Calibration of GA Parameters

Abstract

This chapter focuses on a critical dilemma faced in many GA applications: the optimum selection of the different GA parameters to ensure the GA’s rapid convergence—both on-line and off-line (Section 2.9). We address this task early in this text because the satisfactory resolution of it can either make or break the efficacy of the GA, for there is no single makeup of a GA that can uniformly solve all global search problems. In this respect the GA is unlike, for instance, the Newton-Raphson method which finds zeros of almost any polynomial. In particular, the stochastic process that develops as the GA is executing is noted to be affected by the values chosen for the elite fraction (ε) participating in reproduction, crossover probability (p_c), mutation probability (p_m), population size (ps), etc. The GA’s convergence may be impacted also by the interaction among these parameters. However, no general methodology is presently available to optimize the selection of these parameters. What is even more troublesome is the growing evidence that such “optimum” parameter values may be problem-specific. This chapter presents a robust parameterization procedure based on the statistical design of experiments (DOE) approach (Bagchi and Deb, 1996). A multi-factor constrained optimization problem with known solution is used to illustrate the steps in this proposed method.