Learnable evolution model performance impaired by binary tournament survival selection

A Cultural Algorithm (CA) is an evolutionary algorithm augmented by a machine learner. The machine learner updates a knowledgebase with each generation based on bes and least fit individuals. Using a stochastic selection operator may eliminate critical "best"and "worst"individuals via genetic drift thus radically changing the inferred rules. The new rules may be so different from the previous generation as to switch the convergence trajectory to a different optimum. Overall performance may be severely impaired by this optima "thrashing"from generation to generation. This research demonstrates this phenomena using a variant of a CA, Learnable Evolution Model, on a simple problem. The performance between truncation survival selection and binary tournament survival selection operators is compared. It was expected that the latter survival selection operator would introduce pathological effects due to genetic drift because of its stochastic nature. Comparatively binary tournament selection converged more slowly, had larger convergence variations, and converged to inferior solutions than truncation selection. Practitioners that use evolutionary algorithm/machine learner hybrids need to be aware of this problem. Selection operators will have to be judiciously chosen accordingly to avoid deleterious effects of genetic drift on the machine learning component.