Interactive testing for evaluating the performance of collision avoidance algorithms in multi-ship encounter scenarios

Cooperative and adversarial among different traffic participants in multi-ship encounter scenarios is difficult to predict, leading to a lack of robustness in maritime collision avoidance algorithms. In this paper, an interactive testing method is proposed for the performance evaluation of collision avoidance algorithms in multi-ship encounter scenarios. First, a theoretical model for a multi-ship encounter test scenario is proposed based on the decision-making process of ship collision avoidance. Subsequently, the initial state parameter of the test scenarios is generated through the combinatorial testing strategy and filtered using multi-ship encounter event coverage. Finally, the collision avoidance process is modeled as a multi-agent Markov game process, and the multi-agent deep deterministic policy gradient method is used to update the scenarios. By leveraging dynamic game reward functions, the interactive agents are trained to push test scenarios to risk states and methodically explore operational limits under extreme conditions. The performance boundaries and weaknesses of the tested algorithm were identified and evaluated in multi-ship encounter scenarios involving typical crossing channels. The proposed method can self-adjust based on the characteristics of the algorithm and is suitable for simulating complex encounter scenarios, which indicates its potential for evaluating collision avoidance algorithms. The core contribution lies in applying multi-agent reinforcement learning to train intelligent agents as expert testers that actively discover worst-case scenarios for collision avoidance algorithms, shifting from passive scenario-based testing to active adversarial search for failure points.