Cutting Planes for First Level RLT Relaxations of Mixed 0-1 Programs

Similar documents
Cutting Planes for RLT Relaxations of Mixed 0-1 Polynomial Programs

Cutting Planes for RLT Relaxations of Mixed 0-1 Polynomial Programs

A new family of facet defining inequalities for the maximum edge-weighted clique problem

A Note on Representations of Linear Inequalities in Non-Convex Mixed-Integer Quadratic Programs

Nonconvex Quadratic Programming: Return of the Boolean Quadric Polytope

RLT-POS: Reformulation-Linearization Technique (RLT)-based Optimization Software for Polynomial Programming Problems

linear programming and approximate constraint satisfaction

An RLT Approach for Solving Binary-Constrained Mixed Linear Complementarity Problems

Convexification of Mixed-Integer Quadratically Constrained Quadratic Programs

Section Notes 9. IP: Cutting Planes. Applied Math 121. Week of April 12, 2010

Lift-and-Project Inequalities

Comparing Convex Relaxations for Quadratically Constrained Quadratic Programming

Quadratic reformulation techniques for 0-1 quadratic programs

Decomposition Methods for Quadratic Zero-One Programming

Notes on the decomposition result of Karlin et al. [2] for the hierarchy of Lasserre by M. Laurent, December 13, 2012

CSC Linear Programming and Combinatorial Optimization Lecture 12: The Lift and Project Method

Hierarchies. 1. Lovasz-Schrijver (LS), LS+ 2. Sherali Adams 3. Lasserre 4. Mixed Hierarchy (recently used) Idea: P = conv(subset S of 0,1 n )

L R I R A P P O R T D E R E C H E R C H E KNAPSACK PROBLEM WITH PROBABILITY CONSTRAINTS GAIVORONSKI A A / LISSER A / LOPEZ R

Lift-and-Project Techniques and SDP Hierarchies

Semidefinite Programming Basics and Applications

AM 121: Intro to Optimization! Models and Methods! Fall 2018!

Cutting Planes in SCIP

MINI-TUTORIAL ON SEMI-ALGEBRAIC PROOF SYSTEMS. Albert Atserias Universitat Politècnica de Catalunya Barcelona

Moments and Positive Polynomials for Optimization II: LP- VERSUS SDP-relaxations

On Non-Convex Quadratic Programming with Box Constraints

Cuts for mixed 0-1 conic programs

Duality. Geoff Gordon & Ryan Tibshirani Optimization /

c 2000 Society for Industrial and Applied Mathematics

CORC REPORT Approximate fixed-rank closures of covering problems

Separation Techniques for Constrained Nonlinear 0 1 Programming

Travelling Salesman Problem

CSCI 1951-G Optimization Methods in Finance Part 10: Conic Optimization

On the Polyhedral Structure of a Multi Item Production Planning Model with Setup Times

Online generation via offline selection - Low dimensional linear cuts from QP SDP relaxation -

A notion of Total Dual Integrality for Convex, Semidefinite and Extended Formulations

Polyhedral Approach to Integer Linear Programming. Tepper School of Business Carnegie Mellon University, Pittsburgh

Lecture 2. Split Inequalities and Gomory Mixed Integer Cuts. Tepper School of Business Carnegie Mellon University, Pittsburgh


Reduced rst-level representations via the reformulation-linearization technique: results, counterexamples, and computations

Moments and Positive Polynomials for Optimization II: LP- VERSUS SDP-relaxations

Comparisons and enhancement strategies for linearizing mixed 0-1 quadratic programs

Successive Convex Relaxation Methods for Nonconvex Quadratic Optimization Problems

arxiv: v15 [math.oc] 19 Oct 2016

Cutting planes from extended LP formulations

Convex Optimization M2

PORTA, NEOS, and Knapsack Covers. Cover Inequalities. Prof. Jeff Linderoth. October 29, June 23, 2004 DIMACS Reconnect Conference on MIP Slide 1

Solving Box-Constrained Nonconvex Quadratic Programs

MIP reformulations of some chance-constrained mathematical programs

CORC Technical Report TR Cuts for mixed 0-1 conic programming

Easy and not so easy multifacility location problems... (In 20 minutes.)

Sequential pairing of mixed integer inequalities

Cutting Plane Methods I

A General Framework for Convex Relaxation of Polynomial Optimization Problems over Cones

Introduction to LP and SDP Hierarchies

On the knapsack closure of 0-1 Integer Linear Programs. Matteo Fischetti University of Padova, Italy

On Valid Inequalities for Quadratic Programming with Continuous Variables and Binary Indicators

6-1 The Positivstellensatz P. Parrilo and S. Lall, ECC

E5295/5B5749 Convex optimization with engineering applications. Lecture 5. Convex programming and semidefinite programming

The Trust Region Subproblem with Non-Intersecting Linear Constraints

arxiv: v2 [math.oc] 3 Nov 2016

Two-stage stochastic (and distributionally robust) p-order conic mixed integer programs: Tight second stage formulations

Deciding Emptiness of the Gomory-Chvátal Closure is NP-Complete, Even for a Rational Polyhedron Containing No Integer Point

Lecture: Examples of LP, SOCP and SDP

Introduction to Semidefinite Programming I: Basic properties a

Lagrange Duality. Daniel P. Palomar. Hong Kong University of Science and Technology (HKUST)

LMI Methods in Optimal and Robust Control

Numerical Optimization. Review: Unconstrained Optimization

Copositive Programming and Combinatorial Optimization

Hypergraph Matching by Linear and Semidefinite Programming. Yves Brise, ETH Zürich, Based on 2010 paper by Chan and Lau

Lecture 18: Optimization Programming

Sherali-Adams Relaxations of Graph Isomorphism Polytopes. Peter N. Malkin* and Mohammed Omar + UC Davis

Applications of semidefinite programming in Algebraic Combinatorics

l p -Norm Constrained Quadratic Programming: Conic Approximation Methods

On the Lovász Theta Function and Some Variants

5 Integer Linear Programming (ILP) E. Amaldi Foundations of Operations Research Politecnico di Milano 1

Tightening linearizations of non-linear binary optimization problems

Convex Optimization. (EE227A: UC Berkeley) Lecture 6. Suvrit Sra. (Conic optimization) 07 Feb, 2013

Lecture #21. c T x Ax b. maximize subject to

LP formulations for mixed-integer polynomial optimization problems Daniel Bienstock and Gonzalo Muñoz, Columbia University, December 2014

Polynomial and Synthetic Division

Relations between Semidefinite, Copositive, Semi-infinite and Integer Programming

A finite branch-and-bound algorithm for nonconvex quadratic programming via semidefinite relaxations

Operations Research Lecture 6: Integer Programming

Copositive Programming and Combinatorial Optimization

An Introduction to Linear Matrix Inequalities. Raktim Bhattacharya Aerospace Engineering, Texas A&M University

There are several approaches to solve UBQP, we will now briefly discuss some of them:

Lecture 9: Dantzig-Wolfe Decomposition

The Retail Planning Problem Under Demand Uncertainty. UCLA Anderson School of Management

Network Flows. 6. Lagrangian Relaxation. Programming. Fall 2010 Instructor: Dr. Masoud Yaghini

Computational Integer Programming. Lecture 2: Modeling and Formulation. Dr. Ted Ralphs

arxiv:math/ v1 [math.co] 23 May 2000

minimize x x2 2 x 1x 2 x 1 subject to x 1 +2x 2 u 1 x 1 4x 2 u 2, 5x 1 +76x 2 1,

The Traveling Salesman Problem: Inequalities and Separation

13. Convex programming

Extended Linear Formulation for Binary Quadratic Problems

3.7 Strong valid inequalities for structured ILP problems

15.083J/6.859J Integer Optimization. Lecture 2: Efficient Algorithms and Computational Complexity

An exact reformulation algorithm for large nonconvex NLPs involving bilinear terms

Polyhedral Results for A Class of Cardinality Constrained Submodular Minimization Problems

Distributed Computation of Minimum Time Consensus for Multi-Agent Systems

Transcription:

Cutting Planes for First Level RLT Relaxations of Mixed 0-1 Programs 1 Cambridge, July 2013 1 Joint work with Franklin Djeumou Fomeni and Adam N. Letchford

Outline 1. Introduction 2. Literature Review 3. Cutting Planes 4. Separation 5. Computational Experiments 6. Concluding Remarks

Introduction The Reformulation-Linearization Technique (RLT), due to Adams and Sherali, is a technique for constructing hierarchies of LP relaxations for mixed-integer nonlinear programs. In practice, the first-level relaxation is usually the most useful. One can construct a first-level relaxation of any mixed 0-1 program with a linear or quadratic objective and any mixture of convex and quadratic constraints. We present several families of cutting planes, which can be used to strengthen such first-level relaxations.

Literature Review The seminal paper, Adams & Sherali (1986), was concerned with 0-1 quadratic programs of the form: { max x T Qx : Ax b, x {0,1} n}, where Q Q n n, A Q m n and b Q m. (Without loss of generality, we assume that Q is symmetric.) We will let V denote {1,...,n} and E denote {S V : S = 2}.

Literature Review (cont.) Let us introduce a new variable y ij for all {i,j} E, taking the value 1 if and only if both x i and x j do. The 0-1 QP can then be reformulated as the 0-1 LP: max s.t. i V Q ii x i + 2 e E Q e y e Ax b y ij x i ({i,j} E) y ij x j ({i,j} E) x i + x j y ij 1 ({i,j} E) (x,y) {0,1} V + E (In fact, this was already noted by Fortet, 1959.)

Literature Review (cont.) Replacing the condition (x,y) {0,1} V +E with the condition (x,y) [0,1] V +E yields an LP relaxation of the 0-1 QP. Adams and Sherali propose to strengthen the relaxation by adding 2mn additional inequalities as follows. Let α T x β be a constraint in the linear system Ax b, and let x k be one of the original variables. Then we have: (α T x)x k βx k = α i y ik (β α k )x k i V \{k} (α T x)(1 x k ) β(1 x k ) = α i (x i y ik ) β(1 x k ). i V \{k}

Literature Review (cont.) Sherali & Adams (1990) showed that, if the same approach is applied to a 0-1 LP, the LP relaxation of the reformulated problem dominates that of the original. They called the procedure the Reformulation-Linearization Technique. They also defined an entire hierarchy of relaxations, where the kth level RLT relaxation has variables representing products of up to k + 1 variables, and constraints that are derived by multiplying k + 1 of the original inequalities together.

Literature Review (cont.) There have been many developments since those early days: Lovász & Schrijver (1991) adapted the first-level relaxation to 0-1 problems with convex constraints, and also showed how to derive SDP relaxations that dominate the LP relaxation. In a series of papers, Adams & Sherali extended RLT to any bounded MINLP having a polynomial objective function and a mixture of convex and polynomial constraints. Padberg and others have studied the polytopes associated with various specific 0-1 QPs.

The New Cutting Planes The number of variables grows very rapidly with the level k. To keep the number of variables manageable, we concentrate on the first-level relaxation. We derive cutting planes using a two-phase approach: 1. Construct a valid cubic inequality; 2. Weaken it to make it valid for the first-level relaxation.

The New Cutting Planes: Phase 1 To construct valid cubic inequalities, we simply multiply triples of valid linear inequalities together. Suppose that the linear inequalities α i x β i, for i = 1,2,3, are valid for the original problem. We write them as β i α i x 0, and thereby derive the valid cubic inequality: (β 1 α 1 x)(β 2 α 2 x)(β 3 α 3 x) 0.

The New Cutting Planes: Phase 2 The key to phase 2 is the following lemma: Lemma 1 Suppose x i [0,1] for all i V, and y ij = x i x j for all e E. Then, for any triple {i,j,k} V, we have the following bounds on x i x j x k : x i x j x k min{y ij, y ik, y jk, 1 x i x j x k + y ij + y ik + y jk } x i x j x k max{0, y ij + y ik x i, y ij + y jk x j, y ik + y jk x k }.

The New Cutting Planes: Phase 2 (cont.) Due to the above lemma, we can convert a valid cubic inequality into a valid inequality for the first-level relaxation by replacing each term of the form x i x j x k with: any of 0, y ij + y ik x i, y ij + y jk x j or y ik + y jk x k (if it has a positive LHS coefficient). any of y ij, y ik, y jk or 1 x i x j x k + y ij + y ik + y jk (if the term has a negative LHS coefficient);

A special case: (s, t) inequalities Consider a single linear inequality α T x β and two binary variables x s and x t. We can form the following cubic inequality: (β α T x)x s x t 0. Since x 2 s x t = x s x 2 t = y st, we can re-write this as: α i x i x s x t (β α s α t )y st. i V \{s,t} To make further progress, we need more notation: for any S V, we let S + = {i S : α i > 0} and S = {i S : α i < 0}.

A special case: (s, t) inequalities (cont.) Then, for any partition of V into sets S,T,W and R, we can derive the (s, t) inequality: i S W α i y is + i T W α i y it α i x i α(w ) + α(s + W )x s i W + α(t + W )x t + ( β α({s,t} S + T + W R ) ) y st. Here, membership of i in S, T, W or R determines which of the weakenings x i x s x t min{y is, y it, y st, 1 x i x s x t + y st + y is + y it } x i x s x t max{0, y is + y st x s, y it + y st x t, y is + y it x i } we choose for the term x i x s x t.

More special cases: mixed and reverse (s, t) inequalities Following the same line of argument, we can derive: mixed (s, t) inequalities, by applying Lemma 1 to the cubic inequalities of form (β α T x)(1 x s )x t 0. reverse (s, t) inequalities, by applying Lemma 1 to the cubic inequalities of form (β α T x)(1 x s )(1 x t ) 0.

Remark I Using a simple disjunctive argument, it is possible to strengthen all three kinds of (s, t) inequalities. Basically, we compute upper bounds on the value that the LHS can take under the following four (mutually exclusive) conditions: x s = x t = y st = 0, x s = y st = 0 and x t = 1, x t = y st = 0 and x s = 1 x s = x t = y st = 1. We then use those bounds to tighten the RHS coefficients of x s, x t and y st.

Remark II The method can be applied to problems having: 1. A mixture of linear and quadratic constraints. 2. Non-linear constraints as long as they are convex.

Separation All three families of (s, t) inequalities are exponentially-large. So we need to think about separation. If the original problem has only linear constraints, then the separation problems for all three families can be solved in O(n 3 m) time.

Separation of standard (s, t) inequalities There are m ways to choose an inequality from Ax b. There are O(n 2 ) ways to select s and t. Then, for all i V +, we place i in S, T, W or R according to which of the following is largest: y is + y st x s y it + y st x t y is + y it x i 0. And, for all i V, we place i in S, T, W or R according to which of the following is largest: y is y it 1 x i x s x t + y is + y it + y st y st.

Separation (cont.) The separation problem for the mixed and reverse (s, t) inequalities can be tackled in a similar way. And the separation problems for the strengthened versions of all three families can be solved in O(n 3 m) time as well. For the cutting planes derived by multiplying together two linear inequalities and a variable or its complement, the separation time increases to O(n 3 m 2 ). For the ones derived by multiplying three linear inequalities, the separation time increases further to O(n 3 m 3 ).

Separation (cont.) Now consider the case in which there is a convex constraint x C, and we have a polynomial-time separation oracle for C (à la Lovász-Schrijver). We can show that the separation problems for all three kinds of strengthened (s, t) inequalities can be solved in polynomial time also in this case. On the other hand, we do not know how to solve the separation problem for the other inequalities in this case.

Computational Experiments We apply our scheme to the Quadratic Knapsack Problem: { max x T Qx : w T x c, x {0,1} n}, where Q Z n n +, w Zn + and c Z +. We randomly generate 5 instances for each different combination of size n {10,20,30,40,50} and density {25,50,75,100}. We separate all three families of strong (s, t) inequalities. We compare the resulting bounds against the benchmark upper bound given by the first-level RLT relaxation.

Computational Results Instances Gap (%) n RLTs RLTs + STs Improvement (%) 10 6.33 4.01 36.64 20 3.76 2.95 21.53 100 30 3.44 2.79 18.92 40 1.30 1.21 6.85 50 1.38 1.28 7.73 10 4.61 3.27 29.12 20 3.31 2.23 32.75 75 30 0.66 0.55 16.25 40 1.72 1.41 18.04 50 0.65 0.60 8.37

Computational Results Instances Gap (%) n RLTs RLTs + STs Improvement (%) 10 7.60 4.28 43.66 20 1.93 1.36 29.59 50 30 2.29 1.03 55.31 40 1.84 1.55 15.87 50 1.03 0.41 60.51 10 8.76 1.28 85.37 20 2.08 0.85 59.16 25 30 1.65 1.04 37.24 40 4.46 1.85 58.49 50 3.64 0.92 74.65

Concluding Remarks We have presented a general procedure for generating cutting planes for first-level RLT relaxations of mixed 0-1 programs. For the case in which all constraints are either linear or quadratic, we have polynomial-time separation routines for all families. For the case of convex constraints, we have polynomial-time separation routines only for the strengthened (s, t) inequalities. The computational results so far are quite encouraging.