ATT-D3RQN: An Attention-Based Deep RL Framework for Urban Pursuit-Evasion Games

Addressing the challenge of the pursuit-evasion games in one-on-one situations within urban traffic, the Attention-Double Deep Recurrent Q-Network (ATT-D3RQN) is proposed through three methodological innovations. First, we implement an asynchronous update strategy between target and evaluation network by incorporating the Double DQN mechanism. Subsequently, a Dueling architecture is developed to decouple the state value function from the advantage function, enabling more precise value estimation. Finally, we integrate a Long ShortTerm Memory (LSTM) module with a self-attention mechanism to effectively capture spatiotemporal correlation features in complex game scenarios. The comprehensive simulations demonstrate that the proposed algorithm achieves an improvement of at least $\mathbf{2 1. 5 \%}$ in capture rate compared to the standard DQN across multiple test scenario.