You are on page 1of 3

3/7/2008

Yang Liu (PhD student) Yang Liu (PhD student)

Agenda
Blekinge Institute of Technology
About the report
Competition Results
Agent 111 in Tac Competition Game
Design of Agent 111
Design, Strategy and His Secret of Winning
St t
Strategy in Agent 111
Improvement of Agent 111
Authors: Yang Liu
Date: 2008-02-23 Conclusion

Note: Some slides and/or pictures in the following are adapted


from slides ©IBMCRL, ©TAC Classic and ©MIT Courses.
© 2008 Yang Liu
© 2008 Yang Liu 2 Blekinge Institute of Technology © 2008 Yang Liu

Yang Liu (PhD student) Yang Liu (PhD student)

About the report Competition Results (1)


Objectives Average score is 3000
Online results
– Application of Multi-Agent Systems
– Position 90
– Group work & Competition
– Agent Agent 111
g with Java
– Problem Solving
– Avg Score 2493.26
2493 26
Reference – Games Played 11
– http://www.sics.se/tac/server/, TAC Classic website. – Zero Games 1
– RoxyBot-06: An (SAA)2 Tac Travel Agent, Seong Jae Lee,
et, al. IJCAI, 2007. Figure 1: Results online of Agent
111

3 Blekinge Institute of Technology © 2008 Yang Liu 4 Blekinge Institute of Technology © 2008 Yang Liu

Yang Liu (PhD student) Yang Liu (PhD student)

Competition Results (2) Design of Agent 111


In BTH (3 groups, 3 rounds) Based on DummyAgent
– Score 4216, -2223, 2613 – The framework of agent implementation
– Rank 2 in 1st round – Supported by TacAgent class
– Rank 1 in 2nd round Mainly tasks in DummyAgent
– Rank 1 in 3rd round – Initialization of 28 auctions prices
– Total Rank is 1 for Agent 111 – Construct concerned auctions based on client preferences
– First round bidding price decision
– Update bidding price based on previous results

5 Blekinge Institute of Technology © 2008 Yang Liu 6 Blekinge Institute of Technology © 2008 Yang Liu

1
3/7/2008

Yang Liu (PhD student) Yang Liu (PhD student)

Strategy Selection of auctions


Strategy is decomposed into three strategies Flight auctions
according to different type of bidding in Agent 111 – Bid flight tickets that only preferred by clients
Strategy is based on the objective utility functions of Hotel auctions
clients
– Bid hotel according to flight [InFlightDay, OutFlightDay)
– Bid TT hotel if and only if HP > 75
– where
Entertainment auctions
– Try to sell useless tickets (not involved in travel days)
– Try to buy useful tickets (involved in travel days)

7 Blekinge Institute of Technology © 2008 Yang Liu 8 Blekinge Institute of Technology © 2008 Yang Liu

Yang Liu (PhD student) Yang Liu (PhD student)

Strategy for bidding flight tickets Results of bidding flight tickets


Constraints
– Unlimited tickets, no competition with other agents
– Random prices [150, 800]
y one-side auction (buy
– Only y bid for agents)
g
– Close at the end of the game
Strategy in Agent 111
Ensure buy all tickets that we needed
– Buy needed flight tickets when price is only lower than 400
Avoid buying with a high price which is
encountered by DummyAgent

9 Blekinge Institute of Technology © 2008 Yang Liu 10 Blekinge Institute of Technology © 2008 Yang Liu

Yang Liu (PhD student) Yang Liu (PhD student)

Strategy for bidding hotel Results of bidding hotel


Constraints 1. Threshold 251 in
– Two type of hotel TT(better one) & SS
Agent 111 is well
performed to buy all
– Each has only 16 rooms for one day, no asking price
tickets we wanted.
– Package of flight-hotel is available only when rooms are available
according to flight arrival and departure time 2. What if there is no
2
vacant rooms for
Strategy
TT? ( Try to buy
– Basic SS) This is
• Try agent’s best to buy hotel according to flight time intelligent decision.
– Set an optimistic price to win bids - 251 in Agent 111

11 Blekinge Institute of Technology © 2008 Yang Liu 12 Blekinge Institute of Technology © 2008 Yang Liu

2
3/7/2008

Yang Liu (PhD student) Yang Liu (PhD student)

Strategy for bidding entertainment tickets Reasons why we get the negative score in 2nd
(Key decision about game score) round
Constraints We did have changed our strategy in 2nd round about
– Initially distribute to each agent a number of tickets bidding entertainment tickets
– Agents can exchange tickets between each other, – Do not too greedy in 2nd round, ☺
unused tickets in the final will have penalties
But, I faced compilation problems about our new
• C
Complex
l - two
t side
id biddi
bidding strategy
t t strategy in 2nd round
• Score - do not buy any useless tickets any more, and
try agents’ best to sell useless tickets, or negative score – Not correctly used javac complier in Dos/WinXP, the right
will be occur! command is as follows:
• Javac –classpath src –d bin bth\multiagents\GroupAAgent.java
Strategy ( Do not too greedy ! )
Fortunately, I fixed the problem in the 3rd round.
– Be careful of buy tickets
And finally we win.
– Try to sell disastrous tickets at a lower price!

13 Blekinge Institute of Technology © 2008 Yang Liu 14 Blekinge Institute of Technology © 2008 Yang Liu

Yang Liu (PhD student) Yang Liu (PhD student)

What can be improved in Agent 111 Summary


Multiple bidding format Team work
– Simultaneously bid an auction using different prices, but – It’s very hard to make a project
very complex for modeling and bidding. teams work correctly
– But we need to try
Introduce more intelligence in Agent 111
– Communication
C i i iis iimportant
– Price prediction, wait to buy until the price is the lowest
price Can a project be efficiently
– Robust to be cheated by other competitors partitioned?
• Avoid negative score at least ( basic win requirement ) – Functional - API based partition
– Task - Strategy, Design and
Testing

15 Blekinge Institute of Technology © 2008 Yang Liu 16 Blekinge Institute of Technology © 2008 Yang Liu

Yang Liu (PhD student) Yang Liu (PhD student)

Reference Thank you!


Executing code of BTHGPA is available at Q & A
– http://www.esnips.com/web/liuyang2006-Development

This
his document can be found by
– liuyang2006@gmail.com
– http://www.esnips.com/web/liuyang2006-Documentation

17 Blekinge Institute of Technology © 2008 Yang Liu 18 Blekinge Institute of Technology © 2008 Yang Liu

You might also like