A Bayesian network approach to cognitive alignment and trust in human-unmanned ship collision avoidance
Abstract
To address the cognitive differences between the human intelligence of manned ships and the artificial intelligence of unmanned ships in understanding collision avoidance scenarios, a Bayesian network-based cognitive alignment model for unmanned ships is developed. The model aims to simulate the reasoning of human crew members and enhance the comprehension capabilities of unmanned ships for collision avoidance situations. To equip the model with rich prior knowledge, the proposed approach incorporates navigational habits and the International Regulations for Preventing Collisions at Sea (COLREGs) to classify navigation situations and construct a situational understanding model. To establish a human-like reasoning chain for unmanned ships, a three-layer inference architecture based on Bayesian networks including navigational situations recognition, responsibility assessment and action generation is designed to realize the judgment of situational awareness. The effectiveness of the model is validated through experiments involving four representative navigation scenarios. Participants are asked to independently assess the navigation situations, and their judgments are compared with the model's outputs to evaluate its accuracy. Building on this, a human-machine trust comparison experiment is conducted to examine the level of trust generated by the model during task execution. The Mann-Whitney U test is used to analyze the participants' trust ratings. Experimental results indicate that the model is capable of effectively simulating the thinking process of human crew members and gaining their trust. Based on these findings, the paper further explores the influence and challenges of human-machine trust in negotiating collision avoidance decisions. Nevertheless, the current model has limitations in handling highly dynamic and complex navigational environments, which will be the focus of future enhancements.