An alternative model of genetic algorithms as learning machines

摘要

In this paper we report some results related to the use of genetic algorithms as learning machines (GLM) via a new regeneration/crossover model. In the past GAs have been applied as learning systems in traditional classifier systems. An alternative method to learning is the one where GAs are treated as computing devices where the genetic structure encodes some sort of automaton. That such an approach is theoretically possible results directly from Turing's thesis. We discuss a new model where the standard regeneration mechanism via biased individual survival and random crossover is replaced by a deterministic crossover scheme; furthermore, standard crossover is replaced by ring crossover. We statistically establish that such a genetic process results, indeed, in systems which learn. We also report on the way that crossover probabilities affect the learning convergence of such systems. After a statistical analysis, we found that GLMs, after being trained, show predictive behavior. Finally, we conclude from a statistical analysis, that the probability that a GLM shows a predictive behavior is better that 0.97, proving conclusively their viability as trainable learning systems.