Double Broad Q-Network for Overtaking Control of Autonomous Driving

Overtaking and lane-changing are essential maneuvers in autonomous driving. A correct and timely operation can reduce the probability of collisions, enable vehicles to avoid potential hazards on the road, and enhance overall safety. In addition, they also play a vital role in ensuring the efficient flow of traffic and timely arrival at destinations. Currently, the dominant research on this problem is based on the Deep Q-Network (DQN) approach and some of its improved variants to solve it. However, limited by deep networks, there are inevitably problems such as complex network structures and numerous parameters to be optimized, which will prolong the response time for self-driving vehicles to make decisions, and further increase the safety risks for drivers. The emergence of the Broad Learning System (BLS) offers a new scheme, which can overcome the drawbacks of the deep network mentioned above. Considering the importance of response time for vehicle safety decisions, we proposed the Double Broad Q-Network (DBQN) algorithm, which replaces the deep network inside Double Deep Q-Network (DDQN) with BLS, to solve lane-changing and overtaking problems. Finally, we designed a new reward function based on three scenarios, and experiments have shown that our method can reduce response times and collision rates, thereby improving vehicle driving safety.