Hybrid Decision-making Strategy Combining Reinforcement Learning and Decision Rules for Highway Unmanned Driving

In recent years, reinforcement learning has been employed to address the challenge of unmanned driving decision-making, due to its ability to handle uncertainty and continuous optimization. However, compared with rule-based algorithms, its training time is considerable and the reliability of its results is difficult to guarantee. To enhance the security of decision-making processes, this paper proposes a hybrid decision-making algorithm designed to accelerate the training process of reinforcement learning and enhance the reliability of decision-making. The method first evaluates the confidence levels of both rule-based and learning-based strategies by leveraging the data collected throughout the training phase. Subsequently, an activation function is crafted to determine the final strategy with the greatest level of confidence. This guarantees that the final strategy performs at least on par with the rule-based approach. A simulation focused on highway driving was conducted as a case study to illustrate the effectiveness of the proposed method. The results demonstrate that the proposed approach offers greater profitability than the rule-based method and higher reliability compared to the learning-based method.