An Introduction to Dual Ascent Heuristics

Similar documents
MVE165/MMG630, Applied Optimization Lecture 8 Integer linear programming algorithms. Ann-Brith Strömberg

56:272 Integer Programming & Network Flows Final Examination -- December 14, 1998

Algorithms for Integer Programming

3 INTEGER LINEAR PROGRAMMING

Integer Programming. Xi Chen. Department of Management Science and Engineering International Business School Beijing Foreign Studies University

ACO Comprehensive Exam October 12 and 13, Computability, Complexity and Algorithms

Introduction to Mathematical Programming IE406. Lecture 20. Dr. Ted Ralphs

56:272 Integer Programming & Network Flows Final Exam -- December 16, 1997

5.3 Cutting plane methods and Gomory fractional cuts

Last topic: Summary; Heuristics and Approximation Algorithms Topics we studied so far:

February 19, Integer programming. Outline. Problem formulation. Branch-andbound

Outline. Column Generation: Cutting Stock A very applied method. Introduction to Column Generation. Given an LP problem

Column Generation: Cutting Stock

Introduction to Mathematical Programming IE496. Final Review. Dr. Ted Ralphs

Linear Programming. Course review MS-E2140. v. 1.1

Fundamentals of Integer Programming

and 6.855J Lagrangian Relaxation I never missed the opportunity to remove obstacles in the way of unity. Mohandas Gandhi

Part 4. Decomposition Algorithms Dantzig-Wolf Decomposition Algorithm

ME 391Q Network Flow Programming

15.082J and 6.855J. Lagrangian Relaxation 2 Algorithms Application to LPs

MVE165/MMG631 Linear and integer optimization with applications Lecture 9 Discrete optimization: theory and algorithms

EXERCISES SHORTEST PATHS: APPLICATIONS, OPTIMIZATION, VARIATIONS, AND SOLVING THE CONSTRAINED SHORTEST PATH PROBLEM. 1 Applications and Modelling

Solutions for Operations Research Final Exam

Integer Programming Theory

Final Exam Spring 2003

The Size Robust Multiple Knapsack Problem

TIM 206 Lecture Notes Integer Programming

Primal-Dual Methods for Approximation Algorithms

Randomized rounding of semidefinite programs and primal-dual method for integer linear programming. Reza Moosavi Dr. Saeedeh Parsaeefard Dec.

Simulation. Lecture O1 Optimization: Linear Programming. Saeed Bastani April 2016

Unit.9 Integer Programming

Lagrangean Methods bounding through penalty adjustment

2 is not feasible if rounded. x =0,x 2

Optimizing the positioning of medical facilities using linear programming techniques

Crash-Starting the Simplex Method

MVE165/MMG631 Linear and integer optimization with applications Lecture 7 Discrete optimization models and applications; complexity

Overview. H. R. Alvarez A., Ph. D.

lpsymphony - Integer Linear Programming in R

Network Flows. 7. Multicommodity Flows Problems. Fall 2010 Instructor: Dr. Masoud Yaghini

Reload Cost Trees and Network Design

1. Lecture notes on bipartite matching February 4th,

CS 473: Algorithms. Ruta Mehta. Spring University of Illinois, Urbana-Champaign. Ruta (UIUC) CS473 1 Spring / 36

A survey on benders decomposition applied to xed-charge network design problems

x ji = s i, i N, (1.1)

DETERMINISTIC OPERATIONS RESEARCH

Application of Bounded Variable Simplex Algorithm in Solving Maximal Flow Model

Cloud Branching MIP workshop, Ohio State University, 23/Jul/2014

Civil Engineering Systems Analysis Lecture XIV. Instructor: Prof. Naveen Eluru Department of Civil Engineering and Applied Mechanics

Methods and Models for Combinatorial Optimization Exact methods for the Traveling Salesman Problem

Parallel Branch & Bound

Size of a problem instance: Bigger instances take

Approximation Algorithms: The Primal-Dual Method. My T. Thai

Toward the joint design of electronic and optical layer protection

Lagrangian Relaxation: An overview

Computational Complexity CSC Professor: Tom Altman. Capacitated Problem

Solutions to Assignment# 4

Gate Sizing by Lagrangian Relaxation Revisited

Computational Integer Programming. Lecture 12: Branch and Cut. Dr. Ted Ralphs

Mathematical and Algorithmic Foundations Linear Programming and Matchings

Lecture 3: Totally Unimodularity and Network Flows

to the Traveling Salesman Problem 1 Susanne Timsj Applied Optimization and Modeling Group (TOM) Department of Mathematics and Physics

Mathematical Tools for Engineering and Management

A. Atamturk. G.L. Nemhauser. M.W.P. Savelsbergh. Georgia Institute of Technology. School of Industrial and Systems Engineering.

Outline. Modeling. Outline DMP204 SCHEDULING, TIMETABLING AND ROUTING. 1. Models Lecture 5 Mixed Integer Programming Models and Exercises

Combinatorial Optimization

Approximation Algorithms

Lagrangean Relaxation of the Hull-Reformulation of Linear Generalized Disjunctive Programs and its use in Disjunctive Branch and Bound

Two-stage column generation

Problem set 2. Problem 1. Problem 2. Problem 3. CS261, Winter Instructor: Ashish Goel.

Algorithms for Decision Support. Integer linear programming models

Department of Mathematics Oleg Burdakov of 30 October Consider the following linear programming problem (LP):

Assignment 3b: The traveling salesman problem

arxiv: v1 [cs.dm] 6 May 2009

Modeling and Solving Location Routing and Scheduling Problems

Lecture 4: Primal Dual Matching Algorithm and Non-Bipartite Matching. 1 Primal/Dual Algorithm for weighted matchings in Bipartite Graphs

Minimum Weight Constrained Forest Problems. Problem Definition

An Improved Subgradiend Optimization Technique for Solving IPs with Lagrangean Relaxation

3 No-Wait Job Shops with Variable Processing Times

Heuristic Optimization Today: Linear Programming. Tobias Friedrich Chair for Algorithm Engineering Hasso Plattner Institute, Potsdam

Homework 2: Multi-unit combinatorial auctions (due Nov. 7 before class)

Advanced Operations Research Techniques IE316. Quiz 2 Review. Dr. Ted Ralphs

DISTRIBUTED NETWORK RESOURCE ALLOCATION WITH INTEGER CONSTRAINTS. Yujiao Cheng, Houfeng Huang, Gang Wu, Qing Ling

Approximation Algorithms

6.854 Advanced Algorithms. Scribes: Jay Kumar Sundararajan. Duality

ONLY AVAILABLE IN ELECTRONIC FORM

LP-Modelling. dr.ir. C.A.J. Hurkens Technische Universiteit Eindhoven. January 30, 2008

NATCOR Convex Optimization Linear Programming 1

Surrogate Gradient Algorithm for Lagrangian Relaxation 1,2

On the selection of Benders cuts

Column Generation II : Application in Distribution Network Design

Location Routing Problems on Trees

Steiner Trees and Forests

A Randomized Algorithm for Minimizing User Disturbance Due to Changes in Cellular Technology

15-854: Approximations Algorithms Lecturer: Anupam Gupta Topic: Direct Rounding of LP Relaxations Date: 10/31/2005 Scribe: Varun Gupta

Linear Programming. Linear programming provides methods for allocating limited resources among competing activities in an optimal way.

Outline. Combinatorial Optimization 2. Finite Systems of Linear Inequalities. Finite Systems of Linear Inequalities. Theorem (Weyl s theorem :)

Algorithmic expedients for the S-labeling problem

Column Generation in the Integral Simplex Method

Treatment Planning Optimization for VMAT, Tomotherapy and Cyberknife

6. Lecture notes on matroid intersection

Transcription:

An Introduction to Dual Ascent Heuristics

Introduction A substantial proportion of Combinatorial Optimisation Problems (COPs) are essentially pure or mixed integer linear programming. COPs are in general NP-hard. Hence, algorithms guaranteeing optimality need most often to be used on some kind of implicit enumeration, typically within the framework of branch-and-bound (BB). A successful BB-algorithm can be viewed as a lucky compromise between the quality of the bounds and the computational effort invested in obtaining them.

Bounding Techniques The most common bounding techniques are: Lagrangian Relaxation Linear programming relaxation Dual based procedures They are computationally more efficient than both Lagrangian and LP-relaxation techniques.

Primal min z p =cx s.t Ax b LP-Relaxation min z LP p =cx s.t Ax b Dual max z LP D =wb s.t wa c x {0,1} x 0 w 0 In dual ascent heuristics, an approximate solution to the primal problem and a feasible solution to the dual of an LP relaxation are constructed simultaneously; the performance guarantee is proved by comparing the values of both solutions.

Characteristics of a dual ascent technique Generating good lower bounds for the dual problem (maximization) relatively fast by solving the linear programming dual problem approximately. Identifying feasible network designs that serve as starting solutions for local improvement heuristics. Reducing the problem by eliminating some design variables.

General Discription Dual ascent heuristics solve the dual of the LPrelaxation approximately. In fact: They start with an initial solution. The solution will then iteratively improve according to an ascent strategy. In every iteration, the obtained solution is at least as good as the former one.

Researchers have successfully applied dual ascent to several network design related models including: Uncapacitated facility location problem (Bilde and Krarup, 1977; Erlenkotter, 1978; Van Roy and Erlenkotter, 1982). The generalized assignment problem (Fisher et al, 1986). The Steiner tree problem (Wong, 1984). The set covering problem (Kedia and Fisher, 1986). The set partitioning problem (Fisher and Kedia, 1986). In 1989, Balakrishnan et al, introduced a network design model formulation, together with a general dual-ascent framework, which generalized the former proposed dualascent methods.

In 2001, Rosenberg used the improved formulation by Balakrishnan et al, and introduced a different dual ascent strategy. In 2007, Canovas et al, proposed a dual ascent technique to solve a four indexed dual formulation of the uncapacitated multiple allocation hup location problem. In 2008, Boschetti et al, improved the former obtained lower bounds for the set partitionning problem, on the basis of a new dual ascent heuristic. In 2010, Bardossy and Raghavan proposed a heuristic that combines dual ascent and local search, which together yield strong lower and upper bounds for facility location problem.

Problem P min Problem DP max s.t (1) s.t (5) (2) (6) (3a) (3b) (4) (7) (8)

Dual Ascent Framework For any given vector w={w k ij, w k ji} that satisfies constraints (7) of DP, the best v-values are obtained by solving (5)-(6). Doing so, assume the subproblem SP k (w) corresponding to commodity k: max (9) s.t (10) Where The optimal value of this subproblem is the length of the shortest path from origin O(k) to destination D(k).

The ascent strategy Iteratively increase one or more w-values, so that: Constraints (7) remain feasible ( ) The shortest O-D path length v k D(k) increases for at least one commodity k K at each stage. Therefore, the dual ascent procedure seeks to selectively allocate the unabsorbed fixed charges s ij in order to increase the length of the shortest O-D path for one or more commodities at each iteration.

Question How to select the arc(s) whose slack must be allocated and how to allocate this slack to the various commodities? Different arc selection and slack allocation schemes give rise to different implementations of the dualascent method.

Alternative Implementation (1) Path Diversion: changes a single w-value at each iteration. Initialization: all w-values are set to zero, hence v k D(k) for all k K is the length of the shortest path from O(k) to D(k), using c k ij and c k ji as arc lengths. Ascent Iterations: Select a directed arc (i, j) and commodity k satisfying s ij > 0 & l k ij > v k D(k) Increase w k ij & v k D(k) by min{s ij, [l k ij-v k D(k)]}.

A point Selecting more arcs offers the potential for reducing the operating (or routing) costs at the expense of higher fixed costs. On the other hand, with fewer arcs in the design, the fixed costs are lower but the routing costs increase. Let l k ij denote the length of the shortest O-D path excluding arc (i, j) for each directed arc (i, j) and commodity k. Then l k ij > v k D(k)

Alternative Implementation (2) Labeling Method: simultaneously increases w-values that correspond to several arcs. Initialization: {i, j} A and k K, w k ij=0 & w k ji=0. {i, j} A, s ij =F ij. i N and k K, v k i=shortest path length from O(k) to node i. k K, N 1 (k)=n \{D(k)}, N 2 (k)={d(k)}. z D = k v k D(k). Set CANDIDATES = {k K :O(k) N 1 (k)}.

Ascent Iterations A. calculate the amount of w-increase (Step 1) Select a commodity k CANDIDATES Set, A(k)={(i, j) A: i N 1 (k), j N 2 (k)} A (k)={(i, j): c k ij + w k ij - (v k j-v k i) = 0, (i, j) A(k)} δ 1 =min{s ij : (i, j) A (k)} δ 2 =min{c k ij + w k ij v k j + v k i: (i, j) A (k) =A(k)\A (k)} δ=min{δ 1, δ 2 }

Ascent Iterations B. update relevant w-values, slacks and shortest path lengths: (i, j) A (k), w k ij=w k ij+δ & s ij = s ij - δ l N 2 (k), z D =z D + δ. v k l = v k l + δ C. label a new node: If δ=δ 1, for some (i*, j*) A (k) satisfying s i*j* = 0, set N 1 (k)=n 1 (k)\{i*} N 2 (k)=n 2 (k) {i*}. Remove commodity k from CANDIDATES and repeat Step 1.

Stopping Rule k K, if O(k) N 2 (k), then STOP. Otherwise, set CANDIDATES={k K: O(k) N 1 (k)}, and return to Step 1. When the algorithms terminates, that is, when all origins are labeled, any further increases in w-values, with respect to the final dual solution, cannot improve the final lower bound z D.

Does Dual Ascent Heuristic Cover Every Problem? Dual-based procedures are applicable only if the original problem has a special structure. Hence, other lower bounding techniques can tackle a wider spectrum of problems. How can we modify the method to solve a wider range of problems? By modifying the mathematical formulation of the considered problems. By revising the ascent strategy.