Selecting Deep Neuroevolution hyperparameters for reinforcement learning: a challenging task

• Autores: Benito, Rodrigo; Avila, Luis; Cagnina, Leticia Cecilia
• Año de publicación: 2025
• Idioma: inglés
• Tipo de recurso: documento de conferencia
• Materia: Ciencias Informáticas; Neuroevolution; Reinforcement Learning; Hyperparameters setting; Atari game; Deep Neural Networks
• Institución: Universidad Nacional de La Plata

Deep Neuroevolution (deep NE) has emerged as a promising alternative to gradient-based optimization in reinforcement learning, offering greater flexibility in handling sparse rewards or non-differentiable operations. However, the performance of these evolutionary approaches is highly sensitive to the selection of hyperparameters, which govern aspects such as mutation probability, population size, and parents selection, among others. In this work, we investigate the challenges associated with hyperparameter selection in deep NE for reinforcement learning-based tasks. Using the Atari Learning Environment as our proving ground, we analyze the impact of hyperparameters variation on agents performance. Employing Latin Hypercube Sampling to explore the high-dimensional parameter space, we evaluate a population of agents over a large number of episodes. We conclude that significant performance variability is directly associated to NE hyperparameter configurations.
Red de Universidades con Carreras en Informática