Towards Global Reinforcement Learning

Speaker: Milen Pavlov

Sequential decision making under uncertainty is a ubiquitous problem. In everyday situations we are faced with a series of decisions that aim to maximize the probability of achieving some goal. However, decision outcomes are often uncertain and it is not always immediately evident how to determine if one decision is better than another. The Reinforcement Learning framework overcomes this difficulty by learning to make optimal decisions based on interactions with the environment. One drawback of Reinforcement Learning is that it requires too much data (interactions) to learn from scratch. For this reason, current approaches attempt to incorporate prior information in order to simplify the learning process. However, this is usually accomplished by making problem-specific assumptions, which limit generalizability of the approaches to other problems. This thesis presents the first steps towards a new framework that incorporates and exploits a broad prior knowledge in a principled way. It uses Constraint Satisfaction and Bayesian techniques to construct and update a belief over the environment, as well as over good decisions. This allows for incorporating broad types of prior knowledge without limiting generalizability. Preliminary experiments show that the framework's algorithms work well on toy problems in simulation and encourage further research on real-world problems.