## A LARGE SCALE INTEGER AND COMBINATORIAL OPTIMIZER (2000)

### BibTeX

@MISC{Chen00alarge,

author = {Qun Chen},

title = {A LARGE SCALE INTEGER AND COMBINATORIAL OPTIMIZER },

year = {2000}

}

### OpenURL

### Abstract

The topic of this thesis, integer and combinatorial optimization, involves minimizing (or maximizing) a function of many variables, some of which belong to a discrete set, subject to constraints. This area has abundant applications in industry. Integer and combinatorial optimization problems are often difficult to solve due to the large and complex set of alternatives. The objective of this thesis is to present an effective solution to integer and combinato-rial problems by using a traditional reliable branch-and-bound approach as well as a newly developed fast adaptive random search method, namely Nested Partitions. The proposed integer and combinatorial optimizer has two closely related components: FATCOP and NP/GA. FATCOP is a distributed mixed integer program solver written in PVM for Condor’s opportunistic environment. It is different from prior parallel branch-and-bound work by implementing a general purpose parallel mixed integer programming algorithm in an opportunistic multiple processor environment, as opposed to a conventional dedi-cated environment. We show how to make effective use of opportunistic resources while

### Citations

2827 |
Adaptation in natural and artificial systems
- Holland
- 1975
(Show Context)
Citation Context ...nerating new chromosomes from this selection, and finally mutation is used to increase variety by making small changes to the genes of a 96sgiven chromosome. The GA was originally proposed by Holland =-=[38]-=- and has been found to be very effective for a variety of combinatorial optimization problems [27]. A detailed algorithm for the GA method is given in Appendix C. While many algorithms for single prod... |

1809 |
Genetic algorithms in search, optimization, and machine learning
- Goldberg
(Show Context)
Citation Context ...orhood if one can obtained through a defined modification of the other. Neighborhood search procedures that are currently popular are simulated annealing [19], tabu search [26] and genetic algorithms =-=[27]-=-. Simulated annealing is a search procedure that mimics the physical process of annealing in an attempt to escape local optima. In simulated annealing moves to worse solutions are 59sallowed. The reas... |

1038 |
Integer and Combinatorial Optimization
- Nemhauser, Wolsey
- 1988
(Show Context)
Citation Context ...e variables, a large number of interesting examples can be effectively modeled using linear relationships along with integer variables. Such problems are typically called mixed integer programs (MIP) =-=[53]-=-. Mixed integer programming (MIP) problems are difficult and commonplace. For many of these hard problems, only small instances can be solved in a reasonable amount of time on sequential computers, re... |

1011 |
Using MPI: Portable Parallel Programming with the Message-Passing Interface, seconde dition
- Gropp, Lusk, et al.
- 1999
(Show Context)
Citation Context ... coordination. PVM transparently handles all message routing, data conversion and task scheduling across a network of incompatible computer architectures. A similar message passing environment is MPI =-=[36]-=-. Both systems center around a message-passing model, providing point-to-point as well as collective communication between distributed processes. The development of resource management systems and mes... |

772 |
The Analytic Hierarchy Process
- Saaty
- 1995
(Show Context)
Citation Context ... this partition. = 3 cells can be assigned processor 1. See Figure 16 for Example 3.2: Partitions for Weights Recovering Problem in Analytical Hierarchy Process The Analytical Hierarchy Process (AHP) =-=[58]-=- is one of the most 68sinfluential methods in multiple criteria decision making. It involves the following optimization problem to recover weights from a given preference matrix A = (aij). subject to:... |

341 |
AMPL: a modeling language for mathematical programming Thomson/Brooks/Cole
- Fourer, Gay, et al.
- 2003
(Show Context)
Citation Context ... multiple LP solvers (CPLEX, OSL and SOPLEX); • advanced MIP techniques including pseudocost estimation searching, preprocessing, and cutting plane generation; • the ability to process MPS [51], AMPL =-=[22]-=- and GAMS [7] models; • object oriented design We begin in Section 2.1 with a review of the standard MIP algorithm components of FATCOP that are implemented to ensure the branch-and-bound algorithm ge... |

248 |
Simulated Annealing and Boltzmann Machines: A Stochastic Approach to Combinatorial Optimization and Neural Computing
- Aarts, Korst
- 1989
(Show Context)
Citation Context ...the current solution. Two solutions are neighborhood if one can obtained through a defined modification of the other. Neighborhood search procedures that are currently popular are simulated annealing =-=[19]-=-, tabu search [26] and genetic algorithms [27]. Simulated annealing is a search procedure that mimics the physical process of annealing in an attempt to escape local optima. In simulated annealing mov... |

169 | An automatic method of solving discrete programming problems. Econometrica
- Land, Doig
- 1960
(Show Context)
Citation Context ...P/GA procedure Report best solution found Stop Figure 1: The procedure to apply the proposed optimizer 1.1 The Branch-and-Bound Method The branch-and-bound method was first described by Land and Doig =-=[46]-=-, and further developed by many researchers. We will describe the branch-and-bound algorithms for MIP in Chapter 2. Here we give a brief introduction to the general branch-and-bound method. The branch... |

90 | Analyzing Scalability of Parallel Algorithms and Architectures - Kumar, Gupta - 1991 |

76 |
et.al. PVM: Parallel Virtual Machine A Users- Guide and Tutorial for Networked Parallel Computing
- Geist
- 1994
(Show Context)
Citation Context ...highly applicable in the future. A parallel virtual machine (PVM) is a programming environment that allows a heterogeneous network of computers to appear as a single concurrent computational resource =-=[23]-=-. It provides a unified framework within which parallel programs for a heterogeneous collection of machines can be developed in an efficient manner. However PVM is not sufficient to develop an efficie... |

75 |
Parallel branch-and-bound algorithms: survey and synthesis
- Gendron, Crainic
- 1994
(Show Context)
Citation Context ...ten in PVM, that runs in the opportunistic computing environment provided by the Condor resource 8smanagement system. Parallel branch-and-bound algorithms for MIP have attracted many researchers (see =-=[18, 25, 55]-=- and references therein). Most parallel branch-and-bound programs were developed for large centralized mainframes or supercomputers that are typically very expensive. Users of these facilities usually... |

63 | Optimized regional caching for on-demand data delivery - Eager, Ferris, et al. - 1999 |

54 | Broader Product Line: A Necessity of Achieve Success - Kekre, Srinivasan - 1990 |

51 | A computational study of search strategies for mixed integer programming
- Linderoth, Savelsbergh
- 1999
(Show Context)
Citation Context ...teresting [1]. Pseudocost attempts to estimate the change rate in objective function value associated with a particular branch. Since the pseudocost method is widely used 14sand known to be efficient =-=[47]-=-, we set it as the default branching strategy and briefly describe the method here. We associate two quantities φ − j and φ + j with each integer variable xj that attempts to measure the per unit decr... |

45 | Interfacing Condor and PVM to Harness the Cycles of Workstation Clusters
- Pruyne, Livny
- 1996
(Show Context)
Citation Context ...ing environments have been independent of each other for many years. Researchers at the University of Wisconsin have recently developed a parallel programming framework that interfaces Condor and PVM =-=[56]-=-. The reason to select PVM instead of MPI is that the implementation of MPI has no concept of process control, hence cannot handle resource addition and retreat in a opportunistic environment. Figure ... |

45 |
Nested partitioned method for global optimization
- Shi, Ólafsson
- 2000
(Show Context)
Citation Context ...l heuristic and is not guaranteed to converge to a global optimum. The NP method, on the other hand, maintains a global perspective in every iteration and converges in finite time to a global optimum =-=[64]-=-. The NP/GA algorithm retains the global perspective of the NP and the local search capabilities of the genetic algorithm. We proposed the following procedure to use the new optimizer for any applicab... |

36 |
Parallel branch-and-bound algorithms for general mixed integer programming on the CM-5
- Eckstein
- 1994
(Show Context)
Citation Context ...ten in PVM, that runs in the opportunistic computing environment provided by the Condor resource 8smanagement system. Parallel branch-and-bound algorithms for MIP have attracted many researchers (see =-=[18, 25, 55]-=- and references therein). Most parallel branch-and-bound programs were developed for large centralized mainframes or supercomputers that are typically very expensive. Users of these facilities usually... |

33 |
Preprocessing and probing for mixed integer programming problems
- Savelsbergh
- 1994
(Show Context)
Citation Context ...s. In linear programming this typically leads to problem size reductions. FATCOP identifies infeasibilities and redundancies, tighten bounds on variables, and improves the coefficients of constraints =-=[60]-=-. At the root node, FATCOP analyzes every row of the constraint matrix. If, after processing, some variables are fixed or some bounds are improved, the process is repeated until no further model reduc... |

30 |
Solving large scale zero-one linear programming problems
- Crowder, Johnson
- 1963
(Show Context)
Citation Context ...ay mean that the linear subproblem can be fathomed without having to resort to branching. There are many different classes of cutting planes. FATCOP includes two classes – knapsack cover inequalities =-=[14]-=- and flow cover inequalities [57]. Knapsack covers and flow covers inequalities are derived from structures that are present in many, but not all, MIP instances. This implies that for some instances, ... |

27 | FATCOP: A fault tolerant Condor-PVM mixed integer program solver
- Chen, Ferris
(Show Context)
Citation Context ...of result messages queued at the master. The node utilization corresponding to grain size of 100 seconds is satisfactory. Increasing grain size does not improve node utilization further. As stated in =-=[9]-=-, all Condor-PVM programs risk losing the results of their work if a worker is suspended or deleted from the virtual machine. Taking this into consideration, we prefer a smaller worker grain size so t... |

27 |
Computer Solutions of Linear Programs
- Nazareth
- 1987
(Show Context)
Citation Context ... the use of multiple LP solvers (CPLEX, OSL and SOPLEX); • advanced MIP techniques including pseudocost estimation searching, preprocessing, and cutting plane generation; • the ability to process MPS =-=[51]-=-, AMPL [22] and GAMS [7] models; • object oriented design We begin in Section 2.1 with a review of the standard MIP algorithm components of FATCOP that are implemented to ensure the branch-and-bound a... |

25 | Metacomputing and the masterworker paradigm
- Goux, Linderoth, et al.
- 2000
(Show Context)
Citation Context ... specific PVM and Condor primitives. Many of the features, and several extensions, of the resource management and communication procedures in FATCOP have been incorporated into a new software API, MW =-=[29]-=-, that can be used for any master-worker algorithm. Since this abstraction shields all the platform specific details from an application code, FATCOP was designed to use this API, resulting in a much ... |

24 | Genetic algorithms for flowshop scheduling problems - MURATA, ISHIBUCHI, et al. - 1996 |

20 | Cutting planes for integer programs with general integer variables - Ceria, Cordier, et al. - 1998 |

19 | Experiments in mixed-integer linear programming - Bénichou, Gauthier, et al. - 1971 |

17 |
et al., “Condor - a hunter of idle workstations
- Litzkow
- 1988
(Show Context)
Citation Context ...ces in the machine will be “occupied” during the life cycle of the program. This is not a desirable situation when the machine is owned by a 9sperson different from the user of the MIP solver. Condor =-=[20, 48]-=- is a distributed resource management system that can help to overcome these problems. Condor manages large heterogeneous clusters of machines in an attempt to use the idle cycles of some users’ machi... |

16 |
Search: A Tutorial,” Interfaces
- Glover, “Tabu
- 1990
(Show Context)
Citation Context ...on. Two solutions are neighborhood if one can obtained through a defined modification of the other. Neighborhood search procedures that are currently popular are simulated annealing [19], tabu search =-=[26]-=- and genetic algorithms [27]. Simulated annealing is a search procedure that mimics the physical process of annealing in an attempt to escape local optima. In simulated annealing moves to worse soluti... |

15 |
Heuristics for product-line design using conjoint analysis. Management Sci. 36(12) 1464–1478. Pope, and Jain: Pricing Digital Content Product Lines: A Model and Application for the NAP 636 Marketing Science 28(4
- Kohli, Sukumar
- 1990
(Show Context)
Citation Context ...ttribute is determined independently of all other attributes. A detailed algorithm for the GS heuristic is given in Appendix A. The DP heuristic has been used for a variety of product design problems =-=[42, 43, 44]-=-. This heuristic treats attributes as stages in a dynamic program and the levels as states. Thus similarly to the GS heuristic, in each iteration the DP heuristic sets one attribute to its ‘best’ leve... |

12 |
Genetic Algorithms for Product Design
- Balakrishnan, Jacob
- 1996
(Show Context)
Citation Context ... problems have received considerable attention in the literature. These problems may be divided into single product design problems where the objective is to design the attributes of a single product =-=[43, 3]-=-, and product line design where multiple products are offered simultaneously [31, 44, 16, 50]. The product design problem involves determining the levels of the attributes of a new or redesigned produ... |

12 | Optimal equi-partition of rectangular domains for parallel computation
- Christou, Meyer
- 1996
(Show Context)
Citation Context ...t of as partitioning the grid into P equi-area regions(each of area A := MN/P) of minimum total perimeter. The problem is illustrated in Figure 3.1.2.1. An Integer Programming formulation is given in =-=[12]-=-. s.t.: min M� N� P� i,i ′ =1 j,j ′ =1 p,p ′ =1,p�=p ′ M� N� i=1 j=1 P� p=1 x p ij = MN P xiji ′ j ′xp ijx p′ i ′ j ′ 66 (7) (8) x p ij = 1 (9) x p ij ∈ B = 0, 1 (10) where ciji ′ ⎧ 1 if ⎪⎨ j ′ = ⎪⎩ |... |

12 |
Optimal Product Design using Conjoint Analysis: Computational Complexity and Algorithms
- Kohli, Krishnamurti
- 1989
(Show Context)
Citation Context ... problems have received considerable attention in the literature. These problems may be divided into single product design problems where the objective is to design the attributes of a single product =-=[43, 3]-=-, and product line design where multiple products are offered simultaneously [31, 44, 16, 50]. The product design problem involves determining the levels of the attributes of a new or redesigned produ... |

11 | Parallel machine scheduling problems using memetic algorithms - Cheng, Gen - 1997 |

11 | Ordinal optimization of DEDS. Discrete Event Dynamic Systems: Theory and Applications 2 - Ho, Sreenivas, et al. - 1992 |

10 |
Models and heuristics for product line selection. Marketing Sci
- Green, Krieger
- 1985
(Show Context)
Citation Context ...ncy Speed In this section we compare convergency speed of the pure NP algorithm and the NP/GA algorithm, and test the stopping criteria proposed in this Chapter. We select the product design problems =-=[30, 31, 32]-=- as our test problems. We shall discuss this class of problems in more detail in the next Chapter. We select three small problems (P1, P2 and P3), the optimal solutions of which can be found through e... |

10 |
Branch-and-bound and parallel computation: A historical note
- Pruul, Nemhauser
- 1988
(Show Context)
Citation Context ...ten in PVM, that runs in the opportunistic computing environment provided by the Condor resource 8smanagement system. Parallel branch-and-bound algorithms for MIP have attracted many researchers (see =-=[18, 25, 55]-=- and references therein). Most parallel branch-and-bound programs were developed for large centralized mainframes or supercomputers that are typically very expensive. Users of these facilities usually... |

9 |
Heuristics for pricing and positioning a product line using conjoint and cost data
- Dobson, Kalish
- 1993
(Show Context)
Citation Context ... be divided into single product design problems where the objective is to design the attributes of a single product [43, 3], and product line design where multiple products are offered simultaneously =-=[31, 44, 16, 50]-=-. The product design problem involves determining the levels of the attributes of a new or redesigned product in such a way that it maximizes a given objective function. We assume that the part-worths... |

9 | New Parallel Randomized Algorithm for Traveling Salesman Problem
- Shi, Olafsson, et al.
- 1999
(Show Context)
Citation Context ...spective and offers a framework to incorporate various kinds of local search techniques. The NP method has been found to be efficient for a variety of problems, such as the traveling salesman problem =-=[61]-=-, production scheduling problems [54] and a product design problem [62]. Furthermore, the NP method does not have any requirement on the problem’s objective function, hence can be used to solve nonlin... |

8 |
Recent Contributions to Optimal Product Positioning and Buyer Segmentation,” European
- Green, Krieger
- 1989
(Show Context)
Citation Context ...ncy Speed In this section we compare convergency speed of the pure NP algorithm and the NP/GA algorithm, and test the stopping criteria proposed in this Chapter. We select the product design problems =-=[30, 31, 32]-=- as our test problems. We shall discuss this class of problems in more detail in the next Chapter. We select three small problems (P1, P2 and P3), the optimal solutions of which can be found through e... |

6 |
A Heuristic Approach to Product Design
- Kohli, Krishnamurti
- 1987
(Show Context)
Citation Context ...ttribute is determined independently of all other attributes. A detailed algorithm for the GS heuristic is given in Appendix A. The DP heuristic has been used for a variety of product design problems =-=[42, 43, 44]-=-. This heuristic treats attributes as stages in a dynamic program and the levels as states. Thus similarly to the GS heuristic, in each iteration the DP heuristic sets one attribute to its ‘best’ leve... |

5 | Hybrid evolutionary method for capacitated location-allocation problem - Gong, Gen, et al. - 1997 |

4 |
A Method for Scheduling in Parallel Manufacturing Systems with Flexible Resources
- Ólafsson, Shi
- 2000
(Show Context)
Citation Context ...corporate various kinds of local search techniques. The NP method has been found to be efficient for a variety of problems, such as the traveling salesman problem [61], production scheduling problems =-=[54]-=- and a product design problem [62]. Furthermore, the NP method does not have any requirement on the problem’s objective function, hence can be used to solve nonlinear problems. In this thesis, we deve... |

4 |
Solving mixed integer 0–1 programs by automatic reformulation
- Roy, Wolsey
- 1987
(Show Context)
Citation Context ...m can be fathomed without having to resort to branching. There are many different classes of cutting planes. FATCOP includes two classes – knapsack cover inequalities [14] and flow cover inequalities =-=[57]-=-. Knapsack covers and flow covers inequalities are derived from structures that are present in many, but not all, MIP instances. This implies that for some instances, FATCOP will be able to generate u... |

3 |
a Mixed INTeger
- MINTO
- 1994
(Show Context)
Citation Context ...11sFigure 2: Branch-and-bound algorithm where Z denotes the integers, A is an m×n matrix, and I is a set of distinguished indices identifying the integer variables. Most integer programming textbooks =-=[1, 52]-=- describe the fundamental branch-andbound algorithm for the above MIP problem. Basically, the method explores a binary tree of subproblems. The key steps of BB are summarized below and the algorithm f... |

3 | A new hybrid optimization algorithm - Shi, Olafsson, et al. - 1999 |

2 |
Mathematical Programming for Industrial Engineers
- Avriel, Golany
- 1996
(Show Context)
Citation Context ...11sFigure 2: Branch-and-bound algorithm where Z denotes the integers, A is an m×n matrix, and I is a set of distinguished indices identifying the integer variables. Most integer programming textbooks =-=[1, 52]-=- describe the fundamental branch-andbound algorithm for the above MIP problem. Basically, the method explores a binary tree of subproblems. The key steps of BB are summarized below and the algorithm f... |

2 | Report on the session on branch and bound/implicit enumeration, in discrete optimization - Balas, Martin - 1979 |

2 |
XPRESS-MP User Guide. http://www.dashopt.com
- Associates, House, et al.
(Show Context)
Citation Context ...extension would undoubtedly benefit other MW applications such as those outlined in [29]. SOPLEX is an effective linear programming code, but commercial codes such as CPLEX [40], OSL [39], and XPRESS =-=[15]-=- significantly outperform SOPLEX for solving the LP problem relaxations. In many cases, several copies of these solvers are available to a user of FATCOP and so we design the code to allow a variety o... |

2 | et al. Practical solution of large scale mixed integer programming problems with UMPIRE - Forrest - 1979 |

2 |
A simple heuristic for selecting ’good’ products in conjoint analysis
- Green, Krieger
- 1987
(Show Context)
Citation Context ...we can select the best one from this set. A detailed algorithm for the DP heuristic is given in Appendix B. The DC heuristic was proposed for single product design and sequential product lines design =-=[33]-=-. Variants of the DC were later employed by SIMOPT, a conjoint analysis oriented, product positioning model [35]. One version of the DC method can be implemented as follows. Suppose there are K attrib... |

2 |
Conjoint analysis in consumer research: new developments and directions
- Green, Srinivasan
- 1990
(Show Context)
Citation Context ...on. We assume that the part-worths preferences of individual customers, or market segment, have been elicited for each level of every attribute, for example using conjoint or hybrid conjoint analysis =-=[66, 35]-=-. These part-worths preferences are assumed to be independent, but by a simple modification of the objective function dependencies can also be accounted for in the new framework. Furthermore, we assum... |