Evolving neural networks through bio-inspired parent selection in dynamic environments

Abstract

Environmental variability often degrades the performance of algorithms designed to capture the global convergence of a given search space. Several approaches have been developed to challenge environmental uncertainty by incorporating biologically inspired notions, focusing on crossover, mutation, and selection. This study proposes a bio-inspired approach called NEAT-HD, which focuses on parent selection based on genetic similarity. The originality of the proposed approach rests on its use of a sigmoid function to accelerate species formation and contribute to population diversity. Experiments on two classic control tasks were performed to demonstrate the performance of the proposed method. The results show that NEAT-HD can dynamically adapt to its environment by forming hybrid individuals originating from genetically distinct parents. Additionally, an increase in diversity within the population was observed due to the formation of hybrids and novel individuals, which has never been observed before. Comparing two tasks, the characteristics of NEAT-HD were improved by appropriately setting the algorithm to include the distribution of genetic distance within the population. Our key finding is the inherent potential of newly formed individuals for robustness against dynamic environments.