Balancing samples’ contributions on GA learning

Abstract

A main branch in Evolutionary Computation is learning a system directly from input/output samples without investigating internal behaviors of the system. Input/output samples captured from a real system are usually incomplete, biased and noisy. In order to evolve a precise system, the sample set should include a complete set of samples. Thus, a large number of samples should be used. Fitness functions being used in Evolutionary Algorithms usually based on the matched ratio of samples. Unfortunately, some of these samples may be exactly or semantically duplicated. These duplicated samples cannot be identified simply because we do not know the internal behavior of the system being evolved. This paper proposes a method to overcome this problem by using a dynamic fitness function that incorporates the contribution of each sample in the evolutionary process. Experiments on evolving Finite State Machines with Genetic Algorithms are presented to demonstrate the effect on improving the successful rate and convergent speed of the proposed method. © Springer-VerlagBerlin Heidelberg 2001.