Enhancing Generalization in Sparse Reward Environments: A Fusion of Reinforcement Learning and Genetic Algorithms

This paper investigates the viability of merging attributes of reinforcement learning and genetic algorithms to enhance agent learning in reward-sparse environments. The primary challenge lies in the inefficiency of reinforcement learning and genetic algorithms to adapt and excel in reward-sparse environments. This research addresses this challenge by proposing an adaptive learning framework that merges the strengths of both approaches. The proposed framework combines convolutional neural networks and DDQNs with genetic algorithms to evolve and optimize agents for a specific video game, likely Mario. The reinforcement learning component trains neural network-based agents to play the game, while the genetic algorithm component evolves a population of agents with varying strategies and characteristics to further enhance performance. The framework orchestrates the training process through generations, tracking fitness scores, utilizing replay memory, and employing various strategies like mutation and crossover to refine the population. Additionally, the project employs visualization techniques to analyze the learning process.The results indicate that integrating Reinforcement Learning (RL) with Genetic Algorithms (GAs) leads to rapid fitness improvement in evolving agents, followed by a stabilization phase, with population size and time affecting the exploration-exploitation balance, and late-stage variations potentially breaking local optima. The proposed framework offers insights into hybrid learning dynamics and could be generalized to other domains beyond gaming.