A Comparative Study of DE, GA and ES for Evolutionary Reinforcement Learning of Neural Networks in Pendulum Task

Reinforcement learning of neural networks requires gradient-free algorithms as labeled training data are not available. Evolutionary algorithms are well-suited for this purpose since they do not rely on gradients. However, the success of training neural networks with evolutionary algorithms is contingent on the careful selection of appropriate algorithms, given the numerous algorithmic variations available. In this study, the author evaluates the efficacy of Differential Evolution (DE), Genetic Algorithm (GA), and Evolution Strategy (ES) for the reinforcement learning of neural networks, utilizing a pendulum control task. The experimental results indicate that DE exhibits statistically significant superiority over GA and ES. While GA performs better than ES, this difference is not statistically significant. The study highlights DE's ability to effectively balance between exploratory and exploitative search, adapting to the problem at hand. Based on these findings, it is suggested that an algorithm possessing such characteristics is better suited for evolutionary reinforcement learning of neural networks.