Intelligent Adaptive Decision-Making for Autonomous Vehicles: A Learning-Enhanced Game-Theoretic Approach in Interactive Scenarios

This paper introduces an intelligent adaptive decision-making framework for autonomous vehicles (AVs), specifically for complex interactive merging scenarios. Employing the non-cooperative game theory, we develop a vehicle interaction behavior model that delineates critical traffic elements and the reward function. Subsequently, a parameter optimization method rooted in maximum entropy inverse reinforcement learning (IRL) is deployed to ensure the model’s adaptability to a dynamic traffic environment. Optimal parameters matching across varied settings are derived via the designed interaction behavior feature vector, coupled with interaction probability. Additionally, a dynamic environmental-adaptive behavioral decision-making strategy is proposed. By forging a mapping model between environmental variables and model parameters, enables the real-time identification of AVs’ interactive behavior probabilities. Empirical assessments employing naturalistic driving datasets (highD and exiD) verify the performance of our method with human behaviors. A remarkable 81.73% alignment with human decisions is achieved on average in 188 extracted interactive cases, with a standout 83.12% in the highD dataset. Moreover, within the 145 cases showcasing dynamic interactional behaviors, our model maintains a 77.12% congruence, culminating in 6913 instances of alignment. The results demonstrate our advanced model’s capability in enabling AVs to make adaptive decisions with human-logic in dynamic interactive environments.