A Collective and Continuous Rule Learning Framework towards Enhanced Driving Safety and Decision Transparency

Deep reinforcement learning (DRL)-based driving policies enable connected and autonomous vehicles (CAVs) to adapt to complex environments but lack guarantees of safety and decision transparency. For this reason, we propose a hybrid decision-making system that integrates DRL with a learnable, rule-based safety module, wherein the latter offers enhanced driving safety and decision transparency. However, as collision data is limited for a single CAV during the DRL learning process, it becomes necessary for CAVs to collectively learn such a rulebased safety module to accelerate the learning process. Given the potential for conflicting safety rules learned from different CAVs, we propose a two-stage collective rule learning framework for the safety module. In the first stage, the multi-access edge computing (MEC) network aggregates rules from CAVs based on occurrence frequency and generalization capability. In the second stage, CAVs compute rule weights based on local data satisfaction and violations, followed by the MEC network aggregating these weights. The aggregated weights resolve conflicting rules and help filter out insignificant rules to maintain a compact safety module. The safety module makes comprehensive decisions by aggregating all the weighted rules. Simulation results in the highway driving simulator demonstrate that the two-stage collective rule learning framework effectively aggregates learning outcomes from CAVs, and the proposed hybrid decision-making system further reduces collisions compared to the existing approach.