Research on Lane-Changing Decision-Making of Autonomous Vehicles Considering Epistemic Uncertainty

Aiming at the safety issues in lane-changing decision-making for autonomous vehicles in urban environments, this study proposed a new lane-changing decision-making method for autonomous vehicles that considers epistemic uncertainty (EUTAS-DDQN). In the training phase, a Transformer layer was introduced into the Double Deep Q-Network algorithm (DDQN) to extract data features, which guides the fully connected layer in making decisions. Then, a trainable temperature coefficient was incorporated to develop an adaptive Softmax strategy for generating the final action. In the testing phase, the standard deviation of the Q-values was used as an indicator of epistemic uncertainty. An epistemic uncertainty threshold was defined such that when the standard deviation of the agent’s Q-values exceeds the threshold, the agent’s output action was executed; otherwise, the system switched to a rule-based module to execute a corrected action, thereby enhancing safety. Experimental results demonstrate that the proposed EUTAS-DDQN model achieves a collision rate of only 0.75%, representing a 70.47% reduction compared to the DDQN model.