Theory and Practice of Reinforcement Learning

Reinforcement Learning (RL) is closely related to how animals and humans act and learn. Without a teacher, solely from occasional real-valued pain and pleasure signals, RL agents must discover how to interact with a dyniamic environment to maximize their future expected reward. We intend to advance the state of the art in solving very hard RL problems with large, partially observable state spaces, and in combining theoretically asymptotically optimal with practically feasible algorithms. Concrete subgoals include: (1) Develop RL benchmarks that measure both depth and speed of RL algorithms, being fairly easy to solve partially but very hard to solve completely, requiring the policies to map high-dimensional sate descriptions (HDSD) to actions, (2) study the applicability of multi-dimensional recurrent neural networks to RL problems involving HDSD, (3) further develop a new class of hybrid evolutionary RL algorithms that 'memetically' searches topologies and weight sets of RL machines at different levels, (4) develop a new RL algorithm based on expectation maximization, (5) develop a framework for making asymptotically otpimal universal search practical, combining it with RL as well as other algorithms.