Decision making for autonomous vehicles: A mixed curriculum reinforcement learning approach and a novel safety intervention method

Abstract

Reinforcement learning is considered one of the most promising approaches for decision-making in autonomous vehicles within interactive scenarios. However, its implementation faces challenges of insufficient safety and limited learning efficiency due to the stochastic nature of exploration and the complexity of the exploration space. In this paper, a mixed curriculum learning (MCL) approach, incorporating an intervention method called discrepancy-directed Bernoulli intervention (DDBI), is proposed to address these challenges in reinforcement learning. Firstly, the algorithm divides the training process into a safety phase and a performance phase. The agent focuses on accomplishing the safety task first, which is followed by the performance task. Secondly, DDBI introduces an additional safety agent to intervene in hazardous situations using a novel probability-based method, thereby enhancing the safety of the training process while preserving the exploratory nature of reinforcement learning. Finally, the proposed approach is evaluated in a lane change scenario with random traffic flow. Comprehensive comparative experiments with other algorithms demonstrate that the proposed approach outperforms in both safety and learning efficiency.