Analysis of population size in the accuracy and performance of genetic training for rule-based control systems

Abstract In off-line training of a rule-based controller, the significant measure of successful training is the quality of control provided by the generated rule set. In an adaptive or on-line control environment, performance is also measured by the ability to accurately maintain a satisfactory rule set, but within constraints on speed and/or resource availability. Very small population genetic algorithms, or microGAs, have been proposed as a means of capitalizing on the hill climbing characteristics of faster local optimizatiion techniques while requiring less memory and retaining much of the robustness of traditional, larger population genetic search. A traditional genetic algorithm and a similar microGA are developed and applied to two control problems. The performance of these algorithms is analyzed with respect to (1) the quality of the rules learned, (2) the rate at which learning occurs, and (3) the memory resources required during learning.