Incremental Learning through Probabilistic Behavior Prediction

Learning from expert demonstrations effectively reduces the number of interactions required to train a policy between the learning agent and its environment. Where agent-environment interactions can be costly, reinforcement learning is critical, and imitation learning suffers significantly from learning hierarchical policies when the imitative agent encounters an unobserved state by the expert agent. We propose a probabilistic incremental imitation learning method that employs a Dynamic Bayesian Network to encode observed teaching-agent's behaviors. The presented model grows and matures based on imitation loss formulation at a discrete level in the online learning procedure. The learning agent is trained by using long-term predictions from the generative learning model to replicate the teacher's motion while learning how to choose an appropriate action through new experiences. Our results affirm that a Dynamic Bayesian optimal approach provides a principled framework and outperforms conventional reinforcement learning methods.