Final Report Summary - SUPREL (Scaling Up Reinforcement Learning: Structure Learning, Skill Acquisition, and Reward Shaping)
Our project is concerned with developing the methodology to scale Reinforcement Learning (RL) to real-world applications. Broadly defined, RL is the discipline that is concerned with learning how to act in large dynamic environments by interacting with the environment, by learning from a teacher, or by imitating the behaviour of an expert. The goal of the proposed research is to develop the methodology that is required to turn RL from an attractive theoretical framework with relatively few applications to a more widely used and easily applicable methodology. The project focuses on three aspects of the RL methodology that are crucial for scaling it up to a household name: discovering structure in dynamic data, computing skills that can be re-used for different problems, and calculating a reward function that best fits the particular task at hand.