A Tutorial on Reinforcement Learning Techniques

Reinforcement Learning (RL) is learning through direct experimentation. It does not assume the existence of a teacher that provides examples upon which learning of a task takes place. Instead, in RL experience is the only teacher. With historical roots on the study of conditioned reflexes, RL soon attracted the interest of Engineers and Computer Scientists because of its theoretical relevance and potential applications in fields as diverse as Operational Research and Robotics. Computationally, RL is intended to operate in a learning environment composed by two subjects: the learner and a dynamic process. At successive time steps, the learner makes an observation of the process state, selects an action and applies it back to the process. The goal of the learner is to find out an action policy that controls the behavior of this dynamic process, guided by signals (reinforcements) that indicate how well it is performing the required task. These signals are usually associated to some dramatic condition — e.g., accomplishment of a subtask (reward) or complete failure (punishment), and the learner’s goal is to optimize its behavior based on some performance measure (a function of the received reinforcements). The crucial point is that in order to do that, the learner must evaluate the conditions (associations between observed states and chosen actions) that lead to rewards or punishments. In other words, it must learn how to assign credit to past actions and states by correctly estimating costs associated to these events. Starting from basic concepts, this tutorial presents the many flavors of RL algorithms, develops the corresponding mathematical tools, assess their practical limitations and discusses alternatives that have been proposed for applying RL to realistic tasks, such as those involving large state spaces or partial observability. It relies on examples and diagrams to illustrate the main points, and provides many references to the specialized literature and to Internet sites where relevant demos and additional information can be obtained.