Algorithms and Representations for Reinforcement Learning

Machine Learning is a field of research aimed at constructing intelligent machines that gain and improve their skills by learning and adaptation. As such, Machine Learning research addresses several classes of learning problems, including for instance, supervised and unsupervised learning. Arguably, the most ubiquitous and realistic class of learning problems, faced by both living creatures and artificial agents, is known as Reinforcement Learning. Reinforcement Learning problems are characterized by a long-term interaction between the learning agent and a dynamic, unfamiliar, uncertain, possibly even hostile environment. Mathematically, this interaction is modeled as a Markov Decision Process (MDP). Probably the most significant contribution of this thesis is in the introduction of a new class of Reinforcement Learning algorithms, which leverage the power of a statistical set of tools known as Gaussian Processes. This new approach to Reinforcement Learning offers viable solutions to some of the major limitations of current Reinforcement Learning methods, such as the lack of confidence intervals for performance predictions, and the difficulty of appropriately reconciling exploration with exploitation. Analysis of these algorithms and their relationship with existing methods also provides us with new insights into the assumptions underlying some of the most popular Reinforcement Learning algorithms to date.