Reduction of genetic drift in population-based incremental learning via entropy regularization
Ryoki Hamano, Shinichi Shirakawa · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2022
The Population-Based Incremental Learning (PBIL) is a Bernoulli distribution-based evolutionary algorithm for binary black-box optimization. The PBIL updates the distribution parameter according to the samples generated from the current distribution and their rankings. In PBIL, when some distribution parameters are continuously updated randomly, undesirable convergence without sufficient exploration is observed. This behavior is called genetic drift and induces an increasing number of function evaluations and convergence to local optima. In particular, large update strength leads to genetic drift while faster search. The ways to deal with genetic drift are limited, such as decreasing the update strength, and there is a trade-off between search efficiency and stability. This paper proposes a method to reduce genetic drift in PBIL based on entropy regularization widely used in reinforcement learning. We introduce entropy regularization into PBIL as a penalty term or constraint. The experimental results on well-known benchmark problems show that the proposed entropy regularization can efficiently suppress genetic drift, decrease the number of function evaluations, and improve stability.