Generational Information Transfer with Neuroevolution on Control Tasks
Maximilien Le Clei, Stav Bar-Sheshet, Pierre Bellec · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2024
Traditional genetic algorithms compute fitness scores every generation for all agents in a population, which typically requires agents to perform a task until they either fail or succeed. These evaluations can turn into a computational bottleneck when tasks are either time-consuming or infinite. A common workaround is to set an expiration time on agent trials, i.e. to evaluate agents on a subset of the task, which can bias the true task objective. We propose to address this bias through a novel information inheritance genetic algorithm where three distinct attributes can be passed down generations: the task environment state (agents resume from where their parents reached task expiration), the memory state (agents inherit internal representations from their parents); and the fitness (ancestors evaluation scores are compounded with an agent's own score for selection). We benchmark the various combinations of these three inherited attributes by running a genetic neuroevolution algorithm on popular feature-based control tasks. We report that information inheritance can lead to substantial increases in both data and runtime efficiency, suggesting it may greatly benefit a variety of genetic algorithm techniques and applications in the future. We provide the relevant source code.