Imitation Learning Through Prior Injection in Markov Decision Processes

Specification of an appropriate reward function is vital to Reinforcement Learning (RL). This paper proposes a technique to learn the reward function from observed tracklets to clone the behavior of pedestrians moving in an urban area. The information is then incorporated in a Markov Decision Process (MDP)-driven agent represented using Factor Graph paradigm. We show how this approach can be formalized in a completely probabilistic framework treating the estimated behavior as prior information to be injected into the RL model. We have tested the proposed approach on real scenes and using several path planning algorithms.