Research on autonomous driving highway decision-making incorporating external spatial attention mechanism and double deep Q-network learning

Reinforcement learning can be used to create agents for autonomous decision making in autonomous driving. However, there are still key challenges such as learning inefficiency in solving high-dimensional dynamic driving task problems. To solve the above problems, this paper proposes a deep reinforcement learning decision-making method (ESA-DDQN) that incorporates the external spatial attention mechanism with double deep Q-network learning. State relations are first represented using an external spatial attention mechanism (ESA) and then combined with a double deep Q-network (DDQN) learning algorithm for strategy optimization. Experimental results show that ESA-DDQN achieves higher learning efficiency in learning compared to the baseline algorithm.