Constraint Programming

Michael Trick
(actually 75% Pascal Van Hentenryck, 20% Irv Lustig, 5% Trick)

Carnegie Mellon

Outline

Motivation An Overview of Constraint Programming Constraint Programming at Work

Getting Started Sports Scheduling Manufacturing

Perspectives

Combinatorial Optimization

Many, many practical applications

Resource allocation, scheduling, routing Computationally difficult Technical and modeling expertise needed Experimental in nature Important ($$$) in practice Integer programming Specialized methods Local search/metaheuristics Constraint programming

Properties

Many solution techniques

Constraint programming

Began in 1980s from AI world

Prolog III (Marseilles, France) CLP(R) CHIP (ECRC, Germany) Scheduling, sequencing, resource and personnel allocation, etc. etc.

Application areas

Active research area

Specialized conferences (CP, CP/AI-OR, ) Journal (Constraints) Companies

Constraint Programming

Two main contributions

A new approach to combinatorial optimization

Orthogonal and complementary to standard OR methods Combinatorial versus numerical Rich language for constraints Language for search procedures Vertical extensions

A new language for combinatorial optimization

The Tutorial

Goal: to provide an introduction

What is constraint programming? What is it good for? How does it compare to integer programming? How easy is it to use? What is the underlying technology?

Constraint Programming

Constraint programming by example

Illustrate rich language Contrast with integer programming Illustrate some underlying technologies Cant cover all of CP I want to make you curious Could use many; choose OPL

Disclaimers

Language/system used

Modeling in Constraint Programming

A rich constraint language

Arithmetic, higher-order, logical constraints Global constraints for natural substructures Definition of search tree to explore Specification of search strategy

Specification of a search procedure

Comparison of CP/IP

Branch and Prune

Branch and Bound

Prune: eliminate infeasible configurations Branch: decompose into subproblems

Bound: eliminate suboptimal solutions Branch: decompose into subproblems

Use (linear) relaxation of problem (+ cuts) Use information from relaxation

Prune

Bound

Carefully examine constraints to reduce possible variable values

Branch

Branch

Use heuristics based on feasibility info

Main focus:constraints and feasibility

Main focus: objective function and optimality

Illustrative artificial example

Color a map of (part of) Europe: Belgium, Denmark, France, Germany, Netherlands, Luxembourg No two adjacent countries same color Is four colors enough?

OPL example
enum Country {Belgium,Denmark,France,Germany,Netherlands,Luxem bourg}; enum Colors {blue,red,yellow,gray}; var Colors color[Country]; solve { color[France] <> color[Belgium]; color[France] <> color[Luxembourg]; color[France] <> color[Germany]; color[Luxembourg] <> color[Germany]; color[Luxembourg] <> color[Belgium]; color[Belgium] <> color[Netherlands]; color[Belgium] <> color[Germany]; color[Germany] <> color[Netherlands]; color[Germany] <> color[Denmark]; };

Variables nonnumeric Constraints are non-linear Looks nothing like IP! Perfectly legal CP

Constraint Programming

Domain store

Constraints

For each variable: what is the set of possible values? If empty for any variable, then infeasible If singleton for any variable, then solution

Capture interesting and well studied substructures Need to

Determine if constraint is feasible WRT the domain store Prune impossible values from the domains

Constraints

Can have differing techniques to handle a constraint type: 3x+10y+2z + 4w = 4

x in {0,1}, y in {0,1,2}, z in {0,1,2}, w in {0,1}

Simple bound on sizes gives y in {0} More complicated handling gives x in {0}, y in {0}, z in {0,2}, w in {0,1}

Constraint Solving

General algorithm is

Repeat select a constraint c if c is infeasible wrt domain store return infeasible else apply pruning algorithm of c Until no value can be removed

Branching

Once the constraint solving is done, if the problem is not infeasible nor are the domains singletons, then apply the search method

Choose a variable x with non-singleton domain (d1, d2, di) Foreach d in (d1, d2, di) add constraint x=di to problem and solve

Show OPL solving coloring problem

Strength of CP

Since there is no need for a linear relaxation, the language can represent much more directly (no need for big-M IP formulations.

Examples of formulation abilities

Facility location: want a constraint that customer j can be assigned to warehouse i only if warehouse open. (y[i]=1 if warehouse i open) IP: x[i,j] is 1 if cust j assigned to i x[i,j] <= y[i] CP:x[j] is the warehouse cust j assigned to (not a 0,1 variable) y[x[j]] = 1;

Similar example
Routing type constraints. Let x[i] be the ith customer visited and d[i,j] be distance from i to j

sum (i in 1..n) d[x[i],x[i+1]]

gives total distance traveled

Formulation strengths

Logical requirements: if A=1 and B<=2 then either C>=3 or D=1. Really painful in IP. Straightforward in CP:

((A=1) &(B<=2)) => ((C>=3)\/(D=1))

Global Constraints

Recognize that some types of constraints come up often

Create specialized routines to handle

Strong pruning Efficient handling

Extend system to include these

Global constraint: alldifferent

Most well known and studied constraint. alldifferent(x,y,z) states that x, y, and z take on different values. So x=2, y=1, z=3 would be ok, but not x=1, y=3, z=1.

Clear uses in routing (x[i] is ith customer visited, alldifferent[x] says each customer visited at most once), very useful in many other situations.

Alldifferent feasibility and pruning

Feasibility? Given domains, create domain/variable bipartite graph

x1 x2 x3 x4 x5

1
2 3 4

Alldifferent feasibility and pruning

Pruning? Which edges are in no matching?

x1 x2 x3 x4 x5

1
2 3 4

Domain is sharply reduced

Global constraints

Many different types of constraints have specialized routines

distribute(card,value,base): the number of times value[i] appears in base is card[i] circuit(succ) : the values in succ form a hamiltonian circuit (so if you follow the sequence 1, succ[1], succ[succ[1]] etc, you will get a loop through 1..n.

Global constraints

Many others, and new ones being created all the time

Strengthen and expand the language Make modeling easier and more natural System is faster at finding solutions Details hidden to user

Vertical language extensions

Can add constraints and definitions to make modeling even more natural Ideas remain the same: there are domains and constraints; constraints check for feasibility and prune domains; a search strategy guides the system in finding solutions

Scheduling

Want concepts of jobs, machines, before, after, jobs requiring machines, and so on. Easy to extend

Example of scheduling
forall(j in Jobs) forall(t in 1..nbTasks-1) task[j,t] precedes task[j,t+1]; forall(j in Jobs)

forall(t in Tasks)
task[j,t] requires tool[resource[j,t]];

Search Strategy

Combined with model, search strategies are integral to constraint systems. Allow choice of branching variables or more powerful search strategies Can be key in solving problems Two steps

Specify tree to search Specify how to explore the tree

Example of Search Strategies

forall(s in Stores ordered by increasing regretdmax(cost[s])) tryall(w in Warehouses ordered by increasing supplyCost[s,w]) supplier[s] = w; };

implements a maximum regret ordering (find a store with maximum regret then order the warehouses by increasing cost)

Example Problem

Painting cars (from Magnanti and Sokel). Sequence cars to minimize paint changeover Cars cannot be sequenced too far out of order

Small example
Small Example: 10 cars in sequence. The order for assembly is 1, 2, ..., 10. A car must be painted within 3 positions of its assembly order. For instance, car 5 can be painted in positions 2 through 8 inclusive. Cars 1, 5, and 9 are red; 2, 6, and 10 are blue; 3 and 7 green; and 4 and 8 are yellow. Initial sequence 1, 2, ... 10 corresponds to color pattern RBGYRBGYRB and has 9 purgings. The sequence 2,1,5,3,7,4,8,6,10,9 corresponds to color pattern BRRGGYYBBR and has 5 purgings.

Constraint Program
int n=; int rnge=; int ncolor=; range Slots 1..n; var Slots slot[1..n]; var Slots revslot[1..n]; int color[1..n]= ; minimize sum (j in 1..n-1) (color[revslot[j]] <> color[revslot[j+1]]) subject to { forall (i in Slots) i-rnge<=slot[i] <= i+rnge; /*Must be in range */ alldifferent(slot); /*must choose different slots */ forall (i in Slots) revslot[slot[i]] = i; };

Personal use

Tremendous help in my work on sports scheduling: much easier to formulate idiosyncratic constraints Very fast to create prototypes Competitive (at least!) to IP approaches

Result

Formulation is much easier than IP formulation Gets good solutions much faster than IP Is competitive in proving optimality

Finding optimal solutions

Constraint programs can find optimal solutions. Typically works by finding a feasible solution and adding a constraint that future solutions must be better than it. Repeat until infeasible: the last solution found is optimal

Perspectives

Many solution techniques

Integer programming Constraint programming Local search Combinations

Which to use?

Comparing IP and CP

Complementary technologies

Integer programming

Objective function: relaxations

Constraint programming

Feasibility: domain reductions

Might need to experiment with both CP particularly useful when IP formulation is hard or relaxation does not give much information

Combining Methods

Local and Global Search

Use CP/IP for very large neighborhood search (take a solution, remove large subset, find optimal completion) Use LP as constraint handler Use CP as subproblem solver in branch and price

Combining CP and IP

Conclusions

Constraint programming should become a part of every OR persons toolkit Combinations of CP and IP represent a big thing in future techniques Blurring of lines between optimization and heuristics This talk at http://mat.gsia.cmu.edu/INFORMS/cp.ppt