A deep reinforcement learning approach integrating LSTM-GAT spatiotemporal fusion for autonomous driving decision-making

To address the limitations of traditional decision-making models in capturing dynamic interaction relationships between vehicles within complex traffic scenarios, this paper proposes a novel deep reinforcement learning approach integrating long short-term memory and graph attention network (LSTM-GAT) spatiotemporal fusion for autonomous driving decision-making. First, an encoder–decoder architecture is constructed via LSTM to predict the future motion trajectories and dynamic intentions of vehicles. Second, combined with temporal features, a vehicle interaction graph is built, and a dynamic GAT network based on a safety distance threshold is utilized for interaction feature extraction, enabling faster and safer identification of potential conflict relationships between vehicles. Subsequently, to enhance decision-making performance, a decision-making model and reward function are constructed based on deep reinforcement learning (DRL), integrated with the interaction feature extraction model. Moreover, we propose an action masking strategy to accelerate model training. Finally, the effectiveness and superiority of the proposed approach are verified through a series of experiments. The experimental results demonstrate that compared with state-of-the-art DRL algorithms, the DRL algorithm incorporating the LSTM-GAT spatiotemporal fusion-based interaction feature extraction model and the action masking strategy yields improvements in safety, traffic efficiency, and comfort while ensuring decision-making reliability.