Statistics & OR: The Interface

SE04.1 Algorithm Fine-Tuning with OptQuest

• James P. Kelly; University of Colorado, Coll. of Bus., Boulder, CO 80309; james.kelly@colorado.edu
• Fred Glover; University of Colorado, Coll. of Bus., Boulder, CO 80309-0419; fred.glover@colorado.edu
• Manuel Laguna; University of Colorado, Coll. of Bus., CB 419, Boulder, CO 80309; manuel.laguna@colorado.edu

Many algorithms require fine-tuning to optimize their performance. We discuss a technique based on scatter search and TS called OptQuest that can automate this time-consuming process. OptQuest often reveals parameter sets that are superior to those derived from manual explorations. We compare experimental design results to those obtained using OptQuest.

SE04.2 Fractional Factorial & Incomplete Block Designs for Algorithm Analysis

• M. Coffin; Clemson University, Math. Sci. Dept., Clemson, SC 29634-1907; mcoffin@clemson.edu
• M. J. Saltzman; Clemson University, Math. Science Dept., Clemson, SC 29634-1907; mjs@clemson.edu

Incomplete block and fractional factorial designs are powerful statistical tools for comparing algorithm performance under a variety of experimental conditions. Several examples are provided to demonstrate how these and other experimental design ideas can be used to efficiently compare algorithms and heuristics.

SE04.3 Using Statistical Experimental Design Principles for Empirical Comparison of Network Optimization Software

• Mohammad M. Amini; University of Memphis, MIS/DIS Dept., Fogelman Coll. of Bus., Memphis, TN 38152; mamini@cc.memphis.edu
• Richard S. Barr; SMU, Dept. of Comp. Sci. & Eng., PO Box 750122, Dallas, TX 75275-0122; barr@seas.smu.edu

The use of formal experimental designs for the empirical evaluation and comparision of algorithms and their implementations remains a rarity in the OR literature. We present a mini-tutorial and examples of these statistical techniques to assessing the relative efficiency of 5 network codes for reoptimizing pure network problems.

SE04.4 An Experimental Design-Based Method for Finding Effective Parameter Settings for Heuristic Methods

• Steven P. Coy; University of Maryland, Maryland Bus. Sch., College Park, MD 20742; scoy@mbs.umd.edu
• Bruce L. Golden; University of Maryland, Maryland Bus. Sch., College Park, MD 20742; bgolden@umdacc.umd.edu
• George C. Runger; Arizona State University, College of Eng. & Appl. Sci., Tempe, AZ 85287; runger@asu.edu
• Edward A. Wasil; American University, Kogod College of Bus. Admin., Washington, DC 20016; ewasil@american.edu

We propose a procedure that uses experimental design to find high-quality heuristic parameter values. We illustrate how to apply our method in 2 experiments using 2 vehicle routing heuristics. In each experiment, we fine-tune 1 of the heuristics using experimental design on a small number of problems.