Ensemble Reinforcement Learning in Collision Avoidance to Enhance Decision-Making Reliability

Obstacle avoidance algorithms for Unmanned Surface Vehicles can facilitate missions without human intervention. Typically, these models are trained using machine learning algorithms, such as reinforcement learning. However, despite their demonstrable value, it's evident that disturbances to the model, training data, or the environment can significantly impair the model's performance. Aiming for enhanced decision-making robustness, this research suggests integrating ensemble methods with reinforcement learning-trained agents, employing population voting systems. This study highlights the efficacy of the two proposed voting systems in enhancing the decision-making robustness of USVs for obstacle avoidance.