It is well known that the global optimum of a MDP with finite state and action sets can be obtained through methods based on dynamic programming. Unfortunately, these techniques are known to suffer from the curse of dimensionality, which makes them infeasible for many real-world problems of interest. As a result, most research in the reinforcement learning and control theory literature has focused on obtaining approximate or locally optimal solutions. There exists a broad spectrum of such techniques, including approximate dynamic programming methods, tree search methods, local trajectory-optimization techniques, such as differential dynamic programming and iLQG, and policy search methods. In this talk I shall provide an introduction to policy search methods, which are a family of algorithms that have proven extremely popular in recent years, and which have numerous desirable properties that make them attractive in practice.

Our May meetup will be a robotics themed one with two speakers from KCL and Oxford. Inspired by the antagonistic human musculoskeletal system, the current trend in mechatronic design is to include physically compliant elements into the embodiment of robotic devices. The promise of such'soft robotic' systems, includes safety and agility. However, these offerings are tempered by the increased complexity of the system dynamics leading to difficulty in control. Learning (by demonstration or reinforcement) is often advocated as a means of dealing with this complexity, and can allow us to exploit the principles of human sensorimotor control to improve our robotic systems.