Reinforcement Learning Reward Function Evaluator for USV Straight-Path Following

Constructing efficient reward functions has long been one of the main challenges in the pursuit of artificial intelligence. However, in reward shaping the performance comparison and hyperparameter tuning are often complex and inefficient. In this work, we present a reinforcement learning reward function evaluator for the Unmanned Surface Vehicle (USV) straight-path following which, by combining a tree-based search with a maneuvering prediction method for ships, improves the efficiency of reward function comparison and hyperparameter tuning, without relying on any reinforcement learning model. Furthermore, we leverage this evaluator to shape a dynamic reward function to enhance the performance of the USV straight-path following. When tested with different reward functions, our evaluator efficiently demonstrates the trajectories guided by these reward functions, providing insights into how different reward functions influence performance through visualization and data analysis. Our findings introduce a versatile framework and theoretical guidance for developing and optimizing reward functions in reinforcement learning, particularly for intelligent control systems.