Bayesian Optimization of Hyperparameters for Rainbow DQN in the CartPole-v1 Environment

Abstract

This paper presents a Bayesian optimization approach to hyperparameter tuning for the Rainbow DQN reinforcement learning algorithm, using the Hyperopt library and the CartPole-v1 environment as a benchmark. The study investigates the impact of search space definition on the convergence and quality of optimized hyperparameters. Furthermore, it analyzes the effectiveness of different evaluation methods in the context of hyperparameter optimization for deep reinforcement learning. Results demonstrate the efficacy of Bayesian optimization in identifying high-performing hyperparameter configurations for Rainbow DQN in this control task.