Multi-expert Learning Method for Autonomous Lane-changing Decision-making in Multi-scenario Highway Environments

The decision-making process of autonomous vehicles on highways involves a sequence of driving maneuvers aimed at improving safety and efficiency, which, however, results in considerable training time for the learning algorithm. This study proposes a multi-expert learning method(MELM) that integrates multiple actors (experts), each trained using the soft actor-critic(SAC) algorithm under constraints derived from distinct sub-layer scenarios. Each sub-layer scenario is defined according to the distinct properties of the original training scenario. Each expert controls the vehicle in its corresponding sub-layer scenario and is integrated via a classifier that identifies the applicable sub-scenario. As a result, the MELM significantly reduces the model's training time by 62.19% compared to a single SAC model, while also improving driving safety and efficiency, attributed to a remarkable reduction in the training difficulty of SAC. The proposed MELM is compared against several state-of-the-art methods under representative driving scenarios. Simulation results show a 27.06% improvement in driving efficiency compared to the single SAC model, along with high safety performance characterized by zero collision and off-road incidents across 100 testing episodes ( 100 000 timesteps). Furthermore, the adaptability of MELM is validated through simulation in a variety of scenarios with different condition settings.