Relevance Estimation and Value Calibration of Evolutionary Algorithm Parameters
Volker Nannen, A.E. Eiben
The main objective of this paper is to present and evaluate a method that helps to calibrate the parameters of an evolutionary algorithm in a systematic and semi-automated manner. The method for Relevance Estimation and Value Calibration of EA parameters (REVAC) is empirically evaluated in two different ways. First, we use abstract test cases reflecting the typical properties of EA parameter spaces. Here we observe that REVAC is able to approximate the exact (hand-coded) relevance of parameters and it works robustly with measurement noise that is highly variable and not normally distributed. Second, we use REVAC for calibrating GAs for a number of common objective functions. Here we obtain a common sense validation, REVAC finds mutation rate p_m much more sensitive than crossover rate p_c and it recommends intuitively sound values: p_m between 0.01 and 0.1, and 0.6 <= p_c <= 1.0.