᇎֻxi亴⍻Ǆᦒ䀰ѻˈA(xi)ѪᖒᡀxiⲴ䀓䟺Ⲵ亶ฏ⨶䇪㓴ਸ㖁㔌ᇊ
ѹⲴ࠭ᮠǄㅜҼ↕ˈ࠶᷀亶ฏ⨶䇪Ⲵᵳ٬઼◰⍫⣦ᘱԕᣭਆࠪሩᓄxi⇿њ࠶䟿ⲴA(xi)ⲴሬᮠǄ
˄ণ൘x=xi 䇑㇇ⲴA(x)Ⲵ䳵ਟ∄㹼ࡇᔿ˅ǄᣭਆሬᮠⲴ䗷〻㊫լҾ৽ੁՐ᫝㇇⌅ѝ䇑㇇ ¥ 亩
˄㿱Ґ仈 12.5˅ǄᴰਾˈEBNN֯⭘ҶTangentPropⲴᗞሿਈරᶕ䇝㓳ⴞḷ㖁㔌ԕᤏਸл䶒Ⲵ䈟
ᐞ࠭ᮠǄ
ª § wA( x ) wfˆ ( x) ·
2
º
E ¦i ««( f ( xi ) fˆ ( xi )) P i ¦j ¨¨ wx j wx j ¸¸
2
»
»
˄12.2˅
¬ © ¹ (x xi ) ¼
ަѝ
A( xi ) f ( xi )
Pi { 1 ˄12ˊ3˅
c
䘉䟼xiԓ㺘ㅜiњ䇝㓳ᇎֻˈA(x)ԓ㺘䗃‫ޕ‬xⲴ亶ฏ⨶䇪亴⍻Ǆкḷㅖਧxjԓ㺘ੁ䟿xⲴㅜjњ
࠶䟿˄ণ⾎㓿㖁㔌Ⲵㅜjњ䗃‫ޕ‬㔃⛩˅Ǆ㌫ᮠcѪањᖂаॆᑨ䟿ˈᆳⲴ٬ᱟѪҶ‫؍‬䇱ሩᡰᴹiˈ
0 d Pi d 1Ǆ
㲭❦‫ޜ‬ᔿⴻ䎧ᶕᖸ༽ᵲˈަѝⲴᙍᜣᖸㆰঅǄ⭡ᔿ 12.2 㔉ࠪⲴ䈟ᐞ࠭ᮠоᔿ 12.1 ѝ⭡

TangentPropᴰሿॆⲴ䈟ᐞ࠭ᮠᴹ⴨਼Ⲵа㡜ᖒᔿǄᴰᐖ䗩Ⲵ亩ྲ䙊ᑨ䛓ṧˈᱟ䇝㓳٬f(xi)઼ⴞ
wA( x)
ḷ㖁㔌亴⍻٬ fˆ ( xi ) ѻ䰤Ⲵ䈟ᐞᒣᯩǄᴰਣ䗩Ⲵ亩㺑䟿ҶӾ亶ฏ⨶䇪ѝᣭਆⲴ䇝㓳ሬᮠ j
wx
wfˆ ( x )
઼ⴞḷ㖁㔌Ⲵᇎ䱵ሬᮠ ѻ䰤Ⲵ䈟ᐞᒣᯩǄഐ↔ˈᴰᐖ䗩Ⲵ亩ᨀ‫׋‬ᱟᖂ㓣㓖ᶏˈ‫ٷ‬䇮ᗵ享
wx j
ᤏਸ䇝㓳ᮠᦞ˗㘼ᴰਣ䗩Ⲵ亩ᨀ‫Ⲵ׋‬ᱟ࠶᷀㓖ᶏˈণ‫ٷ‬䇮ᗵ享ᤏਸӾ亶ฏ⨶䇪ѝᣭਆⲴ䇝㓳ሬ
wfˆ ( x ) wfˆ ( s j (D , xi ))
ᮠǄ⌘᜿൘ᔿ 12.2 ѝⲴሬᮠ ਚᱟᔿ 12.1 ѝ㺘䗮ᔿ Ⲵа⿽⢩↺ᖒᔿˈަѝ
wx j wD
ሶ s j (D , xi ) ѝⲴxijᴯѪҶxij+¢ˈEBNN֯⭘Ⲵ㋮⺞Ⲵᵳ٬䇝㓳⌅ࡉ⭡Thrun˄1996˅᧿䘠Ǆ
EBNNѝᖂ㓣઼࠶᷀ᆖҐ䜘࠶Ⲵ⴨ሩ䟽㾱ᙗ⭡ᑨ䟿i⺞ᇊˈᆳ⭡ᔿ 12-3 ᇊѹǄiⲴ٬ᱟ

⭡亶ฏ⨶䇪Ⲵ亴⍻A(xi) ઼䇝㓳٬f(xi)Ⲵᐞᔲ⺞ᇊⲴǄᆖҐⲴ࠶᷀ᡀ࠶ሩҾ㜭㻛亶ฏ⨶䇪↓⺞亴
⍻Ⲵ䇝㓳ṧֻަᵳ٬㻛࣐䟽ˈ㘼ሩҾн㜭↓⺞亴⍻Ⲵṧֻᵳ٬߿䖫Ǆ䘉а࣐ᵳ੟ਁᔿ㿴ࡉ‫ٷ‬ᇊ
൘䇝㓳٬㜭ཏ㻛亶ฏ⨶䇪↓⺞亴⍻ᰦˈӾ亶ฏ⨶䇪ѝᣭਆⲴ䇝㓳ሬᮠᴤᴹਟ㜭ᱟ↓⺞ⲴǄ㲭❦
ਟ㜭ᶴ䙐ࠪ↔੟ਁᔿ㿴ࡉཡ䍕Ⲵᛵߥˈ൘ᇎ䐥ѝࠐњ亶ฏѝ䜭ᐢ䇱᰾ᱟᴹ᭸Ⲵ˄ֻྲˈ㿱
Mitchell & Thrun 1993a; Thrun 1996˅Ǆ
12.4.5 䈪᱄
ᾲᤜൠ䈤ˈEBNN㇇⌅֯⭘Ⲵ亶ฏ⨶䇪㻛㺘⽪Ѫа㓴亴‫ݸ‬ᆖҐࡠⲴ⾎㓿㖁㔌ˈ❦ਾ亶ฏ⨶
䇪о䇝㓳ṧֻа䎧䇝㓳ަ䗃ࠪ‫ٷ‬䇮˄ⴞḷ㖁㔌˅Ǆሩ⇿њ䇝㓳ṧֻˈEBNN֯⭘ަ亶ฏ⨶䇪ᶕ
䀓䟺ᆳˈ❦ਾӾ↔䀓䟺ѝᣭਆ䇝㓳ሬᮠǄሩᇎֻⲴ⇿њ኎ᙗ䇑㇇ࠪањ䇝㓳ሬᮠˈԕ᧿䘠˖᤹
➗亶ฏ⨶䇪ˈⴞḷ࠭ᮠ٬ᱟᘾṧ⭡ަ኎ᙗ٬Ⲵᗞሿਈॆᖡ૽ⲴǄ䘉њ䇝㓳ሬᮠ㻛ᨀ‫׋‬㔉
TangentPropⲴањਈփˈަѝ֯ⴞḷ㖁㔌ᤏਸ䘉Ӌሬᮠ઼䇝㓳ṧֻ٬Ǆᤏਸሬᮠ䲀ࡦҶᆖҐࡠ
㖁㔌ᗵ享ᤏਸ亶ฏ⨶䇪㔉ࠪⲴ‫׍‬䎆‫ޣ‬㌫ˈ㘼ᤏਸ䇝㓳٬䲀ࡦҶ㖁㔌ᗵ享ᤏਸ㿲ሏࡠⲴᮠᦞᵜ
䓛ǄᤏਸሬᮠⲴᵳ٬i ᱟ⭡⇿њ䇝㓳ṧֻ⤜・⺞ᇊⲴˈᆳสҾ亶ฏ⨶䇪亴⍻↔ṧֻ䇝㓳٬Ⲵ
㋮⺞〻ᓖǄ
൘ ཊ њ 亶 ฏ ޵ ˈ EBNN ᐢ 㻛䇱 ᰾ ᱟ Ӿ 䘁 լ 亶 ฏ ⨶ 䇪 ѝ ᆖ Ґ Ⲵ а ⿽ ᴹ ᭸ ᯩ ⌅ Ǆ Thrum
˄1996˅᧿䘠Ҷᆳ൘к䘠䇘䇪Ⲵ Cup ᆖҐԫ࣑ⲴਈփкⲴᓄ⭘ˈᒦфᣕ੺䈤ᆳ∄ḷ߶৽ੁՐ᫝
㇇⌅⌋ॆᴤѪ㋮⺞ˈ⢩࡛ᱟ൘䇝㓳ᮠᦞ㕪ቁⲴᛵߥлǄֻྲˈ൘㓿䗷 30 њṧֻ䇝㓳ਾˈ
EBANN ൘ਖа⍻䈅ᮠᦞ䳶кᗇࡠⲴ൷ᯩṩ䈟ᐞѪ 5.5ˈ㘼৽ੁՐ᫝Ⲵ䈟ᐞѪ 12.0ǄMitchell &
Thrun˄1993a˅᧿䘠Ҷᓄ⭘ EBNN ԕᆖҐ᧗ࡦ⁑ᤏⲴ〫ࣘᵪಘӪˈަѝ亶ฏ⨶䇪⭡⾎㓿㖁㔌ᶴ
ᡀˈᆳԜ亴⍻Ҷн਼ᵪಘӪሩཆ⭼⣦ᘱⲴࣘ֌Ⲵ᭸᷌Ǆަѝ EBNN ҏ֯⭘Ҷ䘁լⲴ亴‫ݸ‬ᆖҐ
Ⲵ亶ฏ⨶䇪ˈᒦ㧧ᗇҶ∄৽ੁՐ᫝ᴤྭⲴᙗ㜭Ǆ䘉䟼৽ੁՐ᫝䴰㾱㓖 90 њ䇝㓳һԦ᡽㜭䗮ࡠ
EBNN ѝ 25 њ䇝㓳һԦਾⲴᙗ㜭ǄO'Sullivan et al.˄1997˅઼ Thrun˄1996˅᧿䘠Ҷ EBNN ᓄ
⭘ࡠަԆࠐ⿽ⵏᇎц⭼ᝏ⸕઼᧗ࡦԫ࣑ˈަѝ亶ฏ⨶䇪⭡㖁㔌㓴ᡀˈᆳ֯⭘༠㓣ǃ㿶㿹઼◰‫ݹ‬
㤳തՐᝏಘ亴⍻Ҷᇔ޵〫ࣘᵪಘӪⲴࣘ֌᭸᷌Ǆ
EBNN оަԆสҾ䀓䟺ⲴᆖҐᯩ⌅ѻ䰤ᆈ൘޵൘Ⲵ㚄㌫ˈྲㅜ 11 ㄐ᧿䘠Ⲵ Prolog-EBGǄ

എᗶ䈕ㄐѝ Prolog-EBG ҏสҾ亶ฏ⨶䇪ᶴ䙐䀓䟺˄ሩṧֻⴞḷ٬Ⲵ亴⍻˅Ǆ൘ Prolog-EBG
ѝˈ䀓䟺Ⲵᶴ䙐ᶕ㠚Ҿ⭡ Horn ᆀਕ㓴ᡀⲴ亶ฏ⨶䇪ˈ㘼ⴞḷ‫ٷ‬䇮Ⲵ㋮ॆᱟ䙊䗷䇑㇇↔䀓䟺ᡀ
・ⲴᴰᕡᶑԦǄഐ↔䀓䟺ѝⲴ⴨‫׍ޣ‬䎆ᙗ൘ᆖҐࡠⲴ Horn ᆀਕ‫ٷ‬䇮ѝ৽᱐ǄEBNN ᶴ䙐Ҷа
њ⴨լⲴ䀓䟺ˈնᆳᱟสҾ⾎㓿㖁㔌ᖒᔿⲴ亶ฏ⨶䇪ˈ㘼нᱟ Horn ᆀਕǄྲ൘ Prolog-EBG
ѝˈ⴨‫׍Ⲵޣ‬䎆ᙗᱟӾ䀓䟺ѝᣭਆⲴˈ㘼ф㻛⭘Ҿ㋮ॆⴞḷ‫ٷ‬䇮Ǆ൘ EBNN ѝˈ䘉Ӌ‫׍‬䎆ᙗ
ᖒᔿѪሬᮠˈഐѪ൘⾎㓿㖁㔌䘉ṧⲴ䘎㔝࠭ᮠѝˈሬᮠᱟ㺘⽪‫׍‬䎆ᙗⲴᖸ㠚❦Ⲵᯩ⌅Ǆ⴨৽ˈ
൘ㅖਧ䀓䟺ᡆ䙫䗁䇱᰾ѝˈ㺘⽪‫׍‬䎆ᙗⲴ㠚❦ᯩ⌅ᱟ᧿䘠↔䇱᰾ᡰᓄ⭘Ⲵṧֻ䳶Ǆ
ㅜ 11 ㄐ᧿䘠ⲴㅖਧⲴสҾ䀓䟺ᯩ⌅о EBNN ൘ަᆖҐ㜭࣋ᯩ䶒ᴹࠐњн਼Ǆѫ㾱н਼൘

Ҿ EBNN 㜭༴⨶нᆼ㖾Ⲵ亶ฏ⸕䇶ˈ㘼 Prolog-EBG н㜭Ǆ䘉ан਼ᱟ⭡Ҿ EBNN ᱟᔪ・൘ᤏ
ਸ㿲ሏ䇝㓳٬Ⲵᖂ㓣ᵪࡦѻкⲴˈ㘼ф亶ฏ⨶䇪ਚ㻛֌ѪᆖҐࡠⲴ‫ٷ‬䇮Ⲵ䱴࣐㓖ᶏǄㅜҼњ䟽
㾱н਼൘Ҿ Prolog-EBG ᆖҐࡠ䙀⑀໎䮯Ⲵ Horn ᆀਕ䳶Ǆ㘼 EBNN ᆖҐࡠപᇊབྷሿⲴ⾎㓿㖁
㔌Ǆྲㅜ 11 ㄐ䇘䇪ⲴˈᆖҐ Horn ᆀਕ䳶Ⲵањ䳮仈ᱟˈ䲿⵰ᆖҐ䗷〻Ⲵ䘋㹼઼ᯠ Horn ᆀਕ
㻛࣐‫࠶ˈޕ‬㊫ᯠᇎֻⲴᔰ䬰нᯝ໎䮯Ǆ❦㘼പᇊབྷሿⲴ⾎㓿㖁㔌ҏᴹ⴨ᓄⲴн䏣ˈᆳਟ㜭ᰐ⌅
㺘⽪䏣ཏ༽ᵲⲴ࠭ᮠˈ㘼໎䮯Ⲵ Horn ᆀਕ䳶ਟԕ㺘⽪䎺ᶕ䎺༽ᵲⲴ࠭ᮠǄMitchell & Thrun
˄1993b˅ᴤ䈖㓶ൠ䇘䇪Ҷ‫ޣ‬Ҿ EBNN ઼ㅖਧ㺘⽪ⲴสҾ䀓䟺ᆖҐᯩ⌅ѻ䰤㚄㌫Ǆ
12.5 ֵ⭞‫ݾ‬僂⸛䇼ᶛᢟኋᩒ㍘㇍ᆆ
ࡽ䶒є㢲㘳ḕҶ‫ݸ‬傼⸕䇶൘ᆖҐѝⲴє⿽н਼䀂㢢Ǆࡍ࿻ॆᆖҐಘⲴ‫ٷ‬䇮ǃ઼᭩ਈⴞⲴ࠭
ᮠԕᕅሬ‫ٷ‬䇮オ䰤кⲴᩌ㍒Ǆᵜ㢲ᡁԜ㘳㲁֯⭘‫ݸ‬傼⸕䇶ᶕ᭩ਈ‫ٷ‬䇮オ䰤ᩌ㍒Ⲵㅜй⿽ᯩ⌅˖
ণ᭩ਈᩌ㍒ѝᇊѹਸ⌅ᩌ㍒↕Ⲵ㇇ᆀ䳶ਸǄ䘉а䙄ᖴ㻛⭘Ҿ FOCL˄Pazzani et al. 1991ˈ
Pazzani & Kibler 1992˅ԕ৺ ML-SMART˄Bergadano & Giordanna 1990˅ㅹ㌫㔏Ǆ䘉䟼ᡁԜ⭘
FOCL ᶕ䈤᰾䘉а䙄ᖴǄ
12.5.1 FOCL ㇍⌋
FOCL ᱟㅜ 10 ㄐ᧿䘠Ⲵ㓟ᖂ㓣Ⲵ FOIL ㌫㔏ⲴањᢙኅǄFOIL ઼ FOCL 䜭ᆖҐа㓴а䱦

Horn ᆀਕԕ㾶ⴆ㿲ሏࡠⲴ䇝㓳ֻǄєњ㌫㔏䜭ᓄ⭘Ҷᒿࡇ㾶ⴆ㇇⌅ᶕᆖҐঅњ Horn ᆀਕˈ〫
৫䛓Ӌ㻛ᯠ Horn ᆀਕ㾶ⴆⲴ↓ֻˈ❦ਾ൘࢙։Ⲵ䇝㓳ṧֻк䟽༽䘉а䗷〻Ǆ൘єњ㌫㔏ѝˈ
⇿њᯠ Horn ᆀਕ䜭ᱟ䙊䗷а㡜ࡠ⢩↺ᩌ㍒ࡋᔪⲴˈᔰ࿻Ҿᴰа㡜Ⲵ Horn ᆀਕ˄ণнਜ਼ࡽԦⲴ
ᆀਕ˅Ǆ❦ਾ⭏ᡀᖃࡽᆀਕⲴࠐњ‫ى‬䘹⢩ॆᔿˈᒦ䘹ᤙަѝ‫ޣ‬Ҿ䇝㓳ṧֻᴹᴰབྷؑ᚟໎⳺Ⲵа
њǄ䟽༽䈕䗷〻ˈ⭏ᡀᴤཊⲴ‫ى‬䘹⢩ॆᔿᒦ䘹ᤙᴰ֣Ⲵˈⴤࡠ㧧ᗇањ┑䏣ᤷᇊᙗ㜭Ⲵ Horn
ᆀਕǄ
FOIL ઼ FOCL ѻ䰤Ⲵ४࡛൘Ҿᩌ㍒অњ Horn ᆀਕⲴа㡜ࡠ⢩↺䗷〻ѝ‫ى‬䘹‫ٷ‬䇮⭏ᡀⲴᯩ

⌅Ǆㅜ 10 ㄐ᧿䘠Ⲵ FOIL ⭏ᡀ⇿њ‫ى‬䘹⢩ॆᔿᱟ䙊䗷࣐‫ޕ‬ањᯠ᮷ᆇࡠᆀਕࡽԦѝᗇࡠⲴǄ
FOCL ֯⭘਼ṧⲴᯩ⌅ӗ⭏‫ى‬䘹⢩ॆᔿˈն䘈สҾ亶ฏ⨶䇪⭏ᡀҶ䱴࣐Ⲵ⢩ॆᔿǄമ 12-8 Ⲵᩌ
㍒ṁⲴᇎ㓯䗩ᱮ⽪Ҷ൘ FOIL ިරⲴᩌ㍒ѝ㘳㲁Ⲵа㡜ࡠ⢩↺ᩌ㍒↕Ǆമ 12-8 ᩌ㍒ṁⲴ㲊㓯䗩
㺘⽪ FOCL ѝสҾ亶ฏ⨶䇪㘳㲁Ⲵ䱴࣐‫ى‬䘹⢩ॆᔿǄ
㲭❦ FOCL ઼ FOIL 䜭ᆖҐа䱦 Horn ᆀਕˈᡁԜ䘉ᰦ╄⽪Ⲵ᫽֌䜭ਚᴹㆰঅⲴભ仈 Horn

ᆀਕ˄ᰐਈ䟿Ⲵ˅Ǆ⢩࡛ൠˈ޽⅑㘳㲁മ 12-3 ѝⲴ Cup ⴞḷᾲᘥǃ䇝㓳ṧ઼ֻ亶ฏ⨶䇪ǄѪ
᧿䘠 FOIL Ⲵ᫽֌ˈᡁԜᗵ享俆‫ݸ‬൘ࠪ⧠Ҿ亶ฏ⨶䇪઼‫ٷ‬䇮㺘⽪ѝⲴє⿽᮷ᆇѻ䰤֌а४࠶Ǆ
ᖃањ᮷ᆇਟ㻛⭘Ҿ᧿䘠ањ䗃ࠪ‫ٷ‬䇮ᰦˈᡁԜ〠ᆳᱟ᫽֌ර˄operational˅Ǆֻྲˈ൘മ 12-
3 Ⲵ Cup ֻᆀѝˈᡁԜ‫ݱ‬䇨䗃ࠪ‫ٷ‬䇮ѝਚ㜭ᕅ⭘᧿䘠䇝㓳ṧֻⲴ 12 њ኎ᙗ˄ྲ˖HasHandle,
HandleOnTop˅ǄสҾ䘉 12 њ኎ᙗⲴ᮷ᆇ㻛䇔Ѫᱟ᫽֌රⲴǄ⴨৽ˈ䛓Ӌਚࠪ⧠൘亶ฏ⨶䇪ѝ
֌Ѫѝ䰤⢩ᖱնнᱟᇎֻⲴ৏ᆀ኎ᙗⲴ᮷ᆇˈ㻛䇔Ѫᱟ䶎᫽֌රǄ൘↔ᛵߥл䶎᫽֌ර኎ᙗⲴ
ањֻᆀᱟ኎ᙗ StableǄ
Generated by the domain theory˖⭡亶ฏ⨶䇪⭏ᡀ
മ 12-8FOCL ѝⲴ‫ٷ‬䇮オ䰤ᩌ㍒
ѪᆖҐањ㿴ࡉˈFOCL Ӿа㡜Ⲵ‫ٷ‬䇮ᔰ࿻ˈᩌ㍒䙀⑀⢩↺Ⲵ‫ٷ‬䇮Ǆᴹє⿽㇇ᆀ⭘Ҿ⭏ᡀᖃࡽ‫ٷ‬
䇮Ⲵ⢩ॆᔿǄа⿽ᱟ໎࣐ањᯠⲴ᮷ᆇ˄മѝⲴᇎ㓯˅Ǆਖа⿽㇇ᆀᱟ䙊䗷໎࣐а㓴᮷ᆇ⢩ॆ↔㿴
ࡉǄ䘉㓴᮷ᆇ᤹➗亶ฏ⨶䇪ᶴᡀҶⴞḷᾲᘥⲴ䙫䗁‫࠶ݵ‬ᶑԦ˄മѝ㲊㓯˅ǄFOCL ൘ᡰᴹ䘉Ӌ‫ى‬䘹ॆ
ᔿѝสҾᆳԜ൘ᮠᦞкⲴᙗ㜭䘋㹼䘹ᤙǄഐ↔ˈнᆼ㖾Ⲵ亶ฏ⨶䇪ਚՊ൘ᴹ䇱ᦞ᭟ᤱ⨶䇪ᰦ᡽Պᖡ૽
‫ٷ‬䇮Ǆ䈕ֻสҾࡽ䶒 KBANN ֻᆀ⴨਼Ⲵ䇝㓳ᮠᦞ઼亶ฏ⨶䇪Ǆ
൘ަа㡜ࡠ⢩↺ᩌ㍒Ⲵ⇿а⛩ˈFOCL ֯⭘л䶒є⿽㇇ᆀᢙኅަᖃࡽ‫ٷ‬䇮 h˖
1ˊሩнᱟ h а䜘࠶Ⲵ⇿њ᫽֌ර᮷ᆇˈࡋᔪ h Ⲵањ⢩ॆᔿˈᯩ⌅ᱟ࣐‫ޕ‬᮷ᆇࡠࡽԦ

ѝǄ䘉ҏᱟ FOIL ѝ⭏ᡀ‫ى‬䘹ਾ㔗Ⲵᯩ⌅Ǆമ 12-8 ᇎ㓯㇝ཤ㺘⽪Ҷ↔⿽㊫රⲴ⢩ॆǄ
2ˊ᤹➗亶ฏ⨶䇪ˈࡋᔪањ᫽֌රⲴˈᒦфᱟⴞḷᾲᘥⲴ䙫䗁‫࠶ݵ‬ᶑԦǄሶ䘉㓴᮷ᆇ࣐
‫ ࡠޕ‬h ⲴᖃࡽࡽԦѝ৫Ǆᴰਾ‫ ࢚؞‬h ⲴࡽԦˈ〫৫ሩҾ䇝㓳ᮠᦞн䴰㾱Ⲵ᮷ᆇǄമ 12-8 ѝ㲊
㇝ཤ㺘⽪Ҷ↔⿽㊫රⲴ⢩ॆǄ
к䶒ㅜ 2 ⿽㇇ᆀⲴ䈖㓶䗷〻ྲлǄFOCL 俆‫ݸ‬䘹ᤙаᶑ亶ฏ⨶䇪ᆀਕˈᆳⲴཤ䜘˄ࡽԦ˅

३䝽ⴞḷᾲᘥǄྲ᷌ᴹཊњ䘉ṧⲴᆀਕˈ䘹ᤙަѝփ䜘˄ਾԦ˅‫ޣ‬Ҿ䇝㓳ṧֻᴹᴰ儈ؑ᚟໎⳺
ⲴǄֻྲˈ൘亶ฏ⨶䇪Ⲵ䇝㓳ᮠᦞѝ˄മ 12-3˅ˈਚᴹањṧⲴᆀਕ˖
CupĕStable, Liftable, OpenVessel
ᡰ䘹ᆀਕⲴࡽԦᖒᡀҶⴞḷᾲᘥⲴањ䙫䗁‫࠶ݵ‬ᶑԦǄ൘䘉Ӌ‫࠶ݵ‬ᶑԦѝˈ޽⅑֯⭘亶ฏ
⨶䇪ˈ⇿њ䶎᫽֌ර᮷ᆇ㻛ᴯᦒᦹˈᒦфሶᆀਕࡽԦԓ‫ࡠޕ‬ᆀਕਾԦѝǄֻྲˈ亶ฏ⨶䇪ᆀਕ
StableĕBottomIsFlat 㻛⭘Ҿሶ᫽֌රⲴ BottomIsFlat ԓᦒ䶎᫽֌රⲴ StableǄ䘉њĀኅᔰā
˄ unfolding˅亶ฏ⨶䇪Ⲵ䗷〻ᤱ㔝ⴤࡠ‫࠶ݵ‬ᶑԦᐢ㻛㺘䘠Ѫ᫽֌ර᮷ᆇǄྲ᷌ᴹཊњਟ䘹Ⲵ亶
ฏ⨶䇪ӗ⭏н਼Ⲵ㔃᷌ˈ䛓Ѹ൘↔ኅᔰ䗷〻Ⲵ⇿а↕⭘䍚ႚⲴᯩ⌅䘹ᤙᴹᴰབྷؑ᚟໎⳺ањǄ
䈫㘵ਟԕ傼䇱൘䘉њֻᆀѝˈ㔉ᇊᮠᦞ઼亶ฏ⨶䇪ˈᴰ㓸Ⲵ᫽֌ර‫࠶ݵ‬ᶑԦѪ˖
BottomIsFlat, HasHandle, Light, HasConcavity, ConcavityPointsUp
֌Ѫ⭏ᡀ‫ى‬䘹⢩ॆᔿⲴᴰਾа↕ˈ↔‫࠶ݵ‬ᶑԦ㻛‫࢚؞‬Ǆሩ㺘䗮ᔿѝⲴ⇿њ᮷ᆇˈ䲔䶎᮷ᆇ
Ⲵ〫䲔Պ䱽վ䇝㓳ֻкⲴ࠶㊫㋮ᓖˈ੖ࡉᆳ㻛〫৫Ǆवਜ਼䘉а↕僔ᱟѪҶӾ䗷⢩ॆ
(overspecialization)Ⲵᛵߥлᚒ༽ˈ䘉ᰦнᆼ㖾Ⲵ亶ฏ⨶䇪ѝवਜ਼н⴨‫Ⲵޣ‬᮷ᆇǄ൘ᡁԜⲴֻᆀ
ѝˈк䘠Ⲵ᮷ᆇ䳶ਸ३䝽єњ↓઼ֻєњ৽ֻǄ‫〫˄࢚؞‬৫˅᮷ᆇ HasHandle Պ֯ᙗ㜭᭩䘋Ǆ
ഐ↔ˈᴰ㓸Ⲵ‫࢚؞‬䗷Ⲵ᫽֌ර‫࠶ݵ‬ᶑԦѪ˖
BottomIsFlat, Light, HasConcavity, ConcavityPointsUp
䘉㓴᮷ᆇ⧠൘㻛࣐‫ࡠޕ‬ᖃࡽ‫ٷ‬䇮ⲴࡽԦѝǄ⌘᜿↔‫ٷ‬䇮ᱟമ 12-8 ѝ㲊㓯㇝ཤᱮ⽪Ⲵᩌ㍒

↕Ⲵ㔃᷌Ǆ
֯⭘Ҷк䶒є⿽᫽֌ਾˈаնᖃࡽ‫ٷ‬䇮Ⲵ‫ى‬䘹⢩ॆᔿᐢ㓿⭏ᡀҶˈᴹᴰབྷؑ᚟໎⳺Ⲵ‫ى‬䘹
㘵㻛䘹ᤙǄ൘മ 12-8 ѝᱮ⽪Ⲵֻᆀѝˈ൘ᩌ㍒ṁⲴㅜаቲ䘹ᤙⲴ‫ى‬䘹㘵Ѫ亶ฏ⨶䇪⭏ᡀⲴ䛓
ањǄᩌ㍒䗷〻㔗㔝㘳㲁䘉њ亶ฏ⨶䇪᧘ѮⲴࡽԦⲴᴤ䘋а↕Ⲵ⢩ॆᔿˈ䘉ṧᆖҐⲴᖂ㓣ᡀ࠶
ਟԕ㋮ॆ亶ฏ⨶䇪ѝሬࠪⲴࡽԦǄ൘↔ֻѝˈ亶ฏ⨶䇪‫ݸ‬ᖡ૽ᩌ㍒ⲴㅜаቲǄ❦㘼ˈᛵߥᒦ䶎
ᙫᱟྲ↔Ǆྲ᷌൘ㅜаቲሩަԆ‫ى‬䘹ᴹᴤᕪⲴ㓿傼ॆⲴ᭟ᤱˈ亶ฏ⨶䇪᧘ѮⲴ᮷ᆇӽਟ㜭൘ᩌ
㍒Ⲵਾ㔝↕僔ѝ㻛࣐‫ޕ‬Ǆᾲᤜൠ䈤ˈFOCL ᆖҐԕлᖒᔿⲴ Horn ᆀਕ˖
c m oi o b o f
ަѝcѪⴞḷᾲᘥˈoi Ѫࡍ࿻Ⲵ᫽֌ර᮷ᆇⲴਸਆˈᆳ⭡ㅜањ䈝⌅㇇ᆀ⇿⅑࣐‫ޕ‬ањ᮷
ᆇˈobᱟสҾ亶ฏ⨶䇪অ↕࣐‫Ⲵޕ‬᫽֌ර᮷ᆇਸਆˈ㘼ofѪㅜањ䈝⌅㇇ᆀ⇿⅑࣐‫ޕ‬ањⲴ᫽
֌ර᮷ᆇⲴਸਆǄ䘉йњ᮷ᆇ䳶ਸ䜭ਟ㜭ѪオǄ
к䘠Ⲵ䇘䇪╄⽪Ҷ֯⭘ભ仈ර亶ฏ⨶䇪൘অњ Horn ᆀਕⲴа㡜ࡠ⢩↺ᩌ㍒ѝࡋᔪ‫ٷ‬䇮Ⲵ

‫ى‬䘹⢩ॆᔿⲴ䗷〻Ǆ䈕㇇⌅ᖸᇩ᱃㻛ᢙኅࡠа䱦㺘⽪ѝ˄ণਜ਼ᴹਈ䟿Ⲵ㺘⽪˅Ǆㅜ 10 ㄐ䈖㓶
䇘䇪Ҷ FOIL ѝ⭏ᡀа䱦 Horn ᆀਕⲴ㇇⌅ˈवᤜк䘠ㅜањ㇇ᆀᢙኅࡠа䱦㺘⽪ⲴᛵߥǄѪ
ᢙኅㅜҼњ㇇ᆀԕ༴⨶а䱦亶ฏ⨶䇪ˈᗵ享൘ኅᔰ亶ฏ⨶䇪ᰦ㘳㲁ਈ䟿ԓᦒǄ䘉ਟԕ䙊䗷⎹৺
ࡠ㺘 11-3 എᖂ䗷〻Ⲵа⿽ᯩ⌅ᆼᡀǄ
12.5.2 䈪᱄
FOCL ֯⭘亶ฏ⨶䇪ᶕ໎࣐⇿↕ᩌ㍒অњ Horn ᆀਕѝ㘳㲁Ⲵ‫ى‬䘹⢩ॆᔿᮠ䟿Ǆമ 12-9 ∄

䖳Ҷ FOCL ᢗ㹼Ⲵ‫ٷ‬䇮オ䰤ᩌ㍒ԕ৺㓟ᖂ㓣Ⲵ FOIL ㇇⌅ᢗ㹼Ⲵᩌ㍒ǄFOCL ѝ亶ฏ⨶䇪᧘Ѯ
Ⲵ⢩ॆᔿሩᓄ FOIL ᩌ㍒ѝⲴањĀᆿĀ˄marcro˅↕ˈަѝཊњ᮷ᆇ൘а↕ѝ㻛࣐‫ޕ‬Ǆ↔䗷
〻ਟ㻛ⴻᡀᱟሶањਟ㜭ԕਾ㻛㘳㲁Ⲵ‫ٷ‬䇮ᨀॷѪ・ণ㻛㘳㲁Ⲵ‫ٷ‬䇮Ǆྲ᷌亶ฏ⨶䇪ᱟ↓⺞
Ⲵˈ䇝㓳ᮠᦞՊᱮ⽪ࠪ↔‫ٷ‬䇮∄ަԆ‫ٷ‬䇮ⲴՈ䎺ᙗˈഐ↔ᆳ㻛䘹ᤙǄྲ᷌亶ฏ⨶䇪н↓⺞ˈሩ
ᡰᴹ‫ى‬䘹Ⲵ㓿傼ॆ䇴ՠՊሶᩌ㍒ሬੁਖཆаᶑ䐟ᖴǄ
Hypotheses that fit training data equally well: ሩ䇝㓳ᮠᦞᴹ਼ㅹᤏਸᓖⲴ‫ٷ‬䇮
FOCL search: FOCL ᩌ㍒

FOIL search: FOIL ᩌ㍒
മ 12-9 ൘ FOCL ѝᩌ㍒Ⲵ‫ٷ‬䇮オ䰤
FOCL ᢙབྷҶ FOIL ѝ֯⭘Ⲵᩌ㍒㇇ᆀ䳶ਸǄFOIL ൘⇿↕ਚ㘳㲁࣐‫ޕ‬অњᯠ᮷ᆇˈ㘼 FOCL 䘈㘳

㲁࣐‫⭡ޕ‬亶ฏ⨶䇪ሬࠪⲴཊњ᮷ᆇǄ
ᾲᤜൠ䈤ˈFOCL ֯⭘䈝⌅⭏ᡀⲴ‫ى‬䘹⢩ॆᔿⲴ਼ᰦˈ䘈֯⭘Ҷ亶ฏ⨶䇪傡ࣘⲴ൘ᩌ㍒ѝ
⇿↕Ⲵ‫ى‬䘹⢩ॆ⭏ᡀǄ䈕㇇⌅䘹ᤙ䘉Ӌ‫ى‬䘹ѫ㾱ᱟสҾᆳԜ൘䇝㓳ᮠᦞкⲴ㓿傼ॆ᭟ᤱǄഐ
↔ˈ亶ฏ⨶䇪֯⭘ᯩᔿᱟ֯ᆖҐಘ‫ٿ‬㖞ˈն䇙ᆳสҾަ൘䇝㓳ᮠᦞкⲴᙗ㜭䘋㹼ᴰ㓸Ⲵ䘹ᤙǄ
⭡亶ฏ⨶䇪ᕅ‫Ⲵޕ‬䘉⿽‫ٿ‬㖞㺘⧠ᖒᔿѪ˖Ո‫ݸ‬䘹ᤙ䘉ṧⲴ Horn ᆀਕˈᆳᴰ⴨լҾ亶ฏ⨶䇪⏥
㮤˄entail˅Ⲵ᫽֌රⲴ䙫䗁‫࠶ݵ‬ᶑԦǄ↔‫ٿ‬㖞о㓟ᖂ㓣Ⲵ FOIL 〻ᒿⲴ‫ٿ‬㖞㔃ਸ൘а䎧Ǆਾ䶒
ањ‫ٿ‬㖞Ո‫ݸ‬䘹ᤙ⸝Ⲵ‫ٷ‬䇮Ǆ
൘䇨ཊнᆼ㖾亶ฏ⨶䇪Ⲵᓄ⭘ѝˈFOCL ᐢᱮ⽪ࠪ∄㓟ᖂ㓣Ⲵ FOIL ㇇⌅ᴹᴤ儈Ⲵ⌋ॆ㋮

ᓖǄֻྲ Pazzanzi & Kibler˄1992˅⹄ウҶᆖҐĀਸ⌅ỻⴈ⣦ᘱāᾲᘥⲴ䰞仈Ǆ㔉ᇊ 60 њ䇝
㓳ṧֻˈ30 њਸ⌅Ⲵ㓸ⴈỻⴈ⣦ᘱˈ30 њѪнਸ⌅ⲴǄFOIL ൘а⤜・⍻䈅ṧֻ䳶кᗇԕҶ
86%Ⲵ㋮ᓖǄFOCL ֯⭘⴨਼Ⲵ 60 њ䇝㓳ṧֻˈԕ৺ањ㋮ᓖѪ 76%Ⲵ䘁լ亶ฏ⨶䇪Ǆ㔃᷌ᗇ
ࡠⲴ‫ٷ‬䇮⌋ॆ㋮ᓖѪ 94ˁüü䈟ᐞ⦷∄ FOIL Ⲵаॺ䘈ሿǄ൘ަԆ亶ฏҏᗇࡠҶ㊫լⲴ㔃᷌Ǆ
ֻྲˈ㔉ᇊ 500 њ⭥䈍㖁䰞仈Ⲵ䇝㓳ṧֻˈԕ৺⭥䈍‫ޜ‬ਨ NYNEX ሩᆳԜⲴ䇺ᯝˈFOIL ㋮ᓖ
Ѫ 90%ˈ㘼 FOCL ൘㔉ᇊ⴨਼䇝㓳ᮠᦞԕ৺ 95ˁ㋮ᓖⲴ亶ฏ⨶䇪ᰦˈᴰ㓸䗮ࡠ㋮ᓖѪ 98ˁǄ
12.6 ⹊ガ⧦⣬
ᵜㄐኅ⽪Ⲵᯩ⌅ਚᱟ㔃ਸ࠶઼᷀ᖂ㓣ᆖҐⲴਟ㜭䙄ᖴѝⲴࠐњֻᆀǄަѝ⇿њᯩ⌅䜭㻛䇱
᰾൘ᡰ䘹亶ฏѝᙗ㜭䎵ࠪ㓟ᖂ㓣ⲴᆖҐᯩ⌅ˈն⋑ᴹањ൘བྷ㤳തⲴ䰞仈亶ฏѝ㻛ᖫᓅ⍻䈅ᡆ
䇱᰾Ǆ㔃ਸᖂ㓣઼࠶᷀ᆖҐⲴѫ仈ӽᱟањ䶎ᑨ⍫䏳Ⲵ⹄ウ亶ฏǄ
x 䘁լⲴ‫ݸ‬傼⸕䇶˄ᡆ亶ฏ⨶䇪˅൘䇨ཊᇎ䱵ᆖҐ䰞仈ѝᱟਟ࡙⭘ⲴǄߣㆆṁ઼⾎
㓿㖁㔌৽ੁՐ᫝䘉ṧⲴ㓟ᖂ㓣ᯩ⌅н㜭࡙⭘䘉ṧⲴ亶ฏ⨶䇪ˈഐ↔൘ᮠᦞ〰㕪ᰦ
ᙗ㜭䖳ᐞǄProlog-EBG 䘉ṧⲴ㓟࠶᷀ᆖҐᯩ⌅㜭ཏ࡙⭘䘉ṧⲴ亶ฏ⨶䇪ˈն൘㔉
ᇊнᆼ㖾‫ݸ‬傼⸕䇶ᰦՊӗ⭏н↓⺞Ⲵ‫ٷ‬䇮Ǆ㔃ਸᖂ㓣઼࠶᷀ᆖҐⲴᯩ⌅ਟԕ㧧ᗇ
є㘵ⲴՈ⛩ˈ߿ሿṧᵜ༽ᵲᓖˈᒦф੖ߣн↓⺞Ⲵ‫ݸ‬傼⸕䇶Ǆ
x ⴻᖵ㔃ਸᖂ㓣઼࠶᷀ᆖҐ㇇⌅Ⲵа⿽ᯩ⌅ᱟˈ㘳㲁亶ฏ⨶䇪ᱟྲօᖡ૽‫ٷ‬䇮オ䰤
ᩌ㍒ⲴǄᵜㄐᡁԜ㘳ḕҶࠐ⿽ᯩ⌅ˈᆳԜ֯⭘нᆼ㖾Ⲵ亶ฏ⨶䇪Ǆ˄1˅ࡋᔪᩌ㍒
ѝⲴࡍ࿻‫ٷ‬䇮ˈ˄2˅ᢙ‫ݵ‬ᖃࡽ‫ٷ‬䇮Ⲵᩌ㍒㇇ᆀ䳶ਸˈ˄3˅᭩ਈᩌ㍒ⴞⲴǄ
x ֯⭘亶ฏ⨶䇪ᶕࡍ࿻ॆ‫ٷ‬䇮Ⲵањ㌫㔏ᱟ KBANNǄ↔㇇⌅֯⭘а྇㕆⸱Ѫભ仈㿴
ࡉⲴ亶ฏ⨶䇪ᶕ࠶᷀ൠࡋᔪㅹԧҾ亶ฏ⨶䇪Ⲵ⾎㓿㖁㔌Ǆ❦ਾ↔㖁㔌㻛৽ੁՐ᫝
㇇⌅ᖂ㓣ൠ㋮ॆˈԕ᭩䘋ަ൘䇝㓳ᮠᦞкⲴᙗ㜭Ǆ㔃᷌ᱟањ㻛৏࿻亶ฏ⨶䇪‫ٿ‬
㖞Ⲵ㖁㔌ˈަᵳ٬㻛สҾ䇝㓳ᮠᦞᖂ㓣㋮ॆǄ
x TangentProp ֯⭘Ⲵ‫ݸ‬傼⸕䇶㻛㺘⽪Ѫⴞḷ࠭ᮠⲴᡰᐼᵋⲴሬᮠǄ൘ḀӋ亶ฏᰦˈ
ྲമ‫ۿ‬༴⨶ˈ䘉ṧ㺘⽪‫ݸ‬傼⸕䇶Ⲵањᖸ㠚❦Ⲵᯩ⌅ǄTangentProp 䙊䗷᭩ਈⴞⲴ
࠭ᮠ֯⭘䘉а⸕䇶ˈ↔࠭ᮠ൘ᩌ㍒‫ٷ‬䇮オ䰤Ⲵởᓖл䱽ѝ㻛ᴰሿॆǄ
x EBNN ֯⭘亶ฏ⨶䇪᭩ਈӪᐕ⾎㓿㖁㔌ᩌ㍒Ⲵ‫ٷ‬䇮オ䰤ⲴⴞḷǄᆳ֯⭘Ⲵ亶ฏ⨶䇪
⭡亴‫ݸ‬ᆖҐⲴ⾎㓿㖁㔌㓴ᡀˈަ֌⭘ᱟᇎ⧠ањ⾎㓿㖁㔌ˈԕ⁑ᤏㅖਧⲴสҾ䀓
䟺ᆖҐǄྲ൘ㅖਧⲴสҾ䀓䟺ⲴᆖҐѝаṧˈ亶ฏ⨶䇪㻛⭘Ҿ䀓䟺অњṧֻˈ㧧
ᗇн਼ṧֻ⢩ᖱⲴ⴨‫ޣ‬〻ᓖⲴؑ᚟Ǆ❦㘼൘⾎㓿㖁㔌㺘⽪ѝˈᴹ‫ޣ⴨ޣ‬ᙗⲴؑ᚟
㻛㺘⽪Ѫⴞḷ࠭ᮠሩᓄҾᇎֻ⢩ᖱⲴሬᮠǄ㖁㔌‫ٷ‬䇮Ⲵ䇝㓳֯⭘Ҷ TangentProp ㇇
⌅Ⲵањਈ⿽ˈަѝ㻛ᴰሿॆⲴ䈟ᐞнӵवਜ਼Ҷ㖁㔌䗃ࠪ٬Ⲵ䈟ᐞˈ䘈वਜ਼ҶӾ
䀓䟺ѝ㧧ᗇⲴ㖁㔌ሬᮠⲴ䈟ᐞǄ
x FOCL ֯⭘亶ฏ⨶䇪ᶕᢙኅ⇿↕ᩌ㍒ѝ㘳㲁ⲴⲴ‫ى‬䘹䳶Ǆᆳ֯⭘㺘⽪Ѫа䱦 Horn
ᆀਕⲴ䘁լ亶ฏ⨶䇪ᶕᆖҐа㓴䙬䘁ⴞḷ࠭ᮠⲴ Horn ᆀਕǄFOCL ᓄ⭘Ҷᒿࡇ㾶
ⴆ㇇⌅ˈ䙊䗷а㡜ࡠ⢩↺ᩌ㍒䗷〻ᶕᆖҐ⇿њ Horn ᆀਕǄ亶ฏ⨶䇪㻛⭘Ҿᢙབྷ↔
ᩌ㍒ѝ⇿↕㘳㲁Ⲵлањᴤ⢩↺‫ى‬䘹‫ٷ‬䇮䳶Ǆ❦ਾ‫ى‬䘹‫ٷ‬䇮สҾަ൘䇝㓳ᮠᦞк
Ⲵᙗ㜭㻛䇴ՠǄԕ䘉⿽ᯩ⌅ˈFOCL 㔃ਸҶ FOIL Ⲵ䍚ႚⲴǃа㡜ࡠ⢩↺ᩌ㍒ㆆ
⮕ˈԕ৺࠶᷀ᯩ⌅ѝⲴ㿴ࡉ䬮࠶᷀᧘⨶Ǆ
x ྲօᴰྭൠ㶽ਸ‫ݸ‬傼⸕䇶ࡠᯠ㿲ሏһ⢙ѝⲴ䰞仈ˈӽᱟᵪಘᆖҐѝѫ㾱Ⲵᖵ䀓ߣ
䰞仈ѻаǄ
䘈ᴹ䇨ཊ⿽㇇⌅䈅മ㔃ਸᖂ㓣઼࠶᷀ᆖҐǄֻྲˈㅜ 6 ㄐ䇘䇪ⲴᆖҐ䍍ਦᯟ㖞ؑ㖁Ⲵᯩ⌅
ᨀ‫׋‬Ҷਖа⿽䙄ᖴǄᵜㄐᵛቮⲴ৲㘳᮷⥞ᨀ‫׋‬Ҷ䘋а↕䰵䈫ⲴֻᆀᶕⓀǄ
Ґ仈
12.1 㘳㲁ᆖҐањⴞḷᾲᘥ GoodCreditRiskˈᆳᇊѹ൘Ḁᇎֻ᧿䘠ѻкˈᇎֻ᧿䘠वਜ਼ 4

њ኎ᙗ HasStudentLoan, HasSavingsAccount, IsStudent, OwnsCarǄሩҾл䶒Ⲵ亶ฏ⨶䇪ˈ㔉ࠪ
KBANN ࡋᔪⲴࡍ࿻㖁㔌ˈवᤜᡰᴹⲴ㖁㔌䘎᧕઼ᵳ٬Ǆ
GoodCreditRiskĕEmpolyed, LowDebt
EmployedĕIsStudent
LowDebtĕHasStudentLoan, HasSavingsAccount
12.2 KBANN ሶа㓴ભ仈 Horn ᆀਕਈᦒѪањࡍ࿻⾎㓿㖁㔌Ǆ㘳㲁а㊫ n-of-m ᆀਕˈ䘉

⿽ Horn ᆀਕࡽԦ˄‫ݸ‬㹼䇽˅ѝवਜ਼ m њ᮷ᆇˈᒦфवਜ਼а‫ޣ‬㚄Ⲵ৲ᮠ nˈnİmǄᖃ m њࡽԦ
ѝ㠣ቁ n њ┑䏣ᰦˈ↔ n-of-m Horn ᆀਕ㻛䇔Ѫ┑䏣Ǆֻྲˈᆀਕ˖
StudentĕLivesInDorm, Young, Studies; n=2
ᯝ䀰ྲ᷌ 3 њࡽԦѝ㠣ቁєњ┑䏣ᰦˈ↔ӪѪ StudentǄ
㔉ࠪо KBANN ⴨լⲴањ㇇⌅ˈᆳ᧕ਇа㓴ભ仈ර n-of-m ᆀਕᒦф㜭ᶴ䙐ࠪо亶ฏ⨶

䇪а㠤Ⲵ⾎㓿㖁㔌Ǆ
12.3 䈅ሶ KBANN ᢙኅˈԕ᧕ਇवਜ਼а䱦 Horn ᆀਕⲴ亶ฏ⨶䇪ˈ㘼нਚᱟભ仈 Horn ᆀ

ਕǄ˄ণ‫ݱ‬䇨 Horn ᆀਕवਜ਼ਈ䟿ˈྲㅜ 10 ㄐѝ䛓ṧ˅Ǆ㔉ࠪањ㇇⌅ԕᶴ䙐ㅹԧҾањ
Horn ᆀਕ䳶Ⲵ⾎㓿㖁㔌ˈྲ᷌н㜭ࡉ䇘䇪ަѝⲴഠ䳮ᡰ൘Ǆ
12.4 ↔Ґ仈㾱≲᧘ሬࠪ㊫լҾ TangentProp ֯⭘Ⲵởᓖл䱽⌅ࡉǄ㘳㲁ᇎֻオ䰤 X ⭡ᇎᮠ

ᶴᡀˈ㘼‫ٷ‬䇮オ䰤 H ⭡ x ⲴҼ⅑࠭ᮠᶴᡀǄণ⇿њ‫ٷ‬䇮ᖒᔿѪ˖
h(x)=w0+w1x+w2x2
˄a˅᧘ሬањởᓖл䱽⌅ࡉˈᆳᴰሿॆ৽ੁՐ᫝ѝ⴨਼Ⲵࡔᦞ˖ণ൘‫ٷ‬䇮઼䇝㓳ᮠᦞⴞ
ḷ٬ѻ䰤Ⲵ䈟ᐞᒣᯩ઼Ǆ
˄b˅᧘ሬањởᓖл䱽⌅ࡉˈᆳᴰሿॆ TangentProp ѝ⴨਼ⲴࡔᦞǄਚ㘳㲁ањਈᦒ

s(¢,x)=x+¢
12.5 EBNNӾ䀓䟺ѝᣭਆ䇝㓳ሬᮠⲴᯩ⌅ᱟˈ㘳㲁ᶴᡀ䀓䟺Ⲵ⾎㓿㖁㔌Ⲵᵳ٬઼◰⍫⣦
ᘱǄ㘳㲁ањㆰঅⲴֻᆀˈަѝ䀓䟺ⲴᖒᔿѪᴹnњ䗃‫Ⲵޕ‬অњsigmoidঅ‫ݳ‬Ǆ᧘ሬањ䗷〻ԕ
wfˆ ( x)
ᣭਆሬᮠ ˈަѝxiѪ䗃‫↔ࡠޕ‬অ‫Ⲵݳ‬⢩ᇊ䇝㓳ᇎֻˈ fˆ ( x ) Ѫsigmoidঅ‫ݳ‬䗃ࠪˈᒦф
wx j x xi
x ԓ㺘sigmoidঅ‫ݳ‬ㅜjњ䗃‫ޕ‬Ǆҏਟԕ֯⭘䇠ਧxijԓ㺘xiⲴㅜjњ࠶䟿Ǆᨀ⽪˖䈕ሬᮠо৽ੁՐ᫝
j
䇝㓳⌅ࡉѝⲴሬᮠ⴨լǄ
12.6 ޽⅑㘳㲁മ 12-8 ѝᱮ⽪Ⲵ FOCL Ⲵᩌ㍒↕僔Ǆ‫ྲٷ‬൘ᩌ㍒Ⲵㅜаቲ䘹ᤙⲴ‫ٷ‬䇮᭩
Ѫ˖
CupĕHasHandle
᧿䘠 FOCL ⭏ᡀⲴ֌Ѫ↔‫ٷ‬䇮ਾ㔗ⲴㅜҼቲ‫ى‬䘹‫ٷ‬䇮Ǆਚ䴰㾱वᤜ䛓Ӌ⭡ FOCL ⲴㅜҼњ

ᩌ㍒㇇ᆀ⭏ᡀⲴ‫ٷ‬䇮ˈণ֯⭘亶ฏ⨶䇪⭏ᡀⲴ‫ٷ‬䇮Ǆн㾱ᘈ䇠ሩ‫࠶ݵ‬ᶑԦ䘋㹼ਾ‫࢚؞‬Ǆ֯⭘㺘
12-3 ѝⲴ䇝㓳ᮠᦞǄ
12.7 ᵜㄐ䇘䇪Ҷ 3 ⿽䙄ᖴᶕ֯⭘‫ݸ‬傼⸕䇶ԕᖡ૽‫ٷ‬䇮オ䰤Ⲵᩌ㍒Ǆ֐䇔Ѫྲօ䳶ᡀ䘉й⿽

ᯩ⌅˛㜭੖ᨀࠪањ⢩↺㇇⌅ˈᆳ䳶ᡀҶ㠣ቁє⿽㇇⌅ԕ䪸ሩḀ⿽⢩↺Ⲵ‫ٷ‬䇮㺘⽪Ǆ൘↔䳶ᡀ
ѝᴹӰѸṧⲴՈ㕪⛩˛
12.8 ޽⅑㘳㲁 12.2.1 㢲ѝⲴ䰞仈ˈণᖃᮠᦞ઼‫ݸ‬傼⸕䇶䜭ᆈ൘ᰦˈᓄ֯⭘ӰѸṧⲴ߶ࡉ൘

‫ٷ‬䇮ѝ䘋㹼䘹ᤙǄ㔉ࠪ֐൘䘉њ䰞仈кⲴ㿱䀓Ǆ
ㅢ13ㄖ ໔ᕰᆜҖ
໎ᕪᆖҐ㾱䀓ߣⲴᱟ䘉ṧⲴ䰞仈˖ањ㜭ཏᝏ⸕⧟ຳⲴ㠚⋫ agentˈᘾṧᆖҐ䘹ᤙ㜭䗮ࡠ
ަⴞḷⲴᴰՈࣘ֌Ǆ䘉њᖸާᴹᲞ䙽ᙗⲴ䰞仈ᓄ⭘ҾᆖҐ᧗ࡦ〫ࣘᵪಘӪǃ൘ᐕলѝᆖҐ䘋㹼
ᴰՈ᫽֌ᐕᒿǃԕ৺ᆖҐỻ㊫ሩᔸㅹǄᖃ agent ൘ަ⧟ຳѝ֌ࠪ⇿њࣘ֌ᰦˈᯭᮉ㘵Պᨀ‫྆׋‬
䍿ᡆ᜙㖊ؑ᚟ˈԕ㺘⽪㔃᷌⣦ᘱⲴ↓⺞о੖Ǆֻྲˈ൘䇝㓳 agent 䘋㹼ỻ㊫ሩᔸᰦˈᯭᮉ㘵ਟ
൘⑨ᠿ㜌࡙ᰦ㔉ࠪ↓എᣕˈ㘼൘⑨ᠿཡ䍕ᰦ㔉ࠪ䍏എᣕˈަԆᰦ‫ى‬Ѫ䴦എᣕǄAgent Ⲵԫ࣑ቡ
ᱟӾ䘉њ䶎ⴤ᧕Ⲵǃᴹᔦ䘏ⲴഎᣕѝᆖҐˈԕ‫ׯ‬ਾ㔝Ⲵࣘ֌ӗ⭏ᴰབྷⲴ㍟〟എᣕǄᵜㄐ⵰䟽ӻ
㓽ањ〠Ѫ Q ᆖҐⲴ㇇⌅ˈᆳਟӾᴹᔦ䘏Ⲵഎᣕѝ㧧ਆᴰՈ᧗ࡦㆆ⮕ˈণ֯ agent ⋑ᴹᴹ‫ަޣ‬
ࣘ֌Պሩ⧟ຳӗ⭏ᘾṧⲴ᭸᷌Ⲵ‫ݸ‬傼⸕䇶Ǆ໎ᕪᆖҐоࣘᘱ㿴ࡂ˄dynamic programming˅㇇⌅
ᴹ‫ˈޣ‬ਾ㘵ᑨ㻛⭘Ҿ䀓ߣᴰՈॆ䰞仈Ǆ
13.1 ԁ㔃
㘳㲁ᔪ䙐ањਟᆖҐᵪಘӪǄ䈕ᵪಘӪ˄ᡆ agent˅ᴹаӋՐᝏಘਟԕ㿲ሏަ⧟ຳⲴ⣦ᘱ
˄state˅ᒦ㜭‫ࠪڊ‬а㓴ࣘ֌˄action˅ᐢ᭩ਈ䘉Ӌ⣦ᘱǄֻྲˈ〫ࣘᵪಘӪާᴹ䮌ཤ઼༠㓣ㅹ
Րᝏಘˈᒦਟԕ‫ࠪڊ‬Āⴤ䎠ā઼Ā䖜ᕟāㅹࣘ֌ǄᆖҐⲴԫ࣑ᱟ㧧ᗇањ᧗ࡦㆆ⮕
˄policy˅ˈԕ䘹ᤙ㜭䗮ࡠⴞⲴⲴ㹼ѪǄֻྲˈ↔ᵪಘӪⲴԫ࣑ᱟ൘ަ⭥⊐⭥䟿䖜վᰦ᢮ࡠ‫ݵ‬
⭥ಘ䘋㹼‫⭥ݵ‬Ǆ
ᵜㄐ‫ޣ‬ᗳⲴቡᱟ˖䘉ṧⲴ agent ᘾṧ൘ަ⧟ຳѝ‫ڊ‬ᇎ傼ᒦᡀ࣏ൠᆖҐࡠ᧗ࡦㆆ⮕Ǆ䘉䟼‫ٷ‬

ᇊ agent Ⲵⴞḷਟ㻛ᇊѹѪањഎᣕ˄reward˅࠭ᮠˈᆳሩ agent Ӿн਼Ⲵ⣦ᘱѝ䘹ਆн਼Ⲵࣘ
֌䍻Ҹањᮠᆇ٬ˈণ・ণ␵㇇˄immediate payoff˅Ǆֻྲ˖ራ᢮⭥⊐‫⭥ݵ‬ಘⲴⴞḷਟ⭘䘉ṧ
Ⲵഎᣕ࠭ᮠᤷᇊ˖ሩ䛓Ӌ㜭ཏ䘎᧕ࡠ‫⭥ݵ‬ಘⲴ⣦ᘱ-ࣘ֌䖜ᦒ䍻Ҹ↓എᣕ˄ྲ+100˅ˈሩަԆ
Ⲵ⣦ᘱࣘ֌䖜ᦒ䍻Ҹ 0 എᣕǄ䘉њഎᣕ࠭ᮠਟ޵፼൘ᵪಘӪѝ˗ᡆ㘵ਚᴹањཆ䜘ᯭᮉ㘵⸕
䚃ˈ⭡ᆳሩᵪಘӪⲴ⇿њࣘ֌㔉ࠪഎᣕ٬ǄᵪಘӪⲴԫ࣑ᱟᢗ㹼а㌫ࡇࣘ֌ˈ㿲ሏަਾ᷌ˈ޽
ᆖҐ᧗ࡦㆆ⮕ǄᡁԜᐼᵋⲴ᧗ࡦㆆ⮕ᱟ㜭ཏӾԫօࡍ࿻⣦ᘱ䘹ᤙᚠᖃⲴࣘ֌ˈ֯ agent 䲿ᰦ䰤
Ⲵ㍟〟Ⲵഎᣕ䗮ࡠᴰབྷǄ䘉њᵪಘӪᆖҐ䰞仈Ⲵа㡜Ṷᷦ൘മ 13-1 ѝᾲ㾱ࡇࠪǄ
Agent: Agent
State: ⣦ᘱ
Reward: എᣕ
Action: ࣘ֌
Environment: ⧟ຳ
Goal:Learn to choose actions that maximize˖ ⴞḷ˖ᆖҐ䘹ᤙࣘ֌֯лᔿᴰབྷॆ
where: ަѝ
മ 13-1 ањо⧟ຳӔӂⲴ agent
↔agent⭏ᆈⲴ⧟ຳ㻛᧿䘠ѪḀਟ㜭Ⲵ⣦ᘱ䳶ਸSǄᆳਟᢗ㹼ԫ᜿Ⲵਟ㜭ࣘ֌䳶ਸAǄ⇿⅑൘Ḁ⣦
ᘱstлᢗ㹼аࣘ֌atˈ↔agentՊ᭦ࡠањᇎ٬എᣕrtˈᆳ㺘⽪↔⣦ᘱ-ࣘ֌䖜ᦒⲴ・ণ٬Ǆྲ↔ӗ⭏Ҷа
㌫ࡇⲴ⣦ᘱsiˈࣘ֌ai઼・ণഎᣕriⲴ䳶ਸˈྲമᡰ⽪ǄAgentⲴԫ࣑ᱟᆖҐањ᧗ࡦㆆ⮕±:SėAˈᆳ
֯䘉ӋഎᣕⲴ઼Ⲵᵏᵋ٬ᴰབྷॆˈަѝਾ䶒Ⲵ≷ᣕ٬䲿⵰ԆԜⲴᔦ䘏ᤷᮠ߿ሿǄ
മ 13-1 ѝਟ␵ᾊൠⴻࡠˈᆖҐ᧗ࡦㆆ⮕ԕ֯㍟〟എᣕᴰབྷॆ䘉њ䰞仈䶎ᑨᲞ䙽ˈᆳ㾶ⴆ
ҶᵪಘӪᆖҐԫ࣑ԕཆⲴ䇨ཊ䰞仈Ǆа㡜ൠˈ↔䰞仈ᱟањ䙊䗷ᆖҐᶕ᧗ࡦᒿࡇ䗷〻Ⲵ䰞仈Ǆ
ֻྲ⭏ӗՈॆ䰞仈ˈަѝ㾱䘹ᤙа㌫ࡇⲴ⭏ӗࣘ֌ˈ㘼֯⭏ӗࠪⲴ䍗⢙߿৫ަᡀᵜ䗮ࡠᴰབྷ
ॆǄ޽ྲаӋᒿࡇ䈳ᓖ䰞仈ˈ‫ۿ‬൘ањབྷ෾ᐲѝ䘹ᤙࠪ』䖖䘀䖭҈ᇒˈަѝഎᣕ࠭ᮠѪ҈ᇒㅹ
ᖵⲴᰦ䰤઼ࠪ』䖖䱏Ⲵᮤփ⋩㙇Ǆа㡜ᶕ䈤ˈᡁԜᝏ‫ޤ‬䏓Ⲵ䰞仈㊫රᱟ˖ањ agent 䴰㾱䙊䗷
ᆖҐ઼䘹ᤙࣘ֌ᶕ᭩ਈ⧟ຳ⣦ᘱˈ㘼ަѝ֯⭘Ҷањ㍟〟എᣕ࠭ᮠᶕᇊѹԫ᜿ࣘ֌ᒿࡇⲴ䍘
䟿Ǆ൘↔㊫䰞仈ѝǄᡁԜ㘳㲁ࠐ⿽⢩↺ⲴṶᷦ˖वᤜࣘ֌ᱟ੖ާᴹ⺞ᇊᙗⲴ䗃ࠪ˗agent ᱟ੖
ᴹަࣘ֌ሩ⧟ຳⲴ᭸᷌Ⲵ‫ݸ‬傼⸕䇶Ǆ
൘ᵜҖࡽ䶒ˈᡁԜᐢ㓿᧕䀖ࡠҶ䙊䗷ᆖҐᶕ᧗ࡦᒿࡇ䗷〻Ⲵ䰞仈Ǆ൘ㅜ 11.4 㢲ѝǄᡁԜ

䇘䇪Ҷ⭘สҾ䀓䟺Ⲵᯩ⌅ᆖҐ㿴ࡉˈԕ᧗ࡦ䰞仈䀓ߣѝⲴᩌ㍒Ǆ൘ަѝ agent ⲴⴞⲴᱟ൘ᩌ㍒
ަⴞḷ⣦ᘱᰦⲴ⇿а↕Ӿਟ䘹ࣘ֌ѝ‫ࠪڊ‬ᢹᤙǄᵜㄐ䇘䇪Ⲵᢰᵟн਼Ҿ 11.4 㢲ˈഐѪ䘉䟼㘳
㲁Ⲵ䰞仈ѝ㹼Ѫਟ㜭ᴹ䶎⺞ᇊᙗⲴ䗃ࠪˈ㘼фᆖҐಘ㕪ቁ᧿䘠ަ㹼Ѫ䗃ࠪⲴ亶ฏ⨶䇪Ǆ൘ㅜ 1
ㄐˈᡁԜ䇘䇪Ҷ൘㾯⌻ৼ䱶ỻሩᔸѝⲴᆖҐ䰞仈Ǆަѝᾲ䘠ⲴᆖҐᯩ⌅䶎ᑨ㊫լҾᵜㄐⲴᆖҐ
ᯩ ⌅ Ǆ ᇎ 䱵 к ᵜ ㄐ Ⲵ ໎ ᕪ ᆖ Ґ ㇇ ⌅ Ⲵ а њ ᴰ ᡀ ࣏ Ⲵ ᓄ ⭘ ቡ ᱟ ㊫ լ Ⲵ ঊ ᔸ 䰞仈 Ǆ Tesauro
˄1995˅᧿䘠Ⲵ TD-Gammon 〻ᒿˈᆳ֯⭘໎ᕪᆖҐᡀѪҶц⭼㓗Ⲵ㾯⌻ৼ䱶ỻ䘹᡻Ǆ䘉њ〻
ᒿ㓿䗷Ҷ 150 зњ㠚⭏ᡀⲴሩᔸ䇝㓳ਾˈᐢ䘁լ䗮ࡠҶӪ㊫ᴰྭ䘹᡻Ⲵ≤ᒣˈᒦф൘ഭ䱵㾯⌻
ৼ䱶ỻ㚄䎋ѝо亦ቆỻ᡻ሩᔸਆᗇҶ㢟ྭⲴᡀ㔙Ǆ
ᆖҐ᧗ࡦㆆ⮕ԕ䘹ᤙࣘ֌Ⲵ䰞仈൘Ḁ⿽〻ᓖк㊫լҾަԆㄐ䇘䇪䗷Ⲵ࠭ᮠ䙬䘁䰞仈Ǆ䘉䟼
ᖵᆖҐⲴⴞḷ࠭ᮠѪ᧗ࡦㆆ⮕±:SėAǄᆳ൘㔉ᇊᖃࡽ⣦ᘱ S 䳶ਸѝⲴ s ᰦˈӾ䳶ਸ A ѝ䗃ࠪ
ањਸ䘲Ⲵࣘ֌ aǄ❦㘼ˈ໎ᕪᆖҐ䰞仈оަԆⲴ࠭ᮠ䙬䘁䰞仈ᴹࠐњ䟽㾱н਼˖
x ᔦ䘏എᣕ˄delayed reward˅Ǆ Agent Ⲵԫ࣑ᱟᆖҐањⴞḷ࠭ᮠ±Ǆᆳᢺᖃࡽ⣦
ᘱ s ᱐ሴࡠᴰՈࣘ֌ a=±(s)Ǆ൘ࡽ䶒ㄐ㢲ѝˈᡁԜᙫᱟ‫ٷ‬ᇊ൘ᆖҐ±䘉ṧⲴⴞḷ
࠭ᮠᰦˈ⇿њ䇝㓳ṧֻᱟᒿ‫Ⲵڦ‬ᖒᔿ<s, ±(s)>Ǆ❦㘼൘໎ᕪᆖҐѝˈ䇝㓳ؑ᚟н
㜭ԕ䘉⿽ᖒᔿᗇࡠǄ⴨৽ˈᯭᮉ㘵ਚ൘ agent ᢗ㹼ަᒿࡇࣘ֌ᰦᨀ‫׋‬ањᒿࡇ・ণ
എᣕ٬ˈഐ↔ agent 䶒Ѥањᰦ䰤ؑ⭘࠶䝽˄temporal credit assignment ˅Ⲵ䰞仈˖
⺞ᇊᴰ㓸എᣕⲴ⭏ᡀᓄᖂ࣏Ҿަᒿࡇѝଚањࣘ֌Ǆ
x ᧒㍒˄exploration ˅Ǆ൘໎ᕪᆖҐѝˈagent 䙊䗷ަ䘹ᤙⲴࣘ֌ᒿࡇᖡ૽䇝㓳ṧֻ
Ⲵ࠶ᐳǄ䘉ӗ⭏Ҷањ䰞仈˖ଚ⿽ᇎ傼ㆆ⮕ਟӗ⭏ᴰᴹ᭸ⲴᆖҐǄᆖҐಘ䶒ѤⲴ
ᱟањᣈѝⲴ䰞仈˖ᱟ䘹ᤙ᧒㍒ᵚ⸕Ⲵ⣦ᘱ઼ࣘ֌˄ԕ᭦䳶ᯠؑ᚟˅ˈ䘈ᱟ䘹ᤙ
ᆳᐢ㓿ᆖҐ䗷ǃՊӗ⭏儈എᣕⲴ⣦ᘱ઼ࣘ֌˄ԕ֯㍟〟എᣕᴰབྷॆ˅Ǆ
x 䜘࠶ਟ㿲ሏ⣦ᘱ˄partially observable states˅Ǆ㲭❦ѪҶᯩ‫ׯ‬䎧㿱ˈਟԕ‫ٷ‬ᇊ agent
Րᝏಘ൘⇿а↕ਟᝏ⸕ࡠ⧟ຳⲴ‫ޘ‬䜘⣦ᘱˈն൘ᇎ䱵ⲴᛵߥлՐᝏಘਚ㜭ᨀ‫׋‬䜘
࠶ؑ᚟Ǆֻྲ˖ᑖᴹࡽੁ䮌ཤⲴᵪಘӪн㜭ⴻࡠᆳਾ䶒ⲴᛵߥǄ൘↔ᛵߥлਟ㜭
䴰㾱㔃ਸ㘳㲁ަԕࡽⲴ㿲ሏԕ৺ᖃࡽⲴՐᝏಘᮠᦞԕ䘹ᤙࣘ֌ˈ㘼ᴰ֣Ⲵㆆ⮕ᴹ
ਟ㜭ᱟ䘹ᤙ⢩ᇊⲴࣘ֌ԕ᭩䘋⧟ຳਟ㿲ሏᙗǄ
x 䮯ᵏᆖҐ˄life-long learning˅Ǆн䊑࠶⿫Ⲵ࠭ᮠ䙬䘁ԫ࣑ˈᵪಘӪᆖҐ䰞仈㓿ᑨ㾱
≲↔ᵪಘӪ൘⴨਼Ⲵ⧟ຳл֯⭘⴨਼ⲴՐᝏಘᆖҐཊњ⴨‫ޣ‬ԫ࣑Ǆᘾṧ൘ゴሿⲴ
䎠ᓺѝ㹼䎠ˈԕ৺ᘾṧӾ◰‫ݹ‬ᢃঠᵪѝਆᗇᢃঠ㓨ㅹǄ䘉֯ᗇᴹਟ㜭֯⭘‫ࡽݸ‬㧧
ᗇⲴ㓿傼ᡆ⸕䇶൘ᆖҐᯠԫ࣑ᰦ߿ሿṧᵜ༽ᵲᓖǄ
13.2 ᆜҖԱࣗ
൘ᵜ㢲ѝˈᡁԜᢺᆖҐᒿࡇ᧗ࡦㆆ⮕Ⲵ䰞仈ᴤ㋮⺞ൠᖒᔿॆǄᴹ䇨ཊ⿽ᯩ⌅ਟԕ‫ࡠڊ‬Ǆֻ
ྲ˖ਟ‫ٷ‬ᇊ agent Ⲵ㹼Ѫᱟ⺞ᇊᙗᡆ䶎⺞ᇊᙗⲴ˗‫ٷ‬ᇊ agent ਟԕ亴⍻⇿ањ㹼Ѫᡰӗ⭏Ⲵ⣦
ᘱˈᡆн㜭亴⍻˗‫ٷ‬ᇊ agent ᱟ⭡ཆ䜘уᇦ䙊䗷⽪ֻᴰՈࣘ֌ᒿࡇᶕ䇝㓳ˈᡆᗵ享䙊䗷ᢗ㹼㠚
ᐡ䘹ᤙⲴࣘ֌ᶕ䇝㓳Ǆ䘉䟼ᡁԜสҾ傜ቄਟཛߣㆆ䗷〻ᇊѹ䈕䰞仈Ⲵа㡜ᖒᔿǄ䘉⿽䰞仈ᖒᔿ
䚥ᗚമ 13-1 ⽪ֻⲴ䰞仈Ǆ
൘傜ቄਟཛߣㆆ䗷〻˄Markov decision processˈMDP˅ѝˈagent ਟᝏ⸕ࡠަ⧟ຳⲴн਼

⣦ᘱ䳶ਸSˈᒦфᴹᆳਟᢗ㹼Ⲵࣘ֌䳶ਸAǄ൘⇿њ⿫ᮓᰦ䰤↕tˈagent ᝏ⸕ࡠᖃࡽ⣦ᘱst ˈ䘹
ᤙᖃࡽࣘ֌at ᒦᢗ㹼ᆳǄ⧟ຳ૽ᓄ↔agentˈ㔉ࠪഎᣕ rt=r(st, at)ˈᒦӗ⭏ањਾ㔗⣦ᘱSt+1=¥
(st, at)Ǆ䘉䟼࠭ᮠ¥઼rᱟ⧟ຳⲴа䜘࠶ˈagent нᗵ⸕䚃Ǆ൘MDPѝˈ࠭ᮠ¥(st, at)઼r(st, at)ਚ
‫׍‬䎆Ҿᖃࡽ⣦ᘱ઼ࣘ֌ˈ㘼н‫׍‬䎆ҾԕࡽⲴ⣦ᘱ઼ࣘ֌ǄᵜㄐѝᡁԜਚ㘳㲁S઼AѪᴹ䲀Ⲵᛵ
ᖒǄа㡜ᶕ䈤ˈ¥઼rਟѪ䶎⺞ᇊᙗ࠭ᮠˈնᡁԜ俆‫ݸ‬Ӿ⺞ᇊᙗⲴᛵᖒᔰ࿻Ǆ
AgentⲴԫ࣑ᱟᆖҐањㆆ⮕±:SėAˈԕสҾᖃࡽ㿲ሏࡠⲴ⣦ᘱst䘹ᤙлⲴа↕ࣘ֌at˗ণ
±(st)=atǄྲօ㋮⺞ᤷᇊ↔agent㾱ᆖҐⲴㆆ⮕±઒˛ањ᰾ᱮⲴᯩ⌅ᱟ㾱≲↔ㆆ⮕ሩᵪಘӪӗ
⭏ᴰབྷⲴ〟㍟എᣕǄѪ㋮⺞ൠ㺘䘠䘉њ㾱≲ˈᡁԜᇊѹ˖䙊䗷䚥ᗚањԫ᜿ㆆ⮕±Ӿԫ᜿ࡍ࿻
±
⣦ᘱst㧧ᗇⲴ㍟〟٬V (st)Ѫ˖
V S ( st ) { rt Jrt 1 J 2 rt 2 ...
f
{ ¦ J i rt i ˄13.1˅
i 0
ަѝഎᣕᒿࡇrt+iⲴ⭏ᡀᱟ䙊䗷⭡⣦ᘱstᔰ࿻ᒦ䟽༽֯⭘ㆆ⮕±ᶕ䘹ᤙк䘠Ⲵࣘ֌˄ྲat=±
(st)ˈat+1=±(st+1)ㅹ˅Ǆ䘉䟼 0İ¤<1 Ѫаᑨ䟿ˈᆳ⺞ᇊҶᔦ䘏എᣕо・ণഎᣕⲴ⴨ሩ٬Ǆ⺞
࠷ൠ䇢ˈ൘ᵚᶕⲴㅜiᰦ䰤↕᭦ࡠⲴഎᣕ㻛ഐᆀ¤iԕᤷᮠ㓗ᣈ㇇Ǆ⌘᜿ྲ᷌䇮㖞¤=0ˈ䛓Ѹਚ
㘳㲁・ণഎᣕǄᖃ¤㻛䇮㖞Ѫ᧕䘁 1 Ⲵ٬ᰦˈᵚᶕⲴഎᣕ⴨ሩҾ・ণഎᣕᴹᴤབྷⲴ䟽㾱〻ᓖǄ
±
⭡ᔿ 13.1 ᇊѹⲴ䟿V (s)ᑨ㻛〠Ѫ⭡ㆆ⮕±Ӿࡍ࿻⣦ᘱs㧧ᗇⲴᣈ㇇㍟〟എᣕ˄discounted
cumulative reward˅ǄᢺᵚᶕⲴഎᣕ⴨ሩҾ・ণഎᣕ䘋㹼ᣈ㇇ᱟਸ⨶ⲴˈഐѪ൘䇨ཊⲴᛵߥ
лˈᡁԜ㾱ᐼᵋ㧧ᗇᴤᘛⲴഎᣕǄн䗷ˈަԆⲴᮤփഎᣕᇊѹҏ㻛⹄ウ䗷Ǆֻྲ˖ᴹ䲀≤ᒣഎ
¦
h
ᣕ˄finite horizon reward ˅ᇊѹѪ r
i 0 t i
ˈᆳ䇑㇇ᴹ䲀Ⲵh↕޵എᣕⲴ䶎ᣈ㇇઼Ǆਖа⿽ᇊ
1 h
ѹᯩᔿᱟᒣ൷എᣕ˄average raward ˅ lim hof ¦ rt i Ǆᆳ㘳㲁Ⲵᱟagentᮤњ⭏ભᵏ޵⇿ᰦ
h i0
䰤↕Ⲵᒣ൷എᣕǄᵜㄐਚ䲀ࡦҾ㘳㲁ᔿ 13.1 ᇊѹⲴᣈ㇇എᣕǄMahadevan˄1996˅䇘䇪ҶᖃՈ
ॆ߶ࡉѪᒣ൷എᣕᰦⲴ໎ᕪᆖҐǄ
⧠൘ਟԕ㋮⺞䱸䘠agentⲴᆖҐԫ࣑ǄᡁԜ㾱≲agentᆖҐࡠањㆆ⮕±ˈ֯ᗇሩҾᡰᴹ⣦
±
ᘱsˈV (s)ѪᴰབྷǄ↔ㆆ⮕㻛〠ѪᴰՈㆆ⮕˄optimal policy ˅ˈᒦ⭘±*ᶕ㺘⽪Ǆ
S * { arg max V S ( s), (s) ˄13.2˅

S
S*
Ѫㆰॆ㺘⽪ˈᡁԜሶ↔ᴰՈㆆ⮕Ⲵ٬࠭ᮠ V ( s ) 䇠֌V*(s)ǄV*(s)㔉ࠪҶᖃagentӾ⣦ᘱsᔰ
࿻ᰦਟ㧧ᗇⲴᴰབྷᣈ㇇㍟䇑എᣕˈণӾ⣦ᘱsᔰ࿻䚥ᗚᴰՈㆆ⮕ᰦ㧧ᗇⲴᣈ㇇㍟〟എᣕǄ
ѪҶ䈤᰾䘉Ӌᾲᘥˈമ 13-2 Ⲵкᯩᱮ⽪ҶањㆰঅⲴṬ⣦ц⭼⧟ຳǄ↔മѝⲴ 6 њᯩṬ

ԓ㺘 agent Ⲵ 6 ⿽ਟ㜭Ⲵ⣦ᘱᡆս㖞Ǆമѝ⇿њ㇝ཤԓ㺘 agent ਟ䟷ਆⲴਟ㜭ࣘ֌ˈӾањ⣦
ᘱ〫ࣘࡠਖањǄо⇿њ㇝ཤ⴨‫ޣ‬㚄Ⲵᮠ٬㺘⽪ྲ᷌ agent ᢗ㹼⴨ᓄⲴ⣦ᘱࣘ֌䖜ᦒਟ᭦ࡠⲴ
・ণഎᣕ r(s,a)Ǆ⌘᜿൘䘉њ⢩ᇊ⧟ຳлˈᡰᴹⲴ⣦ᘱࣘ֌䖜ᦒˈ䲔Ҷሬੁ⣦ᘱ G Ⲵԕཆˈ䜭
㻛ᇊѹѪ 0ǄѪ‫ׯ‬Ҿ䇘䇪ˈਟሶ⣦ᘱ G ⴻ֌ᱟⴞḷ⣦ᘱˈഐѪ agent ਟ᧕ਇࡠഎᣕⲴᜏаᯩ⌅
ᱟ䘋‫↔ޕ‬⣦ᘱǄ䘈㾱⌘᜿൘↔⧟ຳлˈagent аᰖ䘋‫ޕ‬⣦ᘱ Gˈᆳਟ䘹Ⲵࣘ֌ਚ㜭ᱟ⮉൘䈕⣦
ᘱѝǄഐ↔ˈᡁԜ〠 G Ѫ੨᭦⣦ᘱ˄absorbing state˅Ǆ
ᡁԜᐢ㓿ᇊѹҶ⣦ᘱǃࣘ֌઼・ণഎᣕˈਚ㾱޽䘹ᤙᣈ㇇ഐᆀ¤Ⲵ٬ˈቡਟԕ⺞ᇊᴰՈㆆ
⮕±*઼ᆳⲴ٬࠭ᮠV*(s)ҶǄ൘䘉䟼ᡁԜ䘹ᤙ¤=0.9Ǆമ 13-2 Ⲵлᯩᱮ⽪Ҷ൘↔䇮ᇊлⲴа⿽
ᴰՈㆆ⮕˄䘈ᴹަԆⲴᴰՈㆆ⮕˅Ǆоԫ᜿ㆆ⮕аṧˈ䈕ㆆ⮕⺞࠷ൠᤷᇊҶagent൘ԫ᜿㔉ᇊ
⣦ᘱлᓄ䘹ᤙⲴањࣘ֌Ǆྲᡰᜣ䊑Ⲵ䛓ṧˈ䈕ᴰՈㆆ⮕ᢺagentԕᴰ⸝䐟ᖴሬੁ⣦ᘱGǄ
(immediate reward)values˖ ・ণഎᣕ٬
values: ٬
values: ٬
One optimal policy˖ањᴰՈㆆ⮕
മ 13-2 䈤᰾ Q-ᆖҐⲴสᵜᾲᘥⲴањㆰঅⲴ⺞ᇊᙗц⭼
⇿њᯩṬԓ㺘ањн਼Ⲵ⣦ᘱˈ⇿њ㇝ཤԓ㺘ањн਼Ⲵࣘ֌Ǆ・ণഎᣕ࠭ᮠᢺ䘋‫ⴞޕ‬ḷ⣦ᘱG
Ⲵഎᣕ䍻Ҹ 100ˈަԆⲴ䍻Ҹ 0ǄV*(s)઼Q(s, a)Ⲵ٬ᶕ㠚Ҿr(s, a)ԕ৺ᣈ㇇ഐᆀ¤=0.9ǄሩᓄҾᴰབྷQ٬
Ⲵࣘ֌ⲴањᴰՈㆆ⮕ҏᱮ⽪൘മѝǄ
മ 13-2 Ⲵਣ䗩Ⲵമᱮ⽪⇿⣦ᘱⲴV*٬Ǆֻྲ˖㘳㲁↔മⲴਣл䀂Ⲵ⣦ᘱǄ↔⣦ᘱⲴV*٬
Ѫ 100ˈഐѪ൘↔⣦ᘱлᴰՈㆆ⮕Պ䘹ᤙĀੁкāⲴࣘ֌ˈӾ㘼ᗇࡠ・ণഎᣕ 100Ǆ❦ਾˈ
agentՊ⮉൘੨᭦⣦ᘱѝˈн޽᧕ࡠᴤཊⲴഎᣕǄ㊫լⲴˈѝлᯩⲴ⣦ᘱⲴV*٬Ѫ 90Ǆ䘉ᱟഐ
ѪᴰՈㆆ⮕Պ֯agentӾ䘉䟼ੁਣ〫ࣘ˄ᗇࡠѪ 0 Ⲵ・ণഎᣕ˅ˈ❦ਾੁк˄⭏ᡀѪ 100 Ⲵ・ণ
എᣕ˅Ǆ䘉ṧˈ↔⣦ᘱⲴᣈ㇇䗷ⲴഎᣕѪ˖
0+¤100+¤20+¤30+...=90
എᗶ V*Ⲵᇊѹѝˈᆳᱟ൘ᰐ䲀ᵚᶕкⲴᣈ㇇എᣕ઼Ǆ൘䘉њ⢩ᇊⲴ⧟ຳлˈаն agent ࡠ
䗮Ҷ੨᭦⣦ᘱ Gˈަᰐ䲀ᵚᶕሶ⮉൘↔⣦ᘱѝᒦ㧧ᗇ 0 എᣕǄ
13.3 Q ᆜҖ
ањagent൘ԫ᜿Ⲵ⧟ຳѝྲօ㜭ᆖࡠᴰՈⲴㆆ⮕±*˛ⴤ᧕ᆖҐ࠭ᮠ±*: SėAᖸഠ䳮ˈഐ
Ѫ䇝㓳ᮠᦞѝ⋑ᴹᨀ‫<׋‬s, a>ᖒᔿⲴ䇝㓳ṧֻǄ֌Ѫᴯԓˈᜏаਟ⭘Ⲵ䇝㓳ؑ᚟ᱟ・ণഎᣕ〻
ᒿࡇr(si,ai)ˈi=0,1,2...ǄྲᡁԜሶⴻࡠⲴˈ㔉ᇊҶ䘉⿽㊫රⲴ䇝㓳ؑ᚟ˈᴤᇩ᱃ⲴᱟᆖҐањ
ᇊѹ൘⣦ᘱ઼ࣘ֌кⲴᮠ٬䇴ՠ࠭ᮠˈ❦ਾԕ↔䇴ՠ࠭ᮠⲴᖒᔿᇎ⧠ᴰՈㆆ⮕Ǆ
Agentᓄቍ䈅ᆖҐӰѸṧⲴ䇴ՠ࠭ᮠ˛ᖸ᰾ᱮⲴањ䘹ᤙᱟV* Ǆਚ㾱ᖃV*(s1)>V*(s2)ᰦˈ
agent䇔Ѫ⣦ᘱs1ՈҾs2ˈഐѪӾs1ѝਟᗇࡠ䖳བྷⲴ・ণഎᣕǄᖃ❦agentⲴㆆ⮕㾱䘹ᤙⲴᱟࣘ֌
㘼䶎⣦ᘱǄ❦㘼൘ਸ䘲Ⲵ䇮ᇊѝ֯⭘V* ҏਟ䘹ᤙࣘ֌Ǆ൘⣦ᘱsлⲴᴰՈࣘ֌ᱟ֯・ণഎᣕ
r(s,a)࣐к・ণਾ㔗⣦ᘱⲴV*٬˄㻛¤ᣈ㇇˅ᴰབྷॆⲴࣘ֌aǄ
S * ( s ) arg max[r ( s, a) JV * (G ( s, a))] ˄13.3˅

a
˄എᗶ¥(s,a)ԓ㺘ᓄ⭘ࣘ֌aࡠ⣦ᘱsⲴ㔃᷌⣦ᘱ˅Ǆഐ↔ˈagentਟ䙊䗷ᆖҐV*㧧ᗇᴰՈㆆ
⮕ⲴᶑԦᱟ˖ᆳާᴹ・ণഎᣕ࠭ᮠr઼⣦ᘱ䖜ᦒ࠭ᮠ¥Ⲵᆼ㖾⸕䇶Ǆᖃagentᗇ⸕Ҷཆ⭼⧟ຳ⭘
ᶕ૽ᓄࣘ֌Ⲵ࠭ᮠr઼¥Ⲵᆼ㖾⸕䇶ˈᆳቡਟ⭘ᔿ 13.3 ᶕ䇑㇇ԫ᜿⣦ᘱлⲴᴰՈࣘ֌Ǆ
нᒨⲴᱟˈਚ൘agentާᴹr઼¥ᆼ㖾⸕䇶ᰦˈᆖҐV*᡽ᱟᆖҐᴰՈㆆ⮕Ⲵᴹ᭸ᯩ⌅Ǆ䘉㾱
≲ᆳ㜭ᆼ㖾亴⍻ԫ᜿⣦ᘱ䖜ᦒⲴ・ণ㔃᷌˄ণ・ণഎᣕ઼・ণਾ㔝˅Ǆ൘䇨ཊᇎ䱵Ⲵ䰞仈ѝˈ
∄ྲᵪಘӪ᧗ࡦˈagentԕ৺ᆳⲴ〻ᒿ䇮䇑㘵䜭нਟ㜭亴‫⸕ݸ‬䚃ᓄ⭘ԫ᜿ࣘ֌ࡠԫ᜿⣦ᘱⲴ⺞
࠷䗃ࠪǄֻྲˈሩҾањ⭘᡻㟲䬢൏ⲴᵪಘӪˈᖃ㔃᷌⣦ᘱवਜ਼൏ඇⲴ⣦ᘱᰦˈྲօ᧿䘠¥࠭
ᮠ˛ഐ↔ᖃ¥ᡆr䜭ᵚ⸕ᰦˈᆖҐV*ᱟᰐࣙҾ䘹ᤙᴰՈࣘ֌ⲴˈഐѪagentн㜭⭘ᔿ 13-3 䘋㹼䇴
ՠǄ൘ᴤа㡜Ⲵ䘹ᤙѝˈagentᓄ֯⭘ӰѸṧⲴ䇴ՠ࠭ᮠ઒˛ла㢲ᇊѹⲴ䇴ՠ࠭ᮠQᨀ‫׋‬Ҷㆄ
ṸǄ
13.3.1 Q ࠳ᮦ
䇴ՠ࠭ᮠ Q(s,a)ᇊѹѪ˖ᆳⲴ٬ᱟӾ⣦ᘱ s ᔰ࿻ᒦ֯⭘ࣘ֌ a ֌Ѫㅜањࣘ֌ᰦⲴᴰབྷᣈ

㇇㍟〟എᣕǄᦒ䀰ѻˈQ Ⲵ٬ѪӾ⣦ᘱ s ᢗ㹼ࣘ֌ a Ⲵ・ণഎᣕ࣐кԕਾ䚥ᗚᴰՈㆆ⮕Ⲵ٬
˄⭘¤ᣈ㇇˅Ǆ
Q( s, a ) { r ( s, a ) JV * (G ( s, a )) (13.4)
⌘᜿ Q(s,a)↓ᱟᔿ 13.3 ѝѪ䘹ᤙ⣦ᘱ s кⲴᴰՈࣘ֌ a ᓄᴰབྷॆⲴ䟿ˈഐ↔ਟሶᔿ 13.3 䟽
߉Ѫ Q(s,a)Ⲵᖒᔿ˖
S * ( s ) arg max Q( s, a) ˄13.5˅

a
䟽߉䈕ᔿѪӰѸᖸ䟽㾱˛ഐѪᆳᱮ⽪Ҷྲ᷌agentᆖҐQ࠭ᮠ㘼нᱟV*࠭ᮠˈণ֯൘㕪ቁ࠭
ᮠr઼¥Ⲵ⸕䇶ᰦˈagentҏਟ䘹ᤙᴰՈࣘ֌Ǆᔿ 13.5 ␵ᾊൠᱮ⽪ࠪˈagentਚ享㘳㲁ަᖃࡽⲴ⣦
ᘱsл⇿њਟ⭘Ⲵࣘ֌aˈᒦ䘹ᤙަѝ֯Q(s,a)ᴰབྷॆⲴࣘ֌Ǆ
䘉а⛩ᔰ࿻ⴻ䎧ᶕԔӪ᛺ཷˈਚ享ሩᖃࡽⲴ⣦ᘱⲴ Q Ⲵተ䜘٬䟽༽‫৽ࠪڊ‬ᓄˈቡਟ䘹ᤙ
ࡠ‫ޘ‬ተᴰՈॆⲴࣘ֌ᒿࡇˈ䘉᜿ણ⵰ agent н享䘋㹼ࡽⷫᙗᩌ㍒ˈн享᰾⺞ൠ㘳㲁Ӿ↔ࣘ֌ᗇ
ࡠⲴ⣦ᘱˈቡਟ䘹ᤙᴰՈࣘ֌ǄQ ᆖҐⲴ㖾࿉ѻ༴а䜘࠶൘Ҿަ䇴ՠ࠭ᮠⲴᇊѹ㋮⺞ൠᤕᴹ↔
኎ᙗ˖ᖃࡽ⣦ᘱ઼ࣘ֌Ⲵ Q ٬൘অњⲴᮠ٬ѝᾲᤜҶᡰᴹ䴰㾱Ⲵؑ᚟ˈԕ⺞ᇊ൘⣦ᘱ s л䘹ᤙ
ࣘ֌ a ᰦ൘ሶᶕՊ㧧ᗇⲴᣈ㇇㍟䇑എᣕǄ
Ѫ䈤᰾䘉а⛩ˈ㿱മ 13-2Ǆަѝ൘ㆰঅⲴṬᆀц⭼ѝᱮ⽪Ҷ⇿њ⣦ᘱ઼ࣘ֌ⲴQ٬Ǆ⌘᜿
⇿њ⣦ᘱࣘ֌Ⲵ䖜ᦒⲴQ٬ㅹҾ↔䖜ᦒⲴr٬࣐к㔃᷌⣦ᘱⲴV*٬˄⭘¤ᣈ㇇˅Ǆ䘈㾱⌘᜿മѝ
ᱮ⽪ⲴᴰՈㆆ⮕ሩᓄҾ䘹ᤙᴹᴰབྷⲴQ٬Ⲵࣘ֌Ǆ
13.3.2 ᆜҖ Q ⲺжѠ㇍⌋
ᆖҐ Q ࠭ᮠሩᓄҾᆖҐᴰՈㆆ⮕ǄQ ᘾṧ㜭㻛ᆖҐࡠ˛
‫ޣ‬䭞൘Ҿ㾱᢮ࡠањਟ䶐Ⲵᯩ⌅ˈ൘ਚᴹᰦ䰤кኅᔰⲴ・ণഎᣕᒿࡇⲴᛵߥлՠ䇑䇝㓳
٬Ǆ䘉ਟ䙊䗷䘝ԓ䙬䘁Ⲵᯩ⌅ᆼᡀǄѪ⨶䀓ᘾṧᆼᡀ䘉а䗷〻ˈ⌘᜿Q઼V*ѻ䰤Ⲵᇶ࠷㚄㌫˖
V * ( s) max Q( s, a c)
ac
ᆳਟ㻛⭘Ҿ䟽߉ᔿ 13.4 Ѫ
Q( s, a) { r ( s, a) J max Q (G ( s, a), a c) ˄13.6˅

ac
䘉њ Q ࠭ᮠⲴ䙂ᖂᇊѹᨀ‫׋‬Ҷᗚ⧟䙬䘁 Q ㇇⌅Ⲵส⹰˄Watkins 1989˅ǄѪ᧿䘠↔㇇⌅ˈ

ᡁԜሶ֯⭘ㅖਧ Qˆ ᶕᤷԓᇎ䱵 Q ࠭ᮠⲴᆖҐಘⲴՠ䇑ˈᡆ㘵䈤‫ٷ‬䇮Ǆ൘↔㇇⌅ѝᆖҐಘ㺘⽪
ަ‫ٷ‬䇮 Qˆ ᱟ䙊䗷ањབྷ㺘ˈަѝሩ⇿њ⣦ᘱ-ࣘ֌ሩᴹа㺘亩Ǆ⣦ᘱ-ࣘ֌ሩ<s, a>Ⲵ㺘亩ѝᆈ
ۘҶ Qˆ (s,a)Ⲵ٬ˈণᆖҐಘሩᇎ䱵Ⲵնᵚ⸕Ⲵ Q(s,a)٬Ⲵᖃࡽ‫ٷ‬䇮Ǆ↔㺘ਟ㻛ࡍ࿻ປ‫ݵ‬Ѫ䲿ᵪ
٬˄ᖃ❦ˈྲ᷌䇔Ѫᱟ‫ ޘ‬0 Ⲵࡍ࿻٬ᴤ᱃Ҿ⨶䀓˅ǄAgent 䟽༽ൠ㿲ሏަᖃࡽⲴ⣦ᘱ sˈ䘹ᤙ
Ḁࣘ֌ aˈᢗ㹼↔ࣘ֌ˈ❦ਾ㿲ሏ㔃᷌എᣕ r=r(s,a)ԕ৺ᯠ⣦ᘱ s´=¥(s,a)Ǆ❦ਾ agent 䚥ᗚ⇿њ
䘉ṧⲴ䖜ᦒᴤᯠ Qˆ (s,a)Ⲵ㺘亩ˈ᤹➗ԕлⲴ㿴ࡉ˖
Qˆ ( s, a) m r J max Qˆ ( s c, a c) ˄13.7˅
ac
⌘᜿↔䇝㓳٬֯⭘ agent ሩᯠ⣦ᘱ s´Ⲵᖃࡽ Qˆ ٬ᶕ㋮ॆަሩࡽа⣦ᘱ s Ⲵ Qˆ (s,a)ՠ䇑Ǆ↔
䇝㓳㿴ࡉᱟӾᔿ 13.6 ѝᗇࡠⲴˈн䗷↔䇝㓳٬㘳㲁 agent Ⲵ䘁լ Q̂ ˈ㘼ᔿ 13.6 ᓄ⭘ࡠᇎ䱵Ⲵ Q
࠭ᮠǄ⌘᜿㲭❦ᔿ 13.6 ԕ࠭ᮠ¥(s,a)઼ r(s,a)Ⲵᖒᔿ᧿䘠 Qˈagent н䴰⸕䚃䘉Ӌа㡜࠭ᮠᶕᓄ
⭘ᔿ 13.7 Ⲵ䇝㓳㿴ࡉǄ⴨৽ˈᆳ൘ަ⧟ຳѝᢗ㹼ࣘ֌ˈᒦ㿲ሏ㔃᷌⣦ᘱ s´઼എᣕ rǄ䘉ṧˈᆳ
ਟ㻛ⴻ֌ᱟ൘ s ઼ a Ⲵᖃࡽ٬к䟷ṧǄ
к䘠ሩҾ⺞ᇊᙗ傜ቄਟཛߣㆆ䗷〻Ⲵ Q ᆖҐ㇇⌅൘㺘 13-1 ѝ㻛ᴤ㋮⺞ൠ᧿䘠Ǆ֯⭘↔㇇

⌅ˈagent ՠ䇑Ⲵ Q̂ ൘ᶱ䲀ᰦ᭦ᮋࡠᇎ䱵 Q ࠭ᮠˈਚ㾱㌫㔏ਟ㻛ᔪ⁑Ѫањ⺞ᇊᙗ傜ቄਟཛߣ
ㆆ䗷〻ˈഎᣕ࠭ᮠ r ᴹ⭼ˈᒦфࣘ֌Ⲵ䘹ᤙਟ֯⇿њ⣦ᘱ-ࣘ֌ሩ㻛ᰐ䲀仁⦷Ⲵ䇯䰞Ǆ
㺘 13-1 ൘⺞ᇊᙗഎᣕ઼ࣘ֌‫ٷ‬ᇊлⲴ Q ᆖҐ㇇⌅
ᣈ㇇ഐᆀJѪԫ᜿ᑨ䟿┑䏣 0dJ<1Ǆ
Q ᆖҐ㇇⌅
ˆ (s,a)Ѫ 0
ሩ⇿њ s,aˈࡍ࿻ॆ㺘亩 Q
㿲ሏᖃࡽ⣦ᘱ s
аⴤ䟽༽‫˖ڊ‬
䘹ᤙањࣘ֌ a ᒦᢗ㹼ᆳ
᧕᭦ࡠ・ণഎᣕ r
㿲ሏᯠ⣦ᘱ s´
ˆ (s,a)᤹➗лᔿᴤᯠ㺘亩˖
ሩQ
Qˆ ( s, a) m r J max Qˆ ( s c, a c)
ac
sĕs´
13.3.3 ⽰ׁ
Ѫ䈤᰾ Q ᆖҐ㇇⌅Ⲵ᫽֌䗷〻ˈ㘳㲁മ 13-3 ᱮ⽪ⲴḀњ agent 䟷ਆⲴањࣘ֌઼ሩᓄⲴሩ

Q̂ Ⲵ㋮ॆǄ൘↔ֻѝˈagent ൘ަṬᆀц⭼ѝੁਣ〫ࣘањঅ‫ݳ‬Ṭˈᒦ᭦ࡠ↔䖜ᦒⲴ・ণഎᣕ
Ѫ 0Ǆ❦ਾᆳᓄ⭘䇝㓳㿴ࡉᔿ 13.7 ᶕሩࡊᢗ㹼Ⲵ⣦ᘱ-ࣘ֌䖜ᦒ㋮ॆަ Qˆ Ⲵՠ䇑Ǆ᤹➗䇝㓳㿴
ࡉˈ↔䖜ᦒⲴᯠ Q̂ ՠ䇑Ѫ᭦ࡠⲴഎᣕ˄0˅о⭘J˄0.9˅ᣈ㇇Ⲵо㔃᷌⣦ᘱ⴨‫ޣ‬㚄Ⲵᴰ儈 Qˆ ٬
˄100˅Ⲵ઼Ǆ
⇿⅑ agent Ӿаᰗ⣦ᘱࡽ䘋ࡠаᯠ⣦ᘱˈQ ᆖҐՊӾᯠ⣦ᘱࡠᰗ⣦ᘱੁਾՐ᫝ަ Qˆ ՠ䇑Ǆ

਼ᰦˈagent ᭦ࡠⲴ↔䖜ᦒⲴ・ণഎᣕ㻛⭘Ҿᢙབྷ䘉ӋՐ᫝Ⲵ Q̂ ٬Ǆ
㘳㲁ሶ↔㇇⌅ᓄ⭘ࡠṬᆀц⭼ѝˈަഎᣕ࠭ᮠᱮ⽪൘മ 13-2 ѝˈަѝⲴഎᣕ٬䲔Ҷ䘋‫ޕ‬

ⴞḷ⣦ᘱⲴԕཆ䜭Ѫ 0ǄഐѪ↔ц⭼वਜ਼ањ੨᭦ⴞḷ⣦ᘱǄᡁԜਟ‫ٷ‬ᇊ䇝㓳䗷〻वਜ਼а㌫ࡇ
Ⲵᛵ㢲˄episode˅Ǆ൘⇿њᛵ㢲Ⲵ䗷〻ѝˈagent ӾḀ䲿ᵪ䘹ᤙⲴ⣦ᘱᔰ࿻Ǆᢗ㹼ࣘ֌ⴤࡠަ
ࡠ䗮੨᭦ⴞḷ⣦ᘱǄ䘉ᰦᛵ㢲㔃ᶏˈ❦ਾ agent 㻛䘀䗃ࡠањ䲿ᵪ䘹ᤙⲴᯠࡍ࿻⣦ᘱᔰ࿻ла
њᛵ㢲Ǆ
Initial state: ࡍ࿻⣦ᘱ
Next state: ла⣦ᘱ
മ 13-3 ൘ᢗ㹼অњࣘ֌ਾሩ Q Ⲵᴤᯠ
ᐖ 䗩 Ⲵ മ ᱮ ⽪ Ҷ ᵪ ಘ Ӫ R Ⲵ ࡍ ࿻ ⣦ ᘱ s1 ˈ ԕ ৺ ࡍ ࿻ ‫ ٷ‬䇮 ѝ ࠐ њ ⴨ ‫Ⲵ ޣ‬ Q̂ ٬ Ǆ ֻ ྲ ˈ
Qˆ ( s1 , a right ) =72.9ˈަѝarightᤷԓRੁਣ〫ࣘⲴࣘ֌ǄᖃᵪಘӪᢗ㹼ࣘ֌arightਾˈᆳ᭦ࡠ・ণഎᣕr=
0ˈᒦ䖜ᦒࡠ⣦ᘱs2Ǆ❦ਾᆳสҾަሩᯠ⣦ᘱs2Ⲵ Q̂ ՠ䇑ᴤᯠަ Q ˆ ( s , a ) ՠ䇑Ǆ䘉䟼¤=0.9Ǆ
1 right
൘↔ֻѝˈᖃᓄ⭘ Q ᆖҐ㇇⌅ᰦˈ Q̂ Ⲵ٬ᱟྲօ╄ॆⲴ˛ഐѪࡍ࿻Ⲵ Q̂ ٬䜭Ѫ 0ˈagent

ሩԫ᜿ Q̂ 㺘亩䜭нՊ᭩ਈˈⴤࡠᆳᚠྭࡠ䗮ⴞḷ⣦ᘱᒦф᭦ࡠ䶎䴦ⲴഎᣕǄ൘лањᛵ㢲ѝˈ
ྲ᷌㓿䗷䘉Ӌоⴞḷ⣦ᘱ⴨ѤⲴ⣦ᘱˈަ䶎 0 Ⲵ Q̂ ٬Պሬ㠤оⴞⲴ⴨ᐞє↕Ⲵ⣦ᘱѝ٬Ⲵਈ
ॆˈ‫↔׍‬㊫᧘Ǆ㔉ᇊ䏣ཏᮠ䟿Ⲵ䇝㓳ᛵ㢲ˈؑ᚟ՊӾᴹ䶎䴦എᣕⲴ䖜ᦒੁਾՐ᫝ࡠᮤњ⣦ᘱ-
ࣘ֌オ䰤ˈᴰ㓸ᗇࡠањ Q̂ 㺘ǄަѝⲴ Q ٬ྲമ 13-2 ᡰ⽪Ǆ
൘ла㢲ˈᡁԜ䇱᰾Ҷ൘аᇊᶑԦл㺘 13-1 ⲴQᆖҐ㇇⌅Պ᭦ᮋࡠ↓⺞ⲴQ࠭ᮠǄ俆‫ݸ‬㘳

㲁↔QᆖҐ㇇⌅Ⲵєњ⢩⛩ˈ䘉єњ⢩⛩ᱟሩഎᣕ䶎䍏фᡰᴹ Q̂ ٬ࡍ࿻ॆѪ 0 Ⲵԫ᜿⺞ᇊᙗⲴ
MDP䜭Პ䙽ᆈ൘ⲴǄㅜањ኎ᙗᱟˈ൘к䘠ᶑԦл Q̂ ٬൘䇝㓳ѝ≨䘌нՊл䱽Ǆᴤᖒᔿॆൠ
䇢ˈԔ Q̂ n(s,a)㺘⽪䇝㓳䗷〻Ⲵㅜn⅑ᗚ⧟ਾᆖҐࡠⲴ Q̂ (s,a)٬˄ণagentᡰ䟷ਆⲴㅜnњ⣦ᘱ-ࣘ
֌䖜ᦒѻਾ˅Ǆࡉᴹ˖
(s, a, n)Qˆ n 1 ( s, a ) t Qˆ n ( s, a )
ㅜҼњᲞ䙽⢩⛩ᱟ൘ᮤњ䇝㓳䗷〻ѝˈ⇿њ Q̂ ٬ሶ‫؍‬ᤱ൘䴦઼ⵏᇎ Q ٬Ⲵ४䰤޵˖
(s, a, n)0 d Qˆ n ( s, a ) d Q ( s, a )
13.3.4 ᭬ᮑᙝ
㺘 13-1 Ⲵ㇇⌅ᱟ੖Պ᭦ᮋࡠањㅹҾⵏᇎ Q ࠭ᮠⲴ Q̂ ٬˛൘⢩ᇊᶑԦлˈഎㆄᱟ㛟ᇊ

ⲴǄ俆‫ݸ‬䴰㾱‫ٷ‬ᇊ㌫㔏Ѫа⺞ᇊᙗⲴ MDPǄަ⅑ˈᗵ享‫ٷ‬ᇊ・ণഎᣕ٬䜭ᱟᴹ⭼Ⲵˈণᆈ൘
Ḁ↓ᑨᮠ cˈሩᡰᴹ⣦ᘱ s ઼ࣘ֌ aˈ|r(s, a)|<cǄㅜйˈagent 䘹ᤙࣘ֌ⲴᯩᔿѪᆳᰐ䲀仁㑱ൠ
䇯䰞ᡰᴹਟ㜭Ⲵ⣦ᘱ-ࣘ֌ሩǄ䘉њᶑԦ᜿ણ⵰ྲ᷌ࣘ֌ a ᱟӾ⣦ᘱ s ࠪਁⲴањਸ⌅Ⲵࣘ֌ˈ
䛓Ѹ䲿ᰦ䰤Ⲵ㍟䇑ˈagent Ⲵࣘ֌ᒿࡇ䙀⑀䗮ࡠᰐ䲀䮯Ǆagent ᗵ享ԕ䶎 0 Ⲵ仁⦷䟽༽ൠӾ⣦ᘱ
s ᢗ㹼ࣘ֌ aǄ⌘᜿䘉ӋᶑԦ൘Ḁ⿽〻ᓖкॱ࠶а㡜ˈնᴹᰦ৸⴨ᖃѕṬǄᆳԜ᧿䘠Ҷ∄ࡽа
㢲ᡰѮⲴֻᆀѝᴤа㡜Ⲵ䇮ᇊˈഐѪᆳԜ‫ݱ‬䇨⧟ຳᴹԫ᜿Ⲵ↓ᡆ䍏എᣕˈᒦф⧟ຳѝਟᴹԫ᜿
ᮠ䟿Ⲵ⣦ᘱ-ࣘ֌䖜ᦒਟӗ⭏䶎䴦എᣕǄ䘉ӋᶑԦⲴѕṬᙗ൘Ҿᆳ㾱≲ agent ᰐ䲀仁㑱Ⲵ䇯䰞⇿
њн਼Ⲵ⣦ᘱ-ࣘ֌䖜ᦒǄ䘉൘䶎ᑨབྷⲴ˄ᡆ㘵⭊㠣ᱟ䘎㔝Ⲵ˅亶ฏѝᱟ䶎ᑨᕪⲴ‫ٷ‬ᇊǄᡁԜ
ሶ൘ਾ䶒䇘䇪ᴤᕪⲴ᭦ᮋ㔃᷌Ǆ❦㘼ᵜ㢲᧿䘠Ⲵ㔃᷌ሶѪ⨶䀓 Q ᆖҐⲴ䘀㹼ᵪࡦᨀ‫ⴤ׋‬㿲Ⲵ
⨶䀓Ǆ
ሩ᭦ᮋᙗ䇱᰾Ⲵ‫ޣ‬䭞ᙍ䐟൘Ҿˈᴹᴰབྷ䈟ᐞⲴ㺘亩 Q̂ (s,a)ᗵ享൘ަᴤᯠᰦሶ䈟ᐞ᤹ഐᆀ¤
߿ሿǄ৏ഐ൘ҾᆳⲴᯠ٬Ⲵа䜘࠶‫׍‬䎆Ҿᴹ䈟ᐞ‫ Ⲵੁٮ‬Q̂ ՠ䇑ˈަ։Ⲵ䜘࠶‫׍‬䎆Ҿᰐ䈟ᐞⲴ㿲
ሏࡠⲴ・ণഎᣕ rǄ
ᇊ⨶ 13.1Ǆ⺞ ᇊᙗ傜ቄਟཛߣㆆ䗷〻ѝⲴQᆖҐⲴ᭦ᮋᙗǄ㘳㲁ањQᆖҐagentˈ൘ањ⺞ᇊᙗ
MDPѝˈᒦфᴹᴹ⭼എᣕ (s , a ) r ( s , a ) d c ǄQᆖҐagent֯⭘ᔿ 13.7 Ⲵ䇝㓳㿴ࡉˈሶ㺘 Q̂ (s,a)ࡍ࿻
ॆѪԫ᜿ᴹ䲀٬ˈᒦф֯⭘ᣈ㇇ഐᆀ¤ˈ0İ¤<1ǄԔ Q̂ n(s,a)ԓ㺘൘ㅜn⅑ᴤᯠਾagentⲴ‫ٷ‬䇮 Q̂ (s,a)Ǆ
ྲ᷌⇿њ⣦ᘱ-ࣘ֌ሩ䜭㻛ᰐ䲀仁㑱Ⲵ䇯䰞ˈ䛓ѸሩᡰᴹsˈaˈᖃnėĞᰦ Q̂ n(s,a)᭦ᮋࡠQ(s,a)Ǆ
䇱᰾˖ഐѪ⇿њ⣦ᘱ-ࣘ֌䖜ᦒᰐ䲀仁㑱ਁ⭏ˈ㘳㲁䘎㔝Ⲵ४䰤ˈަѝ⇿њ⣦ᘱ-ࣘ֌䖜ᦒ㠣ቁਁ⭏
䗷а⅑Ǆᡰ䴰㾱䇱᰾Ⲵᱟˈ൘ Q̂ 㺘ѝᡰᴹ㺘亩кⲴᴰབྷ䈟ᐞ൘⇿њ䘉ṧⲴ䘎㔝४䰤޵㠣ቁ᤹ഐᆀ ¤߿
ቁǄ Q̂ nѪn⅑ᴤᯠਾagentՠ䇑ⲴQ٬㺘ǄԔnѪ Q̂ nѝᴰབྷ䈟ᐞˈণ˖
' n { max Qˆ n ( s, a ) Q ( s, a )
s ,a
л䶒ᡁԜ֯⭘s´ᶕԓ㺘¥(s,a)ˈ⧠൘ሩ൘ㅜn+1 ⅑䘝ԓѝᴤᯠⲴԫ᜿㺘亩 Q̂ n(s,a)ˈ൘‫↓؞‬ਾⲴՠ䇑
Q̂ n+1(s,a)ѝⲴ䈟ᐞ䟿Ѫ˖
Qˆ n 1 ( s, a ) Q( s, a) (r J max Qˆ n ( s c, a c)) (r J max Q( s c, a c))

ac ac
J max Qˆ n ( s c, a c) max Q( s c, a c)
ac ac
d J max Qˆ n ( s c, a c) Q( s c, a c)
ac
d J max Qˆ n ( s cc, a c) Q ( s cc, a c)

s cc, ac
Qˆ n 1 ( s, a) Q( s, a) d J' n
к䶒Ⲵㅜй㹼ӾㅜҼ㹼ѝሬࠪˈ৏ഐᱟሩԫ᜿єњ࠭ᮠf1઼f2ᴹлࡇнㅹᔿᡀ・˖
max f1 (a ) max f 2 (a) d max f1 (a ) f 2 (a)

a a a
Ӿㅜй㹼ࡠㅜഋ㹼Ⲵ᧘ሬˈᡁԜᕅ‫ޕ‬Ҷањᯠਈ䟿s´´൘ަкᢗ㹼ᴰབྷॆǄަਸ⨶ᙗ൘ҾᖃᡁԜ‫ݱ‬䇨
䱴࣐Ⲵਈ䟿ਈॆᰦˈ↔ᴰབྷ٬ਚਟ㜭ᴤབྷᡆ㠣ቁᱟ⴨ㅹǄ⌘᜿ˈ䙊䗷ᕅ‫↔ޕ‬ਈ䟿ˈᡁԜ㧧ᗇҶањоn
Ⲵᇊѹ३䝽Ⲵ㺘䗮ᔿǄ
ഐ↔ˈሩԫ᜿s, aˈᴤᯠਾⲴ Q̂ n+1(s,a)Ⲵ䈟ᐞᴰཊѪ Q̂ n㺘ѝᴰབྷ䈟ᐞnⲴ¤‫ؽ‬Ǆ൘ࡍ࿻㺘ѝⲴᴰ
བྷ䈟ᐞ0ᱟᴹ⭼ⲴˈഐѪ Q̂ 0(s,a)઼Q(s,a)Ⲵ٬ሩᡰᴹs, a䜭ᴹ⭼Ǆ⧠൘ˈ൘⇿њs,a䜭㻛䇯䰞䗷Ⲵㅜањ

४䰤޵ˈ↔㺘ѝᴰབྷⲴ䈟ᐞ㠣ཊѪ¤0Ǆ൘kњ४䰤ਾˈ䈟ᐞᴰཊѪ¤k0ǄഐѪ⇿њ⣦ᘱ䜭㻛ᰐ䲀仁㑱
ൠ䇯䰞ˈ䘉ṧⲴ४䰤Ⲵᮠⴞᱟᰐ䲀Ⲵˈഐ↔ᖃnėĞᰦnė0Ǆᇊ⨶ᗇ䇱Ǆ
13.3.5 ᇔ僂ㆌ⮛
⌘᜿㺘 13-1 Ⲵ㇇⌅⋑ᴹᤷᇊ agent ྲօ䘹ᤙࣘ֌Ǆањ᰾ᱮⲴㆆ⮕ᱟˈሩҾ൘⣦ᘱ s Ⲵ

agentˈ䘹ᤙ֯ Q̂ (s,a)ᴰབྷॆⲴࣘ֌ˈӾ㘼࡙⭘ަᖃࡽ䙬䘁Ⲵ Q̂ Ǆ❦㘼ˈ֯⭘↔ㆆ⮕ᆈ൘仾
䲙ˈagent ਟ㜭䗷ᓖᶏ㕊ࡠ൘ᰙᵏ䇝㓳ѝᴹ儈 Qˆ ٬Ⲵࣘ֌ˈ㘼н㜭ཏ᧒㍒ࡠަԆਟ㜭ᴹᴤ儈٬
ᗇࣘ֌Ǆᇎ䱵кˈк䶒Ⲵ᭦ᮋᙗᇊ⨶㾱≲⇿њ⣦ᘱ-ࣘ֌䖜ᦒᰐ䲀仁㑱ൠਁ⭏Ǆᱮ❦ˈྲ᷌
agent ᙫ䘹ᤙ֯ᖃࡽ Q̂ (s,a)ᴰབྷⲴࣘ֌ˈሶн㜭‫؍‬䇱↔ᰐ䲀仁㑱ᙗǄഐ↔ˈ൘ Q ᆖҐѝ䙊ᑨ֯
⭘ᾲ⦷Ⲵ䙄ᖴᶕ䘹ᤙࣘ֌Ǆᴹ䖳儈 Qˆ ٬Ⲵࣘ֌㻛䍻Ҹ䖳儈Ⲵᾲ⦷ˈնᡰᴹࣘ֌Ⲵᾲ⦷䜭䶎 0Ǆ
䍻Ҹ䘉⿽ᾲ⦷Ⲵа⿽ᯩ⌅ᱟ
ˆ
k Q ( s ,ai )
P(ai | s )
¦
Qˆ ( s , a j )
j
k
ަѝP(ai|s)Ѫagent൘⣦ᘱsᰦ䘹ᤙࣘ֌aiⲴᾲ⦷ˈk>0 Ѫаᑨ䟿ˈᆳ⺞ᇊ↔䘹ᤙՈ‫ݸ‬㘳㲁儈
Q̂ ٬Ⲵ〻ᓖǄ䖳བྷⲴk٬Պሶ䖳儈Ⲵᾲ⦷䍻Ҹ䎵ࠪᒣ൷ Q̂ Ⲵࣘ֌ˈ㠤֯agent࡙⭘ᆳᡰᆖҐࡠⲴ
⸕䇶ᶕ䘹ᤙᆳ䇔ѪՊ֯എᣕᴰབྷⲴࣘ֌Ǆ⴨৽ˈ䖳ሿⲴk٬Պ֯ަԆࣘ֌ᴹ䖳儈Ⲵᾲ⦷ˈሬ㠤
agent᧒㍒䛓Ӌᖃࡽ Q̂ ٬䘈н儈Ⲵࣘ֌Ǆ൘ḀӋᛵߥлˈkᱟ䲿⵰䘝ԓ⅑ᮠ㘼ਈॆⲴǄԕ֯agent
൘ᆖҐⲴᰙᵏਟ⭘᧒㍒රㆆ⮕ˈ❦ਾ䙀↕䖜ᦒࡠ࡙⭘රⲴㆆ⮕Ǆ
13.3.6 ᴪ᯦ᓅࡍ
к䶒᭦ᮋᙗᇊ⨶ањ䟽㾱ᙗ᳇⽪൘ҾˈQ ᆖҐн䴰㾱⭘ᴰՈࣘ֌䘋㹼䇝㓳ˈቡਟԕ᭦ᮋࡠ
ᴰՈㆆ⮕Ǆᇎ䱵кˈਚ㾱⇿↕Ⲵ䇝㓳ࣘ֌ᆼ‫ޘ‬䲿ᵪ䘹ᤙˈ֯ᗇ㔃᷌䇝㓳ᒿࡇᰐ䲀仁㑱Ⲵ䇯䰞⇿
њ⣦ᘱ-ࣘ֌䖜ᦒˈቡਟԕᆖҐࡠ Q ࠭ᮠ˄ԕ৺ᴰՈㆆ⮕˅Ǆ䘉аһᇎᔪ䇞᭩ਈ䇝㓳䖜ᦒṧֻ
Ⲵᒿࡇˈԕ᭩䘋䇝㓳᭸⦷㘼нড৺ᴰ㓸Ⲵ᭦ᮋᙗǄѪ䈤᰾䘉а⛩ˈ޽⅑㘳㲁൘ањ MDP ѝᴹ
অњ੨᭦ⴞḷ⣦ᘱⲴᆖҐ䗷〻ˈྲ਼ 13-1 ѝᡰ⽪Ǆྲԕࡽ䛓ṧˈ‫ٷ‬ᇊ֯⭘ᒿࡇॆⲴᛵ㢲
˄episode˅ᶕ䇝㓳 agentǄሩ⇿њᛵ㢲ˈagent 㻛᭮㖞൘ањ䲿ᵪࡍ࿻⣦ᘱˈ❦ਾᢗ㹼ࣘ֌ԕᴤ
ᯠަ Q̂ 㺘ˈⴤࡠᆳࡠ䗮੨᭦⣦ᘱǄ❦ਾᔰ࿻ањᯠⲴ䇝㓳ᛵ㢲ˈ䙊䗷ሶ agent Ӿⴞḷ⣦ᘱ䖜ᦒ
ࡠањᯠⲴ䲿ᵪࡍ࿻⣦ᘱǄྲࡽ䶒ᤷࠪⲴˈྲ᷌ᔰ࿻ᡰᴹ Q̂ ٬Ⲵࡍ࿻ॆѪ 0ˈࡉ൘ㅜањᛵ㢲
ਾˈagent Ⲵ Qˆ 㺘ѝਚᴹањ㺘亩᭩ਈ˖ণሩᓄҾᴰਾ䖜ᦒࡠⴞḷ⣦ᘱⲴ㺘亩Ǆྲ᷌൘ㅜҼњ
ᛵ㢲ѝˈagent ᚠྭӾ⴨਼Ⲵ䲿ᵪࡍ࿻⣦ᘱ⋯⵰⴨਼ࣘ֌ᒿࡇ〫ࣘˈࡉਖа㺘亩ਈѪ䶎 0ˈ‫↔׍‬
㊫᧘Ǆྲ᷌䟽༽ൠԕ⴨਼Ⲵᯩᔿ䘀㹼ᛵ㢲ˈ䶎 0 Qˆ ٬Ⲵ䗩㕈䙀⑀ੁਣ〫ࣘˈӾⴞḷ⣦ᘱᔰ࿻ˈ
⇿њᛵ㢲〫ࣘࡠањᯠⲴ⣦ᘱ-ࣘ֌䖜ᦒǄ⧠൘㘳㲁൘䘉Ӌ⴨਼Ⲵ⣦ᘱ-ࣘ֌䖜ᦒкⲴ䇝㓳ˈն
ሩ⇿њᛵ㢲ԕ৽ੁⲴᰦᒿǄণሩ⇿њ㘳㲁Ⲵ䖜ᦒᓄ⭘ᔿ 13-7 ѝ⴨਼Ⲵᴤᯠ㿴ࡉˈնԕ䘶ᒿᢗ
㹼䘉ӋᴤᯠǄ䘉ṧˈ൘ㅜањᛵ㢲ਾˈagent Պᴤᯠ䗮ࡠⴞḷ䐟ᖴк⇿њ䖜ᦒⲴ Q̂ ՠ䇑Ǆ↔䇝
㓳䗷〻ᱮ❦Պ൘ᴤቁⲴᗚ⧟⅑ᮠ޵᭦ᮋˈ㲭❦ᆳ㾱≲ agent ൘ᔰ࿻↔ᛵ㢲䇝㓳ࡽ֯⭘ᴤཊⲴ޵
ᆈᶕᆈۘᮤњᛵ㢲Ǆ
᭩䘋᭦ᮋ䙏⦷ⲴㅜҼњㆆ⮕ᱟᆈۘ䗷৫Ⲵ⣦ᘱ-ࣘ֌䖜ᦒˈԕ৺⴨ᓄ᭦ࡠⲴ・ণഎᣕˈ❦
ਾઘᵏᙗൠ൘ަк䟽ᯠ䇝㓳Ǆᔰ࿻ਟ㜭Պ䇔Ѫ⭘⴨਼Ⲵ䖜ᦒ䟽ᯠ䇝㓳ᱟ‫ڊ‬ᰐ⭘࣏Ǆն⌘᜿ࡠᴤ
ᯠⲴ Qˆ (s,a)٬ᱟ⭡ਾ㔗⣦ᘱ s´= ¥ (s,a)Ⲵ Qˆ (s´,a)٬⺞ᇊⲴǄഐ↔ˈྲ᷌ਾ㔝Ⲵ䇝㓳᭩ਈҶ
Q̂ (s,a)٬ަѝањˈ൘䖜ᦒ<s,a>к䟽䇝㓳Պᗇࡠ Q̂ (s,a)Ⲵн਼٬Ǆа㡜ൠˈᡁԜᐼᵋ䟽᭮ᰗ
Ⲵ䖜ᦒ⴨∄ҾӾ⧟ຳѝ㧧ᗇᯠ䖜ᦒⲴ〻ᓖਆߣҾ䘉є⿽᫽֌൘⢩ᇊ䰞仈亶ฏѝ⴨ሩᔰ䬰Ǆֻྲ
൘ᵪಘӪሬ㡚ࣘ֌Ⲵ亶ฏˈަࣘ֌ᢗ㹼䴰㾱ᮠ。Ⲵᰦ䰤ˈӾཆ䜘ц⭼᭦䳶ᯠⲴ⣦ᘱ-ࣘ֌䖜ᦒ
Ⲵᔦ䘏Պ∄൘޵䜘䟽᭮ԕࡽ㿲ሏ䗷Ⲵ䖜ᦒⲴᔰ䬰㾱བྷ㤕ᒢᮠ䟿㓗Ǆ⭡Ҿ Q ᆖҐ䙊ᑨ㾱≲ᡀॳ
кзⲴ䇝㓳ᗚ⧟᡽᭦ᮋˈ䘉⿽ᐞ࡛ᱮᗇॱ࠶䟽㾱ⲴǄ
⌘᜿䍟ク൘к䘠䇘䇪ѝⲴєњ‫ٷ‬ᇊᱟˈagent н⸕䚃⧟ຳ⭘ᶕ⭏ᡀਾ㔗⣦ᘱ s´Ⲵ⣦ᘱ䖜ᦒ

࠭ᮠ¥(s,a)ˈҏн⸕䚃⭏ᡀഎᣕⲴ࠭ᮠ r(s,a)Ǆྲ᷌ᆳ⸕䚃Ҷ䘉єњ࠭ᮠˈቡਟ㜭ᴹᴤཊᴹ᭸
Ⲵᯩ⌅Ǆֻྲˈྲ᷌ᢗ㹼ཆ䜘ࣘ֌ᔰ䬰ᖸབྷˈ agent ਟԕㆰঅൠᘭ⮕⧟ຳˈ൘ަ޵䜘⁑ᤏ⧟
ຳˈᴹ᭸⭏ᡀ⁑ᤏࣘ֌ᒦ䍻Ҹ䘲ᖃⲴ⁑ᤏഎᣕˈSutton˄1991˅᧿䘠Ҷ Dyna փ㌫㔃ᶴˈᆳ൘
ཆ䜘ц⭼ѝᢗ㹼Ⲵ⇿↕ࣘ֌ਾᢗ㹼аᇊᮠ䟿Ⲵ⁑ᤏࣘ֌ǄMoore & Atkeson˄1993˅᧿䘠Ҷа
⿽〠ѪՈ‫ݸ‬㓗ᢛ䲔˄prioritized sweeping˅Ⲵ䙄ᖴˈ䘹ᤙᴰਟ㜭Ⲵ⣦ᘱᶕᴤᯠлањˈ⵰䟽Ҿ
ᖃࡽ⣦ᘱᴹ䖳བྷᴤᯠᰦⲴࡽ傡⣦ᘱǄPeng & Williams˄1994˅᧿䘠Ҷањ⴨լⲴ䙄ᖴǄӾࣘᘱ
㿴ࡂ亶ฏᶕⲴབྷ䟿ᴹ᭸㇇⌅ਟ㻛ᓄ⭘Ҿ࠭ᮠ¥઼ r ᵚ⸕ⲴᛵߥǄKaelbling et al.˄1996˅䈳ḕҶ
ަѝⲴࠐ⿽㇇⌅Ǆ
13.4 䶔⺤ᇐᙝഔᣛૂࣞ֒
к䶒ᡁԜ㘳㲁Ҷ⺞ᇊᙗ⧟ຳлⲴ Q ᆖҐǄ䘉䟼ᡁԜ㘳㲁䶎⺞ᇊᙗᛵߥˈަѝഎᣕ࠭ᮠ
r(s,a)઼ࣘ֌䖜ᦒ࠭ᮠ¥(s,a)ਟ㜭ᴹᾲ⦷Ⲵ䗃ࠪǄֻྲˈ൘ Tesauro˄1995˅Ⲵ㾯⌻ৼ䱶ỻሩᔸ
〻ᒿѝˈ䗃ࠪⲴࣘ֌ᱟާᴹപᴹⲴᾲ⦷ᙗⲴˈഐѪ⇿⅑〫ࣘ䴰㾱᧧僠ᆀߣᇊǄ㊫լⲴˈ൘ᴹಚ
༠ⲴՐᝏಘ઼᭸ᓄಘⲴᵪಘӪѝˈሶࣘ֌઼എᣕᔪ᪨Ѫ䶎⺞ᇊᙗ䗷〻䖳Ѫਸ䘲Ǆ൘䘉ṧⲴᛵߥ
лˈ࠭ᮠ¥(s,a)઼ r(s,a)ਟ㻛ⴻ֌ᱟ俆‫ݸ‬สҾ s ઼ a ӗ⭏䗃ࠪⲴᾲ⦷࠶ᐳˈ❦ਾ᤹↔࠶ᐳᣭਆ䲿
ᵪⲴ䗃ࠪǄᖃ䘉Ӌᾲ⦷࠶ᐳѫ㾱‫׍‬䎆Ҿ s ઼ a ᰦ˄ֻྲˈᆳԜн‫׍‬䎆ԕࡽⲴ⣦ᘱ઼ࣘ֌˅ˈᡁ
Ԝਟ〠䘉њ㌫㔏Ѫ䶎⺞ᇊᙗ傜ቄਟཛߣㆆ䗷〻Ǆ
ᵜ㢲ѝᡁԜᢺ༴⨶⺞ᇊ䰞仈Ⲵ Q ᆖҐ㇇⌅ᢙኅࡠ䶎⺞ᇊᙗⲴ MDPǄѪ䗮ࡠ䘉њⴞⲴˈᡁ

Ԝഎ亮൘⺞ᇊᙗᛵߥлⲴ㇇⌅᧘ሬ↕僔ˈ൘䴰㾱ᰦሩަ‫↓؞ࠪڊ‬Ǆ
൘䶎⺞ᇊᙗᛵߥлˈᡁԜᗵ享‫ݸ‬䟽ᯠਉ䘠ᆖҐಘⲴⴞḷˈԕ㘳㲁ࣘ֌Ⲵ䗃ࠪн޽ᱟ⺞ᇊᙗ
±
ⲴᛵߥǄᖸ᰾ᱮˈа⿽а㡜ॆⲴᯩ⌅ᱟᢺањㆆ⮕±Ⲵ٬V 䟽ᇊѹѪᓄ⭘↔ㆆ⮕ᰦ᭦ࡠᣈ㇇㍟
〟എᣕⲴᵏᵋ٬˄൘䘉Ӌ䶎⺞ᇊᙗ䗃ࠪк˅Ǆ
ªf º
V S ( st ) { E «¦ J i rt i »
¬i 0 ¼
ྲԕࡽ䛓ṧˈަѝഎᣕᒿࡇrt+iᱟӾ⣦ᘱsᔰ࿻䚥ᗚㆆ⮕±⭏ᡀǄ⌘᜿↔ᔿᱟᔿ 13.1 Ⲵа㡜

ॆᖒᔿˈਾ㘵㾶ⴆҶ⺞ᇊᙗⲴᛵᖒǄ
±
ྲԕࡽ䛓ṧˈᡁԜᇊѹᴰՈㆆ⮕±*Ѫᡰᴹ⣦ᘱsѝ֯V (s)ᴰབྷॆⲴㆆ⮕±Ǆла↕ᡁԜᢺ
‫ࡽݸ‬ᔿ 13.4 ѝሩQⲴᇊѹа㡜ॆˈ޽а⅑䘀⭘ަᵏᵋ٬Ǆ
Q( s, a) { E[r ( s, a) JV * (G ( s, a))]
E[r ( s, a)] JE[V * (G ( s, a ))]
E[r ( s, a)] J ¦ P ( s c | s, a)V * ( s c) ˄13.8˅

sc
ަѝP(s´|s,a)Ѫ൘⣦ᘱs䟷ਆࣘ֌aՊӗ⭏лањ⣦ᘱѪs´Ⲵᾲ⦷Ǆ⌘᜿ᡁԜ൘䘉䟼ᐢ㓿֯⭘
ҶP(s´|s,a)ᶕ᭩߉V*(¥(s,a))Ⲵᵏᵋ٬ˈᖒᔿѪоᾲ⦷ᙗⲴ¥Ⲵਟ㜭䗃ࠪ⴨‫ޣ‬㚄Ⲵᾲ⦷Ǆ
ྲԕࡽˈਟሶ Q 䟽ᯠ㺘䗮Ѫ䙂ᖂⲴᖒᔿ˖
Q ( s, a ) E[r ( s, a)] J ¦ P( s c | s, a) max Q ( s c, a c) ˄13.9˅

a'
sc
ᆳᱟࡽ䶒ᔿ 13.6 Ⲵа㡜ॆᖒᔿǄᾲᤜൠ䈤ˈᡁԜᢺ䶎⺞ᇊᙗᛵߥлⲴ Q(s,a)ㆰঅൠ䟽ᇊѹ

Ѫ⺞ᇊᙗᛵߥлᇊѹⲴ䟿Ⲵᵏᵋ٬Ǆ
ᡁԜᐢ㓿ᢺQⲴᇊѹа㡜ॆԕ䘲ᓄ䶎⺞ᇊᙗ⧟ຳлⲴ࠭ᮠr઼¥ˈ⧠൘ᡰ䴰㾱Ⲵᱟањᯠ䇝
㓳⌅ࡉǄࡽ䶒ሩ⺞ᇊᙗᛵᖒ᧘ሬⲴ䇝㓳⌅ࡉ˄ᔿ 13.7˅н㜭ཏ൘䶎⺞ᇊᙗᶑԦл᭦ᮋǄֻྲˈ
㘳㲁ањ䶎⺞ᇊᙗഎᣕ࠭ᮠr(s,a)ˈ⇿⅑䟽༽<s,a>䖜ᦒᰦӗ⭏н਼ⲴഎᣕǄ䘉ṧˈণ֯ Q̂ 㺘٬
㻛ࡍ࿻ॆѪ↓⺞ⲴQ࠭ᮠˈ䇝㓳㿴ࡉӽՊнᯝⲴ᭩ਈ Q̂ (s,a)Ⲵ٬Ǆㆰ㾱Ⲵ䈤ˈ↔䇝㓳㿴ࡉн᭦
ᮋǄ↔䳮仈Ⲵ䀓ߣਟ䙊䗷‫؞‬᭩䇝㓳㿴ࡉˈԔަ֯⭘ᖃࡽ Qˆ ٬઼‫Ⲵ↓؞‬ՠ䇑Ⲵањ㺠߿Ⲵ࣐ᵳᒣ
൷Ǆ⭘ Qˆ nᶕԓ㺘ㅜn⅑ᗚ⧟ѝagentⲴՠ䇑ˈл䶒‫؞‬᭩ਾⲴ䇝㓳㿴ࡉ䏣ԕ‫؍‬䇱 Qˆ ᭦ᮋࡠQǄ
Qˆ n ( s, a) m (1 D n )Qˆ n 1 ( s, a ) D n [r J max Qˆ n 1 ( s c, a c)] ˄13.10˅

ac
ަѝ
1
Dn ˄13.11˅
1 visits n ( s, a)
ަѝs઼aѪㅜn⅑ᗚ⧟ѝᴤᯠⲴ⣦ᘱ઼ࣘ֌ˈ㘼фvisitsn(s,a)Ѫ↔⣦ᘱ-ࣘ֌ሩ൘䘉n⅑ᗚ⧟
޵˄वᤜㅜn⅑ᗚ⧟˅㻛䇯䰞Ⲵᙫ⅑ᮠǄ
൘↔‫↓؞‬ҶⲴ㿴ࡉѝ‫ޣ‬䭞ᙍᜣ൘Ҿሩ Q̂ Ⲵᴤᯠ∄⺞ᇊᙗᛵߥлᴤѪᒣ㕃Ǆ⌘᜿ˈྲ᷌൘ᔿ
13.10 ѝᢺ¢n䇮㖞Ѫ 1ˈਟᗇࡠ⺞ᇊᙗᛵᖒлⲴ䇝㓳㿴ࡉǄ֯⭘䖳ሿⲴ¢٬ˈ䈕亩ਟԕ㻛ᖃࡽ
Ⲵ Q̂ (s,a)൷ॆԕӗ⭏ᯠⲴᴤᯠ٬Ǆ൘ᔿ 13.11 ѝ¢nⲴ٬䲿nⲴ໎䮯㘼߿ሿˈഐ↔ᖃ䇝㓳䘋㹼ᰦ
ᴤᯠ〻ᓖ䙀⑀ਈሿǄ൘䇝㓳ѝԕаᇊ䙏⦷߿ሿ¢ˈਟԕ䗮ࡠ᭦ᮋࡠ↓⺞Q࠭ᮠⲴⴞⲴǄк䶒㔉
ࠪⲴ¢nⲴ䘹ᤙᱟ┑䏣᭦ᮋᙗᶑԦⲴ䘹ᤙѻаˈᆳ᤹➗л䶒Ⲵᇊ⨶˄㿱Watkn & Danyan 1992˅
ᇊ⨶ 13.2 ሩ䶎⺞ᇊᙗ傜ቄਟཛߣㆆ䗷〻ⲴQᆖҐ᭦ᮋᙗǄ㘳㲁ањQᆖҐagent൘а䶎⺞ᇊᙗMDP
ѝˈᒦфᴹᴹ⭼Ⲵഎᣕ (s, a ) | r ( s, a ) |d c Ǆ↔QᆖҐagent֯⭘ᔿ 13.10 Ⲵ䇝㓳㿴ࡉˈࡍ࿻ॆ㺘 Q̂ (s,a)
Ѫԫ᜿ᴹ䲀٬ˈᒦф֯⭘ᣈ㇇ഐᆀ 0İ¤<1ˈԔn(i,s,a)Ѫሩᓄࣘ֌aㅜi⅑ᓄ⭘Ҿ⣦ᘱsⲴ䘝ԓǄྲ᷌⇿њ
⣦ᘱ-ࣘ֌ሩ㻛ᰐ䲀仁㑱䇯䰞ˈ0İ¢n<1ˈᒦф
f f
¦ D n (i ,s ,a )
i 1
f, ¦ [D
i 1
n (i , s ,a ) ]2 f
䛓Ѹሩᡰᴹs઼aˈᖃnėĞᰦˈ Q̂ n(s,a) ėQ(s,a)ˈᾲ⦷Ѫ 1Ǆ
㲭❦ Q ᆖҐ઼ᴹ‫Ⲵޣ‬໎ᕪ㇇⌅ਟ㻛䇱᰾൘аᇊᶑԦл᭦ᮋˈ൘֯⭘ Q ᆖҐⲴᇎ䱵㌫㔏
ѝˈ䙊ᑨ䴰㾱ᮠԕॳ䇑Ⲵ䇝㓳ᗚ⧟ԕ᭦ᮋǄֻྲˈTesauro Ⲵ㾯⌻ৼ䱶ỻሩᔸ֯⭘ 150 зњሩ
ᔸỻተ䘋㹼䇝㓳ˈ⇿⅑वᤜᮠॱњ⣦ᘱ-ࣘ֌䖜ᦒǄ
13.5 ᰬ䰪ᐤࡡᆜҖ(temporal difference learning ⷢᰬᐤ࠼ᆜҖ)

Q ᆖҐ㇇⌅ⲴᆖҐ䗷〻ᱟᗚ⧟ൠ߿ሿሩ⴨䛫⣦ᘱⲴ Q ٬Ⲵՠ䇑ѻ䰤ⲴᐞᔲǄ൘䘉њ᜿ѹ
кˈQ ᆖҐᱟᴤᒯ⌋Ⲵᰦ䰤ᐞ࡛˄temporal difference˅㇇⌅ѝⲴ⢩ֻǄᰦ䰤ᐞ࡛ᆖҐ㇇⌅ᆖҐ
䗷〻ᱟ߿ሿ agent ൘н਼Ⲵᰦ䰤‫ࠪڊ‬ՠ䇑ѻ䰤ⲴᐞᔲǄഐѪᔿ 13.10 Ⲵ㿴ࡉ߿ሿҶሩḀ⣦ᘱⲴ
Q̂ ٬ՠ䇑ԕ৺ަ・ণਾ㔗Ⲵ Q̂ ՠ䇑ѻ䰤ⲴᐞˈᡁԜҏਟԕ䇮䇑㇇⌅ᶕ߿ሿ↔⣦ᘱоᴤ䘌Ⲵਾ
㔗ᡆࡽ䎻⣦ᘱѻ䰤ⲴᐞᔲǄ
Ѫ䘋а↕᧒䇘䘉а䰞仈ˈഎᗶалQᆖҐˈᆳⲴ䇝㓳㿴ࡉ䇑㇇ࠪⲴ Q̂ (st,at)Ⲵ䇝㓳٬ᱟԕ
Q̂ (st+1,at+1)㺘⽪Ⲵˈަѝst+1ᱟᓄ⭘ࣘ֌atࡠ⣦ᘱstⲴ㔃᷌ǄԔQ(1)(st,at)Ѫ↔অ↕ࡽⷫ䇑㇇Ⲵ䇝㓳
٬˖
Q (1) ( st , at ) { rt J max Qˆ ( st 1 , a)
a
䇑㇇Q(st,at)䇝㓳٬Ⲵਖа⿽ᯩ⌅ᱟสҾє↕Ⲵ㿲ሏࡠⲴഎᣕ˖
Q ( 2 ) ( st , at ) { rt Jrt 1 J 2 max Qˆ ( st 2 , a)
a
ԕ৺൘а㡜Ⲵᛵߥл n ↕Ⲵഎᣕ˖
Q ( n ) ( st , at ) { rt Jrt 1 J ( n 1) rt n 1 J n max Qˆ ( st n , a)

a
Sutton ӻ㓽Ҷ␧ਸ䘉Ӌн਼䇝㓳ՠ䇑Ⲵа㡜ᯩ⌅ˈ〠Ѫ TD(¬)Ǆ䘉аᜣ⌅ᱟ֯⭘ᑨ䟿 0İ

¬İ1 ᶕਸᒦӾн਼ࡽⷫ䐍⿫ѝ㧧ᗇⲴՠ䇑ˈ㿱лᔿ˖
> @
Q O ( st , at ) { (1 O ) Q (1) ( st , at ) OQ ( 2) ( st , at ) O2 Q (3) ( st , at )
¬
Q ⲴањㅹԧⲴ䙂ᖂᇊѹѪ˖
Q O ( st , at ) rt J [(1 O ) max Qˆ ( s t , at ) OQ O ( st 1 , at 1 )]
a
⌘᜿ྲ᷌ᡁԜ䘹ᤙ¬=0ˈࡉᗇࡠ৏ᶕⲴ䇝㓳ՠ䇑Q(1)ˈᆳਚ㘳㲁 Q̂ ՠ䇑ѝⲴঅ↕ᐞᔲǄᖃ
¬໎བྷᰦˈ↔㇇⌅䟽⛩䙀⑀䖜〫ࡠᴤ䘌Ⲵࡽⷫ↕ѝǄ൘ᶱㄟᛵߥ¬=1 ᰦˈਚ㘳㲁㿲ሏࡠⲴrt+i
¬
٬ˈᖃࡽⲴ Q̂ ՠ䇑ሩަ⋑ᴹ䍑⥞Ǆ⌘᜿ᖃ Q̂ =Qᰦˈ⭡Q 㔉ࠪⲴ䇝㓳٬ሩҾ 0İ¬İ1 Ⲵᡰᴹ
¬٬䜭⴨਼Ǆ
TD(¬)Ⲵࣘᵪᱟˈ൘ḀӋᶑԦлˈྲ᷌㘳㲁ᴤ䘌Ⲵࡽⷫˈ䇝㓳Պᴤᴹ᭸Ǆֻྲˈᖃagent
¬
䚥ᗚᴰՈㆆ⮕䘹ᤙࣘ֌ᰦˈ¬=1 ⲴQ ሶᨀ‫׋‬ሩⵏᇎQ٬Ⲵᆼ㖾ՠ䇑ˈн䇪 Q̂ ᴹཊѸн㋮⺞Ǆ
ਖаᯩ䶒ˈྲ᷌ࣘ֌ᒿࡇⲴ䘹ᤙᱟ⅑ՈⲴˈ䛓ѸሩᵚᶕⲴ㿲ሏrt+iਟ㜭ᴹ䈟ሬᙗǄ
¬
Peng & Williams˄1994˅ᨀ‫׋‬Ҷ䘋а↕Ⲵ䇘䇪઼ᇎ傼㔃᷌ˈᱮ⽪ҶQ ൘ањ䰞仈亶ฏкⲴ
ঃ䎺ᙗ㜭ǄDayan(1992)ᱮ⽪Ҷ൘аᇊᶑԦл㿱㊫լⲴTD(¬)ᯩ⌅ᓄ⭘ࡠᆖҐV*࠭ᮠѝˈሩҾ 0
İ¬İ1 Ⲵԫ᜿¬٬䜭ਟ↓⺞᭦ᮋǄTesauro(1995) ൘ަTD-Gammon〻ᒿ㾯⌻ৼ䱶ỻሩᔸѝ֯
⭘ҶTD(¬)ᯩ⌅Ǆ
13.6 Ԅṭׁѣ⌑ौ
㠣↔ˈ൘ Q ᆖҐѝਟ㜭ᴰާᴹ㓖ᶏᙗⲴ‫ٷ‬ᇊᱟަⴞḷ࠭ᮠ㻛㺘⽪ѪањᱮᔿⲴḕ᢮㺘ˈ
ሩ⇿њн਼䗃‫ޕ‬٬˄ণ⣦ᘱ-ࣘ֌ሩ˅ᴹањ㺘亩Ǆഐ↔ᡁԜⲴ䇘䇪Ⲵ㇇⌅ᢗ㹼а⿽ᵪỠⲴᆖ
Ґᯩ⌅ ˈᒦфнՊቍ䈅䙊䗷ӾᐢⴻࡠⲴ⣦ᘱ-ࣘ֌ሩѝ⌋ॆԕՠ䇑ᵚⴻࡠⲴ⣦ᘱ-ࣘ֌ሩⲴ Q
٬Ǆ䘉њᵪỠᆖҐ‫ٷ‬ᇊ൘᭦ᮋᙗ䇱᰾ѝ৽᱐ࠪᶕˈᆳ䇱᰾Ҷਚᴹ⇿њਟ㜭Ⲵ⣦ᘱ-ࣘ֌㻛ᰐ䲀
仁㑱Ⲵ䇯䰞ˈᆖҐ䗷〻᡽Պ᭦ᮋǄ൘བྷⲴᡆᰐ䲀Ⲵオ䰤ѝˈᡆ㘵ᢗ㹼ࣘ֌Ⲵᔰ䬰ᖸབྷᰦˈ䘉ᱮ
❦ᱟн࠷ᇎ䱵Ⲵ‫ٷ‬ᇊǄ֌Ѫ㔃᷌ˈᴤᇎ䱵Ⲵ㌫㔏䙊ᑨਸᒦҶަԆㄐ䇘䇪Ⲵ࠭ᮠ䙬䘁ᯩ⌅ԕ৺䘉
䟼䇘䇪Ⲵ Q ᆖҐ䇝㓳㿴ࡉǄ
ᖸᇩ᱃ᢺ৽ੁՐ᫝䘉ṧⲴ࠭ᮠ䙬䘁㇇⌅㔃ਸࡠ Q ᆖҐ㇇⌅ѝˈ䙊䗷⭘⾎㓿㖁㔌ᴯԓḕ᢮
㺘ˈᒦфᢺ⇿њ Q̂ (s,a)ᴤᯠ֌Ѫ䇝㓳ṧֻǄֻྲˈᡁԜਟᢺ⣦ᘱ s ઼ࣘ֌ a 㕆⸱Ѫ㖁㔌䗃‫ˈޕ‬
ᒦф䇝㓳㖁㔌ԕ䗃ࠪ Q̂ Ⲵⴞḷ٬ˈ൘㔉ᇊᔿ 13.7 ઼ᔿ 13.10 Ⲵ䇝㓳㿴ࡉⲴᶑԦлˈਖа⿽ᴹᰦ
൘ᇎ䐥ѝᴤᡀ࣏Ⲵᯩ⌅ᱟሩ⇿њࣘ֌䇝㓳ањঅ⤜Ⲵ㖁㔌ˈ֯⭘⣦ᘱ֌Ѫ䗃‫ ˈޕ‬Q̂ Ѫ䗃ࠪǄ䘈
ᴹа⿽䙊ᑨ֯⭘Ⲵᯩ⌅ᱟ䇝㓳ањ㖁㔌ˈᆳԕ⣦ᘱ֌Ѫ䗃‫ˈޕ‬նሩ⇿њࣘ֌䗃ࠪањ Q̂ ٬Ǆഎ
ᗶㅜ 1 ㄐѝᡁԜ䇘䇪Ҷ൘ỻⴈ⣦ᘱк֯⭘㓯ᖒ࠭ᮠ઼ LMS ㇇⌅ᶕ䙬䘁ՠ䇑࠭ᮠǄ
൘ᇎ䐥ѝˈᐢᔰਁࠪҶ䇨ཊᡀ࣏Ⲵ໎ᕪᆖҐ㌫㔏ˈᆳԜ䙊䗷㔃ਸ䘉ṧⲴ࠭ᮠ䙬䘁㇇⌅ᶕԓ
ᴯḕ᢮㺘ǄTesauro Ⲵᡀ࣏Ⲵ TD-Gammon 〻ᒿ֯⭘Ҷ⾎㓿㖁㔌઼৽ੁՐ᫝㇇⌅ˈо TD(¬)䇝
㓳㿴ࡉ⴨㔃ਸǄZhang & Dietterich(1996)֯⭘⴨լⲴ৽ੁՐ᫝о TD(¬)Ⲵ㔃ਸ⭘Ҿ job-shop 䈳
ᓖԫ࣑Ǆ Crites & Barto˄1996˅᧿䘠Ҷањ⾎㓿㖁㔌໎ᕪᆖҐᯩ⌅ˈ⭘Ҿ⭥ở䈳ᓖԫ࣑Ǆ
Thrun(1996)ᣕ੺ҶањสҾ⾎㓿㖁㔌Ⲵ Q ᆖҐˈᆳਟᆖҐᑖᴹ༠㓣઼ᩴ‫ۿ‬ཤՐᝏಘⲴ〫ࣘᵪ
ಘӪⲴสᵜ᧗ࡦ䗷〻ǄMahadevan & Connell(1991)᧿䘠ҶањสҾ㚊㊫⣦ᘱⲴ Q ᆖҐᯩ⌅ˈᓄ
⭘ҾㆰঅⲴ〫ࣘᵪಘӪ᧗ࡦ䰞仈Ǆ
㲭❦䘉Ӌ㌫㔏㧧ᗇҶᡀ࣏ˈሩҾަԆⲴԫ࣑ˈаնᕅ‫ޕ‬Ҷ⌋ॆ࠭ᮠ䙬䘁ಘˈ໎ᕪᆖҐሶн
㜭᭦ᮋǄ䘉ṧⲴᴹ䰞仈Ⲵԫ࣑⭡Boyan & Moore(1995)ˈBaird(1995)઼Gordon(1995)ӻ㓽Ǆ⌘᜿
ᵜㄐࡽ䶒䇘䇪Ⲵ᭦ᮋᙗᇊ⨶ਚᓄ⭘Ҿ Qˆ 㺘⽪Ѫ᰾⺞Ⲵ㺘ᖒᔿᰦˈѪҶⴻࡠഠ䳮ᡰ൘ˈ㘳㲁֯⭘
ањ⾎㓿㖁㔌㘼нᱟ᰾⺞Ⲵ㺘ᶕ㺘⽪ Q̂ Ǆྲ᷌ᆖҐಘᴤᯠ㖁㔌ԕᴤྭൠ३䝽⢩ᇊ䖜ᦒⲴ<si,ai>
Ⲵ䇝㓳Q٬ˈਈॆҶⲴ㖁㔌ᵳ٬ҏՊ‫؞‬᭩ަԆⲴԫ᜿䖜ᦒⲴ Q̂ ՠ䇑ǄഐѪ䘉Ӌᵳ٬ਈॆՊ໎࣐
ަԆ䖜ᦒⲴ Q̂ ՠ䇑Ⲵ䈟ᐞˈ৏ᶕᇊ⨶ѝⲴ䇱᰾↕僔н޽ᡀ・Ǆ‫ޣ‬Ҿᑖᴹ⌋ॆ࠭ᮠ䙬䘁ಘⲴ໎ᕪ
ᆖҐⲴ⨶䇪࠶᷀⭡Gordon(1995)઼Tsitsiklis(1994)֌ࠪǄBaird(1995)ᨀࠪҶสҾởᓖⲴᯩ⌅ˈᆳ
䙊䗷ⴤ᧕ᴰሿॆሩ⴨䛫⣦ᘱⲴՠ䇑ѝⲴᐞᔲᒣᯩ઼ᶕ䀓ߣ䘉а䳮仈˄ҏ㻛〠ѪBellman↻⮉䈟
ᐞBellman residual error˅Ǆ
13.7 фࣞᘷ㿺ࡈⲺ㚊㌱
‫ ۿ‬Q ᆖҐ䘉ṧⲴ໎ᕪᆖҐᯩ⌅ˈо䮯ᵏ⹄ウⲴ⭘Ҿ䀓ߣ傜ቄਟཛߣㆆ䗷〻Ⲵࣘᘱ㿴ࡂᯩ
⌅ᴹ⵰㍗ᇶⲴ㚄㌫Ǆ䘉ӋᰙᵏⲴᐕ֌䙊䗷‫ٷ‬ᇊ agent ᤕᴹᆳᡰ༴⧟ຳⲴ࠭ᮠ¥(s,a)઼ r(s,a)Ⲵᆼ
㖾⸕䇶Ǆഐ↔ˈᆳѫ㾱䀓ߣⲴ䰞仈ᱟ⭘ᴰሿⲴ䇑㇇䟿ᗇࡠᴰՈㆆ⮕ˈ‫ٷ‬ᇊ⧟ຳਟ㻛ᆼ㖾ൠ⁑
ᤏˈн䴰㾱ⴤ᧕ⲴӔӂǄQ ᆖҐⲴ፝ᯠѻ༴൘Ҿᆳ‫ٷ‬ᇊнާᴹ¥(s,a)઼ r(s,a)Ⲵ⸕䇶ˈᆳн㜭
൘޵䜘⁑ᤏⲴ⣦ᘱオ䰤ѝ〫ࣘˈ㘼ᗵ享൘⧠ᇎц⭼ѝ〫ࣘᒦ㿲ሏਾ᷌Ǆ൘ਾа⿽ᛵߥлᡁԜѫ
㾱㘳㲁Ⲵᱟ agent Ѫ᭦ᮋࡠањਟ᧕ਇⲴㆆ⮕ᗵ享ᢗ㹼Ⲵⵏᇎц⭼ࣘ֌ᮠ䟿ˈ㘼нᱟ享㣡䍩Ⲵ
䇑㇇䘝ԓ⅑ᮠǄ৏ഐᱟ൘䇨ཊᇎ䱵Ⲵ亶ฏѝˈ∄ྲ⭏ӗ䰞仈ˈ൘ཆ䜘ц⭼ѝᢗ㹼ࣘ֌Ⲵᰦ䰤઼
䍩⭘ᔰ䬰∄䇑㇇ᔰ䬰ᴤ٬ᗇ‫⌘ޣ‬Ǆ൘ⵏᇎ⧟ຳѝ〫ࣘ䘋㹼ᆖҐˈᒦф㿲ሏަ㔃᷌Ⲵ㌫㔏䙊ᑨ〠
Ѫ൘㓯(online)㌫㔏ˈ㘼ѫ㾱䙊䗷⁑ර⁑ᤏࣘ֌ⲴᆖҐ㻛〠Ѫ⿫㓯˄offline˅㌫㔏Ǆ
䙊䗷㘳㲁 Bellman ㅹᔿˈਟԕ␵ᾊൠⴻࡠᰙᵏⲴᯩ⌅઼䘉䟼䇘䇪Ⲵ໎ᕪᆖҐ䰞仈ѻ䰤Ⲵᇶ

࠷⴨‫ޣ‬ᙗǄBellman ㅹᔿᖒᡀҶ䀓ߣ MDP Ⲵࣘᘱ㿴ࡂᯩ⌅Ⲵส⹰ˈަᖒᔿྲл˖
(s S )V * ( s ) E[r ( s, S ( s)) JV * (G ( s, S ( s)))]
䈧⌘᜿Bellmanㅹᔿ઼ࡽ䶒ᔿ 13.2 ѝᇊѹⲴᴰՈㆆ⮕ѻ䰤䶎ᑨ㍗ᇶⲴ㚄㌫ǄBellman(1957)

䇱᰾ҶᴰՈㆆ⮕±*┑䏣к䘠ㅹᔿˈф┑䏣↔ㅹᔿⲴㆆ⮕±ѪањᴰՈㆆ⮕Ǆࣘᘱ㿴ࡂᯩ䶒Ⲵ
ᰙᵏᐕ֌वᤜBellman-Fordᴰ⸝䐟ᖴ㇇⌅˄Bellman 1958; Ford & Fulkerson 1962˅Ǆᆳ䙊䗷н
ᯝᴤᯠ⇿њമ㔃⛩ࡠ㓸⛩Ⲵՠ䇑䐍⿫ˈᶕᆖҐമѝⲴ䐟ᖴˈสҾ㔃⛩䛫ትⲴ䐍⿫Ǆ൘↔㇇⌅ѝ
മⲴ਴䗩ԕ৺ⴞḷ㔃⛩ᐢ⸕Ⲵ‫ٷ‬ᇊˈㅹԧҾ¥(s,a)઼r(s,a)ᐢ⸕Ⲵ‫ٷ‬ᇊǄBarto et al.˄1995˅䇘
䇪Ҷ໎ᕪᆖҐ઼ࣘᘱ㿴ࡂⲴ㍗ᇶ㚄㌫Ǆ
x ໎ᕪᆖҐ䀓ߣ㠚⋫ agent ᆖҐ᧗ࡦㆆ⮕Ⲵ䰞仈Ǆᆳ‫ٷ‬ᇊ䇝㓳ؑ᚟ⲴᖒᔿѪሩ⇿њ⣦
ᘱ-ࣘ֌䖜ᦒ㔉ࠪⲴᇎ٬എᣕؑਧǄagent ⲴⴞḷᱟᆖҐањ㹼ࣘㆆ⮕ˈᆳ֯ agent
Ӿԫ᜿䎧࿻⣦ᘱ᭦ࡠⲴᙫഎᣕѪᴰབྷǄ
x ᵜㄐӻ㓽Ⲵ໎ᕪᆖҐ㇇⌅䘲ਸа㊫㻛〠Ѫ傜ቄਟཛߣㆆ䗷〻Ⲵ䰞仈Ǆ൘傜ቄਟཛ
ߣㆆ䗷〻ѝˈᓄ⭘ԫ᜿ࣘ֌ࡠԫ᜿⣦ᘱкⲴ䗃ࠪਚਆߣҾ↔ࣘ֌઼⣦ᘱ˄оԕࡽ
Ⲵࣘ֌ᡆ⣦ᘱᰐ‫˅ޣ‬Ǆ傜ቄਟཛߣㆆ䗷〻㾶ⴆҶ㤳തᖸᒯⲴ䰞仈ˈवᤜ䇨ཊᵪಘ
Ӫ᧗ࡦˈᐕল㠚ࣘॆ઼䈳ᓖ䰞仈Ǆ
x Q ᆖҐᱟ໎ᕪᆖҐⲴа⿽ᖒᔿǄަѝ agent ᆖҐⲴᱟа㓴⣦ᘱ઼ࣘ֌кⲴՠ䇑࠭
ᮠǄ⺞࠷ൠ䇢ˈՠ䇑࠭ᮠ Q(s,a)㻛ᇊѹѪ agent ᓄ⭘ࣘ֌ a ࡠ⣦ᘱ s кਟ㧧ᗇⲴᴰ
བྷᵏᵋᣈ㇇〟㍟എᣕǄQ ᆖҐⲴՈ⛩ᱟˈণ֯൘ᆖҐಘнާᴹަࣘ֌ᘾṧᖡ૽⧟ຳ
Ⲵ‫ݸ‬傼⸕䇶ᛵߥлˈ↔㇇⌅ӽਟᓄ⭘Ǆ
x ਟԕ䇱᰾ˈ൘䘲ᖃ‫ٷ‬ᇊлˈྲ᷌ᆖҐಘⲴ‫ٷ‬䇮 Q̂ (s,a)㻛㺘⽪Ѫањḕ᢮㺘ˈфሩ⇿
њ<s,a>ሩᴹঅ⤜Ⲵ㺘亩ˈ䛓Ѹ Q ᆖҐਟԕ᭦ᮋࡠ↓⺞Ⲵ Q ࠭ᮠǄ൘⺞ᇊᙗ઼䶎⺞
ᇊᙗⲴ MDP л↔㇇⌅䜭ਟ᭦ᮋǄ൘ᇎ䐥ѝ Q ᆖҐণ֯൘㿴⁑䘲ѝⲴ䰞仈ѝҏ䴰㾱
ᮠॳ⅑Ⲵ䇝㓳ᗚ⧟Ǆ
x Q ᆖҐᱟа㊫ᴤᒯ⌋Ⲵ〠Ѫᰦ䰤ᐞᔲ㇇⌅ѝⲴа⿽Ǆа㡜䈤ᶕˈᰦ䰤ᐞᔲ㇇⌅䙊䗷
нᯝ߿ሿ agent ൘н਼ᰦ䰤޵ӗ⭏Ⲵՠ䇑Ⲵᐞᔲᶕ䘋㹼ᆖҐǄ
x ໎ᕪᆖҐоᓄ⭘Ҿ傜ቄਟཛߣㆆ䗷〻Ⲵࣘᘱ㿴ࡂᴹ㍗ᇶ㚄㌫Ǆަᐞᔲ‫ޣ‬䭞൘Ҿˈ
শਢк䘉Ӌࣘᘱ㿴ࡂᯩ⌅‫ٷ‬ᇊ agent ᤕᴹ⣦ᘱ䖜ᦒ࠭ᮠ¥(s,a)઼എᣕ࠭ᮠ r(s,a)Ⲵ⸕
䇶Ǆ⴨৽ˈQ ᆖҐ䘉ṧⲴ໎ᕪᆖҐ㇇⌅‫ٷ‬ᇊᆖҐಘ㕪ቁ䘉Ӌ⸕䇶Ǆ
൘໎ᕪᆖҐᯩ䶒䇨ཊᐕ֌ѝˈ䙊ᑨⲴѫ仈ᱟ䘝ԓൠ߿ሿⲴਾ㔗⣦ᘱⲴՠ䇑ѻ䰤ⲴᐞᔲǄ֯
⭘䘉⿽ᯩ⌅ⲴḀӋᴰᰙⲴᐕ֌ਟ㿱 Samuel(1959)ˈᆳⲴ㾯⌻ৼ䱶ỻᆖҐ〻ᒿ䈅മ䙊䗷ਾ㔗⣦ᘱ
Ⲵՠ䇑ᶕ⭏ᡀ‫ࡽݸ‬⣦ᘱⲴ䇝㓳٬ˈӾ㘼ᆖࡠ㾯⌻ৼ䱶ỻⲴՠ䇑࠭ᮠǄࠐѾ਼ᰦˈBellman-Ford
ⲴঅⴞⲴᴰ⸝䐟ᖴ㇇⌅㻛ᔰਁࠪᶕ˄Bellman 1958; Ford & Fulkerson 1962˅ˈᆳᢺࡠⴞⲴⲴ䐍
⿫٬Ӿ㔃⛩Ր᫝ࡠᆳⲴ䛫ትǄ൘ᴰՈ᧗ࡦᯩ䶒Ⲵ⹄ウሬ㠤Ҷ֯⭘⴨լᯩ⌅ᶕ䀓ߣ傜ቄਟཛߣㆆ
䗷〻˄Bellman1961; Blackwell 1965˅ǄHolland˄1986˅ⲴᆖҐ࠶㊫㌫㔏Ⲵ㓴Ầᔿ˄bucket
brigade˅ᯩ⌅֯⭘Ҷ㊫լⲴᯩ⌅൘ᔦ䘏എᣕⲴᛵߥлՐ᫝ؑ⭘ǄBarto et al.˄1983˅䇘䇪а⿽
ᰦ䰤ؑ⭘࠶䝽Ⲵᯩ⌅ˈሬ㠤Ҷ Sutton˄1988˅Ⲵ䇪᮷ˈަѝᇊѹҶ TD(¬)ᯩ⌅ᒦ䇱᰾Ҷ൘¬
=0 ᰦᆳⲴ᭦ᮋᙗǄDayan(1992)ᢺ䘉њ㔃᷌ᢙኅࡠ¬Ⲵԫ᜿٬ǄWatkin˄1989˅ӻ㓽Ҷ⭘ Q ᆖ
Ґ൘എᣕ઼ࣘ֌䖜ᦒ࠭ᮠᵚ⸕Ⲵᛵߥл㧧ਆᴰՈㆆ⮕Ǆ൘䘉Ӌᯩ⌅кⲴ᭦ᮋᙗ䇱᰾ᴹࠐњਈ
⿽Ǆ䲔Ҷᵜㄐኅ⽪Ⲵ᭦ᮋᙗ䇱᰾ཆˈਟ㿱˄Baird 1995; Bertsekas 1987; Tsitsiklis 1994, Singh ઼
Sutton 1996˅Ǆ
໎ᕪᆖҐӽᱟањ⍫䏳Ⲵ⹄ウ亶ฏǄֻྲ McCallum(1995)઼ Littman˄1996˅䇘䇪Ҷ໎ᕪ
ᆖҐⲴᢙኅˈԕ䘲ᓄᴹ䳀㯿⣦ᘱਈ䟿⹤ൿ傜ቄਟཛ‫ٷ‬ᇊⲴᛵߥǄ䇨ཊᖃࡽⲴ⹄ウ㠤࣋Ҿᢺ䘉Ӌ
ᯩ⌅ॷ㓗ࡠᴤᓎབྷᴤᇎ䱵Ⲵ䰞仈ѝǄֻྲ Maclin ઼ Shavlik˄1996˅᧿䘠Ҷа⿽ᯩ⌅ˈަѝ໎
ᕪᆖҐ agent ਟ᧕ਇᯭᮉ㘵Ⲵнᆼ㖾ᔪ䇞ˈสҾ KBANN ㇇⌅˄ㅜ 12 ㄐ˅ⲴањᢙኅǄLin
˄1992˅㘳㲁Ҷ䙊䗷ᨀ‫׋‬ᔪ䇞ࣘ֌ᒿࡇᶕᯭᮉⲴ֌⭘ǄSingh˄1993˅઼ Lin˄1993˅ᔪ䇞֯⭘
ቲ⅑ॆⲴࣘ֌ᶕॷ㓗䘉Ӌ㇇⌅ǄDietterich & Flann˄1995˅᧒㍒ҶสҾ䀓䟺Ⲵᯩ⌅઼໎ᕪᆖҐ
Ⲵ䳶ᡀˈMitchell & Thrun˄1993˅᧿䘠Ҷᓄ⭘ EBNN ㇇⌅˄ㅜ 12 ㄐ˅ࡠ Q ᆖҐѝǄRing
˄1994˅㘳㲁Ҷ agent ൘ཊњԫ࣑ѝⲴᤱ㔝ᆖҐǄ
䘁ᵏ‫ޣ‬Ҿ໎ᕪᆖҐⲴ䈳ḕ⭡ Kaelbling et al.˄1996˅; Barto(1992); Barto et al.(1995); Dean et

al.˄1993˅֌ࠪǄ
Ґ仈
13.1 㔉ࠪമ 13-2 ᡰ⽪䰞仈Ⲵਖа⿽ᴰՈㆆ⮕Ǆ
13.2 㘳㲁лമᱮ⽪Ⲵањ⺞ᇊᙗṬᆀц⭼ˈަѝਜ਼ᴹ੨᭦ⴞḷ⣦ᘱ GǄ䘉䟼֌Ҷḷ䇠Ⲵ䖜

ᦒⲴ・ণഎᣕѪ 10ˈ㘼ަԆ䖜ᦒ䜭Ѫ 0Ǆ
(a)㔉ࠪṬᆀц⭼ѝ⇿њ⣦ᘱⲴV* ٬Ǆ㔉ࠪ⇿њ䖜ᦒⲴQ(s,a)٬Ǆᴰਾˈ߉ࠪањᴰՈㆆ
⮕ˈ֯⭘¤=0.8Ǆ
(b)䈅᭩ਈഎᣕ࠭ᮠr(s,a)ˈ֯Q(s,a)ਈॆˈնн᭩ਈᴰՈㆆ⮕Ǆ䈅‫؞‬᭩r(s,a)ˈ֯Q(s,a)ਈ
ॆˈնн᭩ਈV*(s,a)Ǆ
(c)⧠൘㘳㲁ᓄ⭘ Q ᆖҐࡠ↔Ṭᆀц⭼ˈ‫ٷ‬ᇊ Q̂ ٬㺘㻛ࡍ࿻ॆѪ 0Ǆ‫ٷ‬ᇊ agent ᔰ࿻Ҿᐖл

ⲴᯩṬ❦ਾ亪ᰦ䪸⋯⵰ઘ䗩ⲴṬᆀ〫ࣘˈⴤ㠣䗮ࡠ੨᭦ⴞḷ⣦ᘱˈᆼᡀㅜањ䇝㓳ᛵ㢲Ǆ䈅߉
ࠪ↔ᛵ㢲Ⲵ㔃᷌ሬ㠤ଚӋ Qˆ ٬Ⲵ‫؞‬᭩ˈ㔉ࠪ‫↓؞‬ਾⲴ٬Ǆ⧠ྲ᷌ agent ㅜҼ⅑䘀⭘਼ṧⲴᛵ
㢲ˈ޽⅑എㆄ↔䰞仈Ǆ਼ṧ൘ㅜйњᛵ㢲ਾഎㆄ↔䰞仈Ǆ
13.3 㘳㲁оањ䲿ᵪлỻⲴሩᇦሩᔸ Tic-Tac-ToeǄ⺞࠷ൠ䇢ˈ‫ٷ‬ᇊሩᇦ൘ᴹཊњ䘹ᤙᰦԕ

൷रⲴᾲ⦷䘹ᤙ䎠ỻˈ䲔䶎ᴹањᕪࡦᙗⲴ䎠ỻ˄䘉ᰦᆳ䟷ਆᱮ❦↓⺞Ⲵ↕ᆀ˅Ǆ
(a)൘↔ᛵߥлˈሶᆖҐᴰՈⲴ Tic-Tac-Toe ㆆ⮕ᖒᡀањ Q ᆖҐ䰞仈Ǆ൘↔䶎⺞ᇊᙗ傜ቄ

ਟཛߣㆆ䗷〻ѝˈօѪ⣦ᘱǃࣘ֌ԕ৺എᣕ˛
(b)ྲ᷌ሩᇦ䘹ᤙᴰՈⲴ䎠ỻ㘼нᱟ䲿ᵪ䎠ỻˈ֐Ⲵ〻ᒿ㜭੖㜌࡙˛
13.4 ൘䇨ཊMDPѝˈᴹਟ㜭᢮ࡠєњㆆ⮕±1઼±2ˈྲ᷌agentᔰ࿻Ҿ⣦ᘱs1ˈࡉ±1ՈҾ
± 2 ˗ ྲ ᷌ agent ᔰ ࿻ Ҿ ਖ а ⣦ ᘱ s2 ˈ ࡉ ± 2 Ո Ҿ ± 1 Ǆ ᦒ 䀰 ѻ V S 1 ( s1 ) ! V S 2 ( s1 ) ն
V S 2 ( s 2 ) ! V S 1 ( s 2 ) Ǆ䀓䟺ѪӰѸᙫᆈ൘ањㆆ⮕ˈ㜭ሩҾԫ᜿ањࡍ࿻⣦ᘱs֯V±(s)ᴰབྷॆ
˄ ণ ᴰ Ո ㆆ ⮕ ± * ˅ Ǆ ᦒ 䀰 ѻ ˈ 䀓䟺 Ѫ Ӱ Ѹ а њ MDP ᙫ ᴹ а њ ㆆ ⮕ ± * ˈ ֯
(S , s )V S ( s ) t V S ( s ) Ǆ
*

Mechine Learning CH

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Mechine Learning CH

Uploaded by

Copyright:

Available Formats

ᒿ 䀰

ᇐѿφ 对于某类任务 T 和性能度量 P，如果一个计算机程序在 T(任务)上以 P（性

ቭ㇑൘ֻᆀѝᖸ᰾ᱮᓄᢺ ChooseMove ֌Ѫⴞḷ࠭ᮠˈնᡁԜՊਁ⧠ᆖҐ䘉њⴞḷ࠭ᮠ

4. ྲ᷌ b нᱟᴰ㓸ỻተˈ䛓Ѹ V(b)=V(b)ˈަѝ bᱟӾ b ᔰ࿻ৼᯩ䜭䟷ਆᴰՈሩ

Vtrain(b) Vˆ (Successor(b)) ˄1.1˅

New Problem(initial game board)-ᯠ䰞仈˄ࡍ࿻ỻተ˅

Solution trace(game history)-䀓ㆄ䐟㓯˄ሩᔸশਢ˅

䇮䇑㾯⌻䐣ỻ〻ᒿⲴ⍱〻㻛ᖂ㓣൘മ 1-2 ѝǄ䘉њ䇮䇑ᐢ㓿൘ࠐᯩ䶒ᢺᆖҐԫ࣑䲀ࡦ൘

Determine Type of Training Experience- ߣᇊ䇝㓳㓿傼ᖒᔿ

Games against experts- оуᇦሩᔸ

Games against self- о㠚ᐡሩᔸ

Table of correct moves- ↓⺞䎠ᆀⲴ㺘Ṭ

Determine Target Function- ߣᇊⴞḷ࠭ᮠ

Determine Representation of Learned Function-ߣᇊⴞḷ࠭ᮠⲴ㺘⽪

Linear function of six features- ‫ޝ‬њ৲ᮠⲴ㓯ᙗ࠭ᮠ

Artificial neural network- Ӫᐕ⾎㓿㖁㔌

Determine Learning Algorithm- ߣᇊᆖҐ㇇⌅

Gradient descent- ởᓖл䱽

Linear programming- 㓯ᙗ㿴ࡂ

Completed Design- ᆼᡀⲴ䇮䇑

1.4 മ 1-1 ѝᇎ傼⭏ᡀಘ⁑ඇਟ䟷⭘ަԆаӋㆆ⮕Ǆ⺞࠷ൠ䇢ˈ㘳㲁ᇎ傼⭏ᡀಘ⭘л䶒

㺘 2-1 ⴞḷᾲᘥ EnjoySport Ⲵ↓઼ֻ৽ֻ

Example Sky AirTemp Humidity Wind Water Forecast EnjoySport

ྲ᷌ḀӋᇎֻ x ┑䏣‫ٷ‬䇮 h Ⲵᡰᴹ㓖ᶏˈ䛓Ѹ h ሶ x ࠶㊫Ѫ↓ֻˈ˄h(x)=1 ˅Ǆ∄ྲˈ

<?, Cold, High, ?, ?, ?>

㔬кᡰ䘠ˈEnjoySport 䘉њᾲᘥᆖҐԫ࣑䴰㾱ᆖҐⲴᱟ֯ EnjoySport=Yes Ⲵᰕᆀˈᒦሶ

㺘 2-2 EnjoySport ᾲᘥᆖҐԫ࣑

аᰖ㔉ᇊⴞḷᾲᘥ c Ⲵ䇝㓳ṧֻ䳶ˈᆖҐಘ䶒ѤⲴ䰞仈ቡᱟ‫ٷ‬䇮ᡆՠ䇑 cǄ֯⭘ㅖਧ H

ᵪಘᆖҐⲴԫ࣑ᱟ൘ᮤњᇎֻ䳶ਸ X к⺞ᇊоⴞḷᾲᘥ c ⴨਼Ⲵ‫ٷ‬䇮 hˈ❦㘼ᡁԜሩҾ

定义： Ԕhj ઼hk Ѫ൘XкᇊѹⲴᐳቄ࠭ᮠǄᇊѹањmore-general-than-or-equal-to‫ޣ‬

മ 2-1 ᇎֻǃ‫ٷ‬䇮઼ more-general-than ‫ޣ‬㌫

Ѫ䈤᰾䘉Ӌᇊѹˈ㘳㲁EnjoySportֻᆀѝⲴh1ǃh2ǃh3ˈྲമ 2-1 ᡰ⽪Ǆ䘉йњ‫ٷ‬䇮ᱟྲ

ྲօ֯⭘ more-general-than ‫ٿ‬ᒿᶕᩌ㍒о䇝㓳ṧֻ⴨а㠤Ⲵ‫ٷ‬䇮˛а⿽࣎⌅ᱟӾ H ѝ

Ѫ䈤᰾䘉а㇇⌅ˈ‫ٷ‬ᇊ㔉ҸᆖҐಘⲴа㌫ࡇ䇝㓳ṧֻྲ㺘 2-1 ᡰ⽪ǄFind-S Ⲵㅜа↕ᱟ

൘ᢛ᧿ࡠ㺘 2-1 ѝㅜањ䇝㓳ṧֻᰦˈᆳࡊྭᱟњ↓ֻǄᖸ␵ᾊˈ䘉ᰦⲴ h ཚ⢩↺ҶǄ

hĕ<Sunny, Warm, Normal, Strong, Warm, Same>

hĕ<Sunny, Warm, ?, Strong, Warm, Same>

❦ਾ༴⨶ㅜйњ䇝㓳ṧֻˈ䘉䟼ᱟањ৽ֻˈh нਈǄᇎ䱵кˈFind-S ㇇⌅ㆰঅൠᘭ ᘭ⮕

᧕⵰ᆼᡀ Find-S ㇇⌅ˈㅜഋњ↓ֻ֯ᗇ h ᴤа㡜˖

hĕ<Sunny, Warm, ?, Strong, ?, ?>

Find-S ㇇⌅╄⽪Ҷа⿽࡙⭘ more-general-than ‫ٿ‬ᒿᶕᩌ㍒‫ٷ‬䇮オ䰤Ⲵᯩ⌅Ǆ䘉аᩌ㍒⋯

മ 2-2 Find-S ѝⲴ‫ٷ‬䇮オ䰤ᩌ㍒

Find-S ㇇⌅Ⲵ‫ޣ‬䭞⢩⛩൘Ҿ˖ሩԕ኎ᙗ㓖ᶏⲴਸਆᔿ᧿䘠Ⲵ‫ٷ‬䇮オ䰤˄ྲ EnjoySport

‫ى‬䘹⎸䲔㇇⌅Ⲵᓄ⭘ᴹ˖Ӿॆᆖ䍘䉡࠶᷀˄chemical mass spectroscopy˅ѝᆖҐ㿴ࡉᙗ

定义： ањ‫ٷ‬䇮 h о䇝㓳ṧֻ䳶ਸ D а㠤(consistent)ˈᖃфӵᖃሩ D ѝ⇿ањṧֻ

Ѫ䈤᰾ਈරオ䰤Ⲵ䘉⿽㺘⽪ˈ޽а⅑㘳㲁㺘 2-2 ѝ᧿䘠Ⲵ EnjoySport ᾲᘥᆖҐ䰞仈Ǆሩ

h˙<Sunny, Warm, ?, Strong, ?, ?>

ᇎ䱵кˈ䘉ਚᱟ H ѝо䇝㓳ṧֻа㠤Ⲵᡰᴹ 6 њ‫ٷ‬䇮ѻаǄᡰᴹ 6 њ‫ٷ‬䇮൘മ 2-3 ѝ

定义： ‫ޣ‬Ҿ‫ٷ‬䇮オ䰤 H ઼䇝㓳ᮠᦞ D Ⲵа а㡜䗩⭼˄General boundary˅Gˈᱟ൘ H ѝ

定义： ‫ޣ‬Ҿ‫ٷ‬䇮オ䰤 H ઼䇝㓳ᮠᦞ D Ⲵ⢩ ⢩↺䗩⭼˄Specific boundary˅Sˈᱟ൘ H ѝ

Sį{ sęH | Consistent(s, D)ġ(s´ęH)[(s˚g s´) ġConsistent(s´, D)]}

ਚ㾱䳶ਸ G ઼ S 㻛㢟ྭൠᇊѹҶ˄㿱Ґ仈 2.7˅

ᇊ⨶ 2-1 ਈරオ䰤㺘⽪ᇊ⨶ǄԔ X Ѫаԫ᜿Ⲵᇎֻ䳶ਸˈH оѪ X кᇊѹⲴᐳቄ‫ٷ‬䇮Ⲵ䳶ਸǄ

VSH,D = { hęH | (sęS) (gęG) (gıghıgs)}

മ 2-4 ╄⽪Ҷ‫ى‬䘹⎸䲔㇇⌅ᓄ⭘ࡠ㺘 2-1 ѝཤєњ䇝㓳ṧֻᰦⲴ䘀㹼↕僔Ǆྲкᡰ䘠ˈ

Training examples: 䇝㓳ṧֻ

Training examples: 䇝㓳ṧֻ

ᴹ 6 њ኎ᙗਟԕ⭘ᶕ֯G2⢩ॆˈѪӰѸਚᴹ 3 њ൘G3ѝ઒˛∄ྲh=<?, ?, Normal, ?, ?, ?>

ㅜഋњ䇝㓳ṧֻˈྲമ 2-6 ᡰ⽪ˈ֯ਈරオ䰤Ⲵ S 䗩⭼ᴤа㡜ॆǄᆳҏሬ㠤 G 䗩⭼ѝ

Training examples: 䇝㓳ṧֻ

മ 2-7 EnjoySport ᾲᘥᆖҐ䰞仈ѝⲴᴰ㓸Ⲵਈරオ䰤

޽⅑㘳㲁മ 2-3 ѝᡰ⽪ⲴӾ EnjoySport Ⲵ 4 њṧֻѝᆖҐࡠⲴਈරオ䰤Ǆ䘉ᰦᆖҐಘᘾ

<Sunny, Warm, Normal, Light, Warm, Same>

⌘᜿䘉аᇎֻ┑䏣ਈරオ䰤Ⲵ 6 њ‫ٷ‬䇮ѝⲴ 3 њǄྲ᷌ᯭᮉ㘵ሶᇎֻࡂ࠶Ѫ↓ֻˈਈර

Sį{ sęH | Consistent(s, D)ġ(s´ęH)[(s˚g s´) ġConsistent(s´, D)]}

VSH,D = { hęH | (sęS) (gęG) (gıghıgs)}

(xięX)[ (BġDcġxi) ƅ L(xi, Dc)] (2.1)

Entropy(S) {-plog2p-plog2p ˄3.1˅

Entropy ([9,5]) (9 / 14) log 2 (9 / 14) (5 / 14) log 2 (5 / 14)

Entropy ( S ) (8 / 14) Entropy ( SWeak ) (6 / 14) Entropy ( S Strong )

定义：㒭ᅮϔϾ‫؛‬䆒ぎ䯈 HˈϔϾ‫؛‬䆒 hHˈབᵰᄬ೼݊Ҫⱘ‫؛‬䆒 h´HˈՓ