Transfer Learning for Reinforcement Learning with Dependent Dirichlet Process and Gaussian Process

The ability to transfer knowledge across tasks is important in guaranteeing the performance of lifelong learning in autonomous agents. We propose a flexible Bayesian Nonparametric (BNP) model based architecture for transferring knowledge between reinforcement learning domains. A Dependent Dirichlet Process Gaussian Process hierarchial BNP model is used to cluster different classes of source MDPs in sequential batch mode. Our architecture is flexible and adaptable because it allows for clustering across MDP transition functions or value-function representations, and because it is capable of incorporating new tasks and forgetting old irrelevant tasks. These properties ensure the performance of lifelong learning without requiring excessive onboard resources.