Hyperparameter Optimisation of Reinforcement Learning Algorithms in Webots Simulation Environment
László Schäffer, Zoltán Kincses, Szilveszter Pletl · 2023
Reinforcement learning (RL) has shown great potential for solving complex mobile robotic tasks. However, developing RL algorithms that can perform effectively on a variety of robotic systems and environments can be challenging. In this paper a RL hyperparameter optimisation framework is presented, which designed to be flexible, scalable, and adaptable to various mobile robot scenarios. The proposed framework is built on top of the Webots simulation environment, which provides a realistic 3D physics engine for simulating different mobile robots. Stablebaselines3 is utilized, which is a popular RL library, to train and evaluate RL agents. Moreover, to address the challenge of hyperparameter tuning, a genetic algorithm-based hyperparameter optimization technique is incorporated to automatically tune the hyperparameters of the RL agents. The effectiveness of the proposed framework is demonstrated on two different mobile robot scenarios, including inverted pendulum balancing, and target tracking. The results show that the proposed framework can be used and achieves state-of-the-art performance in all of the scenarios. The proposed framework can provide a solid foundation for developing and testing RL algorithms on mobile robots and has the potential to accelerate the development of intelligent robotic systems.