Hierarchical nonparametric Bayesian Reinforcement Learning

During his PhD, Samuel will work on building scalable sample-efficient autonomous machines. The approach would be to use a hierarchical approach, and thus break down tasks into sub-tasks to achieve, and as a result enable us to scale to larger problems and transfer knowledge between tasks through the knowledge of sub-tasks. He would also bring scale by using spatial abstractions in sub-tasks, and thus reduce the state-space of individual sub-tasks based on their relevant variables. For the above, Samuel would take a model-based Bayesian approach using Bayesian non-parametrics to model the environment and then Value-Function/policy-search methods to find optimal sub and parent policies. One of the main challenges he would like to contribute to solving is making Gaussian processes more scalable, and designing efficient hierarchical frameworks for learning.