We present a simple scheme for the automated, iterative specification of the genetic mutation, crossover, and reproduction (usage) probabilities during run time for a specific genetic algorithm-driven tool. The tool is intended for supporting static scheduling decisions in flexible manufacturing systems. Using a randomly generated (base) test problem instance, we first assess the method by using it to determine the appropriate levels for specific types of mutation and crossover operators. The level for the third operator, reproduction, may then be inferred. We next report on its ability to choose one or more appropriate crossovers from a set of many such operators. Finally, we compare the method's performance with that of approaches suggested in prior research for the base problem and a number of other test problem.Our experimental findings within the specific scheduling domain studied suggest that the simple method could potentially be a valuable addition to any genetic algorithmbased decision support tool. It is, therefore, worthy of additional investigations.
Subject Areas.: MISDSS and Simulation.*We extend our sincere thanks to the two anonymous referees. an associate editor, the assistant editor, and the editor for their invaluable suggestions for improving this manuscript.
749Pakath and Zaveri 75 1 p,., may then be inferred.) In the remainder of this section, we discuss and contrast these efforts.DeJong [4] addressed the problem in the context of GAdriven systems with binaryencoded populations and the following binary genetic operators: simple mutation, simple (single-point) crossover, and multipoint crossover. He set up a test suite of five function optimization problems intended to simulate every conceivable optimization problem that could be modeled using such encoding. He defined two measures of system performance: off-line and on-line. At any iteration, T, the off-line performance is the average of the best-so-far answers identified at each of the iterations 1,2, ..., T. The on-line performance, on the other hand, is the average of all answers identified up to and including iteration T.Based on exhaustive, enumerative testing, DeJong suggested the following values for an acceptable compromise between good on-line and off-line performance: population size, N40;pc4.6; pm4.001; generation gap, G=l. Here, the population size refers to the number of solutions simultaneously under consideration at any iteration. The generation gap determines what proportion of the current population would be replaced with new members from one iteration to the next. In general, N*G members must be replaced, where O
