15.

093: Optimization Methods

Lecture 12: Discrete Optimization

1 Todays Lecture
Modeling with integer variables What is a good formulation? Theme: The Power of Formulations

Slide 1

2 Integer Optimization
2.1 Mixed IO
(MIO) max c x + h y s.t. Ax + B y b n (x 0 x integer) x 2 Z+ y 2 Rm + (y 0)
0 0

Slide 2

2.2 Pure IO

(IO) max c x s.t. Ax b n x 2 Z+ Important special case: Binary Optimization (BO) max c x s.t. Ax b x 2 f0 1gn
0 0

Slide 3

2.3 LO

(LO) max c x s.t. By b y 2 Rn +

0

Slide 4

3 Modeling with Binary Variables

if event occurs x2 1 0 otherwise Example 1: IO formulation of the knapsack problem n : projects, total budget b aj : cost of project j cj : value of project j 1 if project j is selected. xj = 0 otherwise. 1

3.1 Binary Choice

Slide 5

Slide 6

n P max cj xj j 1 P s.t. aj xj b xj 2 f0 1g

3.2 Modeling relations

At most one event occurs

X
j

Slide 7

xj 1

Neither or both events occur x2 ; x1 = 0 If one event occurs then, another occurs 0 x2 x1 If x = 0, then y = 0 if x = 1, then y is uncontrained 0 y Ux x 2 f0 1g

3.3 The assignment problem

cij : cost of assigning person j to job i: person j is assigned to job i xij = 1 0 Pc x min ij ij n P s.t. xij = 1 each job is assigned

n people m jobs

Slide 8

xij 1 each person can do at most one job. i1 xij 2 f0 1g

j1 m P

4 What is a good formulation?

4.1 Facility Location

Data

Slide 9

N = f1 : : :ng potential facility locations I = f1 : : :mg set of clients cj : cost of facility placed at j hij : cost of satisfying client i from facility j:

Decision variables a facility is placed at location j xj = 1 0 otherwise yij = fraction of demand of client i satis ed by facility j:

n m P n P P IZ1 = min cj xj + hij yij j 1 i1 j 1 n P s.t. yij = 1 j 1 yij xj xj 2 f0 1g 0 yij 1: Consider an alternative formulation. n m P n P P h y IZ = min cx +

Slide 10

Slide 11

s.t.

yij m xj i1 xj 2 f0 1g 0 yij

j 1 m P

j 1 n P

j j

yij = 1

i1 j 1

ij ij

1:

Are both valid? Which one is preferable?

4.2 Observations

IZ1 = IZ2 , since the integer points both formulations de ne are the same. P1 = f(x y) :
n X

Slide 12

xj 1 yij = 1 yij xj 0 0 yij 1 j 1

n X j 1

P2 = f(x y) :

yij = 1

m X i1

yij m xj

0 xj 1 0 yij 1

Slide 13

Let

Z1 = min cx + hy (x y) 2 P1 Z2 Z1 IZ1 = IZ2

Z2 = min cx + hy (x y) 2 P2

4.3 Implications

Finding IZ1 (= IZ2 ) is di cult. Solving to nd Z1 Z2 is a LOP. Since Z1 is closer to IZ1 several methods (branch and bound) would work better (actually much better). Suppose that if we solve min cx + hy (x y) 2 P1 we nd an integral solution. Have we solved the facility location problem?

Slide 14

Slide 15

Formulation 1 is better than Formulation 2. (Despite the fact that 1 has a larger number of constraints than 2.) What is then the criterion?

4.4 Ideal Formulations

Let P be a linear relaxation for a problem Let H = f(x y) : x 2 f0 1gng \ P

Slide 16

Consider Convex Hull (H) X iX = fx : x = ix i = 1 i 0 xi 2 H g

i i
Slide 17

The extreme points of CH (H ) have f0 1g coordinates. So, if we know CH (H ) explicitly, then by solving min cx + hy (x y) 2 CH (H ) we solve the problem. Message: Quality of formulation is judged by closeness to CH (H ).

C H (H ) P1 P2

5 Minimum Spanning Tree (MST)

How do telephone companies bill you? It used to be that rate/minute: Boston ! LA proportional to distance in MST Other applications: Telecommunications, Transportation (good lower bound for TSP)

Slide 18

Slide 19

Given a graph G = (V E ) undirected and Costs ce e 2 E . Find a tree of minimum cost spanning all the nodes. 1 if edge e is included in the tree Decision variables xe = 0 otherwise The tree should be connected. How can you model this requirement? Let S be a set of vertices. Then S and V n S should be connected 2S Let (S ) = fe = (i j ) 2 E : i j 2V nS Then, X xe 1

Slide 20

e2 (S )

What is the number of edges in a tree? P Then, xe = n ; 1

e2E

P IZM ST = min cexe E 8 eP x 1 > < eP(S ) e H > x = n;1 >eE e : xe 2 f0 1g: Is this a good formulation?

5.1 Formulation

Slide 21
8S

V S= 6 V

Slide 22

Is Pcut the CH (H )?

e2 (S )

Pcut = fx 2 R E : 0 x e X xe = n ; 1 e E X 6 Vg xe 1 8 S V S =
j j 2

5.2 What is C H (H )?
Let Psub = fx 2 R E : xe = n ; 1 e E X xe jS j ; 1 8 S V S 6= V g
j j 2

Slide 23

e2E (S )

2S E (S ) = e = (i j ) : i j2S Why is this a valid IO formulation?

Slide 24

Theorem: Psub = CH (H ). ) Psub is the best possible formulation. MESSAGE: Good formulations can have an exponential number of constraints.

6 The Traveling Salesman Problem

Given G = (V E ) an undirected graph. V = f1 : : : ng, costs ce 8 e 2 E . Find a tour that minimizes total length.

Slide 25

6.1 Formulation I
xe =

1 if edge e is included in the tour. 0 otherwise. Pcx min e e eP E s:t: xe 2 S E eP (S ) xe = 2 i 2 V e (i) xe 2 f0 1g

Slide 26

6.2 Formulation II
min s:t:

Pc x Pe ex jS j ; 1 S E e eP E (S ) x =2 i2V
e e2 (i) xe 2 f0 1g

Slide 27

Slide 28

TS P = Pcut

fx 2 RjE j
j j

P
e2 (S )
2

xe 2

P
e2 (i)

xe = 2

TS P = Psub

0 xe 1g P fx 2 R E xe = 2 ) P x ejS j(i; 1 e e (S ) 0 xe 1g
2

Slide 29

Theorem: CH (H ) Nobody knows CH (H ) for the TSP

TS P Pcut

TSP 6 = Psub

7 Minimum Matching

Given G = (V E ) ce costs on e 2 E . Find a matching of minimum cost. Formulation: Pc x min P e e s:t: xe = 1 i 2 V e (i) xe 2 f0 1g

Slide 30

Is the linear relaxation CH (H )?

j j 2

P x =1 PM AT = fx 2 R E : e e (i) P 6 xe 1 jS j = 2k + 1 S = e (S ) xe 0g Theorem: PM AT = CH (H )
2

Let

Slide 31

8 Observations

For MST, Matching there are e cient algorithms. CH (H ) is known. For TSP 6 9 e cient algorithm. TSP is an NP ; hard problem. CH (H ) is not known. Conjuecture: The convex hull of problems that are polynomially solvable are explicitly known.

Slide 32

9 Summary

1. An IO formulation is better than another one if the polyhedra of their linear relaxations are closer to the convex hull of the IO. 2. A good formulation may have an exponential number of constraints. 3. Conjecture: Formulations characterize the complexity of problems. If a problem is solvable in polynomial time, then the convex hull of solutions is known.

Slide 33