Explainable Brain-Like Decision-Making for Autonomous Vehicles Trustworthiness Improvement

Explainability is indispensable for safety-critical systems like autonomous vehicles (AVs), particularly in accident scenarios where in-depth clarification of their decision processes is mandatory, and it is still a great challenge to apply explainable AI to these high-dimensional traffic environments. This paper proposes a hierarchical logical reasoning method, Situation-Aware Reinforcement Learning (SARL), which tries to enhance decision-making explainability for AVs. The SARL incorporates logical reasoning modeled via the high-level situation awareness and intuitive reasoning simulated via neural network to express the brain-like human hierarchical decision-making mechanism. To improve the explainability of the intuitive reasoning process, the observation features’ contributions to intuitive reasoning are analyzed, making the decision results of intuitive reasoning easier to be understood. A case study illustrates that the explainable SARL framework outperforms the baseline in terms of lower collision rate and higher episode mean reward with reasonable convergence speed, showing the potential of SARL in enhancing the performance and explainability of time-sequential decision-making for AVs.