Learning faster to perform autonomous lane changes by constructing maneuvers from shielded semantic actions

This paper introduces a new method to solve tactical decision making problems for highway lane changes. In the system design, reference sets for low level controllers are employed to formulate semantic meaningful actions used by reinforcement learning algorithm. Safety is ensured by preemptively shielding the Markov decision process (MDP) from unsafe actions. This frees the agent to focus on learning how to interact efficiently with the surrounding traffic. By introducing human demonstration with supervised loss as better exploration strategy, the learning process and initial performance are boosted further.