A Linear Search Strategy Using Bounds Sharlee Climer and Weixiong Zhang.

29
A Linear Search Strategy Using Bounds Sharlee Climer and Weixiong Zhang

Transcript of A Linear Search Strategy Using Bounds Sharlee Climer and Weixiong Zhang.

A Linear Search Strategy Using Bounds

Sharlee Climer andWeixiong Zhang

Sharlee Climer Washington University in St. Louis 2

Overview Introduction

Example problem

Results for Traveling Salesman Problem

Future Work

Sharlee Climer Washington University in St. Louis 3

Introduction Linear search strategy called cut-and-solve At each step:

A chunk of the solution space is cut away and solved, providing incumbent solutions

A relaxed solution is found for remaining solution space

Iterate until relaxed solution is greater than or equal to incumbent

Cut-and-solve may not be useful for simple problem instances

Sharlee Climer Washington University in St. Louis 4

Introduction

Use cut-and-solve to solve Linear Programs (LPs)

LPs are useful for modeling: Traveling Salesman Problem Constraint Satisfaction Problem Minimum cost flow problem

Sharlee Climer Washington University in St. Louis 5

Introduction

Asymmetric Traveling Salesman Problem (ATSP) can be used to model: No-wait flowshop Stacker crane Tilted drilling machine Computer disk read head Robotic motion Pay phone coin collection

Sharlee Climer Washington University in St. Louis 6

Introduction

Minimize Z = cij xij

s.t.: xij = 1 for j = 1,…,n

xij = 1 for i = 1,…,n

xij <= |W| - 1, for all proper non-empty subsets W of V

xij = {0,1}

Sharlee Climer Washington University in St. Louis 7

Introduction NP-hard Frequently solved using search trees

Branch-and-bound Branch-and-cut

Search strategy Best-first search Depth-first search

Cut-and-solve has minimal memory requirements and no “wrong” subtrees

Sharlee Climer Washington University in St. Louis 8

Introduction

ATSP solution space is a high-dimensional convex polyhedron

See paper for algorithm

Simple 2D problem for example

Sharlee Climer Washington University in St. Louis 9

Example

Minimize Z = y – 4/5 x

s.t.: x >= 0y <= 3y + 13/6 x <= 9y – 5/13 x >= 1/14y + 3/5 x >= 6/5x,y integers

Sharlee Climer Washington University in St. Louis 10

Example

x >= 0y <= 3y + 13/6 x <= 9y – 5/13 x >= 1/14y + 3/5 x >= 6/5x,y integers

Sharlee Climer Washington University in St. Louis 11

Example

Minimize Z = y – 4/5 x

x = 0y = 3Z = 3

Sharlee Climer Washington University in St. Louis 12

Example

Minimize Z = y – 4/5 x

x = 2y = 1Z = -0.6

Sharlee Climer Washington University in St. Louis 13

Example

Minimize Z = y – 4/5 x

x = 3.5y = 1.4Z = -1.4

Sharlee Climer Washington University in St. Louis 14

Example

Add new constraint:

y – 17/3 x >= -14

Sharlee Climer Washington University in St. Louis 15

Example

Incumbent solution:

x = 3y = 2Z = -0.4

Sharlee Climer Washington University in St. Louis 16

Example

Minimize Z = y – 4/5 x

x = 2.6y = 1.0Z = -1.1

Sharlee Climer Washington University in St. Louis 17

Example

Add new constraint:

Sharlee Climer Washington University in St. Louis 18

Example

New incumbent solution:

x = 2y = 1Z = -0.6

Sharlee Climer Washington University in St. Louis 19

Example

Minimize Z = y – 4/5 x

x = 1.0y = 0.6Z = -0.2

Incumbent solution:Z = -0.6

Sharlee Climer Washington University in St. Louis 20

Example

Keep new constraints and incumbent solution from one iteration to the next

Incumbent yields pruning opportunities

No children to choose between Need to decide size of cut

Sharlee Climer Washington University in St. Louis 21

ATSP results

Implemented using cplex with default settings

Compared with Concorde (Applegate, Bixby, Chvatal, & Cook)

Concorde designed for STSP Used 2-node transformation

Sharlee Climer Washington University in St. Louis 22

ATSP results

Testbed: All 27 TSPLIB instances 6 each: rtilt, stilt, crane, disk, coin, shop,

and super (Fischetti, Lodi, & Toth, 2002)

Cut-and-solve required 29.7% as much time as Concorde

Sharlee Climer Washington University in St. Louis 23

ATSP results

rtilt class

100 cities100 trials

Sharlee Climer Washington University in St. Louis 24

ATSP results

stilt class

Sharlee Climer Washington University in St. Louis 25

ATSP results

crane class

Sharlee Climer Washington University in St. Louis 26

ATSP results

disk class

Sharlee Climer Washington University in St. Louis 27

ATSP results

coin class

Sharlee Climer Washington University in St. Louis 28

ATSP results

shop class

Sharlee Climer Washington University in St. Louis 29

Future work

Improve ATSP solver: Automatic determination of cut size Use specialized sparse problem solver

Clustering algorithm