Risk-rule filtered LSTM-dueling DQN for autonomous lane change decision

Autonomous vehicle lane-change decision making has long been a prominent research topic in intelligent transportation systems. To enhance both the safety and efficiency of lane changes in dynamic traffic environments, we propose a Risk-Rule Filtered Long Short-Term Memory Dueling Deep Q-Network (LSTM-DDQN) method for autonomous lane-change decision making under complex traffic scenarios.

This approach achieves intelligent decision making by integrating a risk-rule filter with a reinforcement-learning framework. First, the model employs a Time-to-Collision (TTC)-based risk filtering mechanism to select target vehicles that pose potential collision risks to the ego vehicle as the basis for decision making. Next, the LSTM network processes the filtered observation sequences to extract dynamic traffic-context information with temporal features. These features are then input into a Dueling DQN architecture to separately estimate the state-value function and the action-advantage function, thereby optimizing the final lane-change action selection.

Compared to a standard Dueling DQN baseline implemented under identical simulation settings (same architecture, hyper-parameters, and training protocol but without the LSTM temporal module and TTC-based risk filter), the proposed LSTM-DDQN achieves a 10.9% increase in normalized average single-step reward.

The proposed model was validated on the Simulation of Urban Mobility (SUMO) platform, underscoring its superior performance in improving lane-change safety.