Reinforcement Learning

Reinforcement learning problems involve an agent interacting with an environment. The agent must learn about the environment and must also discover how to act optimally in that environment. Hence, there is both a statistical component (learning about the environment) and a computational component (deciding how to act). In reinforcement learning, the environment is typically modeled as a controllable Markov process, so the agent must solve a Markov decision problem. Reinforcement learning can therefore be viewed as the study of tractable approximation algorithms for solving MDPs.

There are two main challenges in reinforcement learning research: (a) scaling up to large problems and (b) handling partially-observable Markov decision problems (where the agent cannot sense the entire state of the environment). In addition addressing these fundamental problems, we are also interested in finding innovative applications of RL.

Large Problems

Partially-Observable MDPs

A good source of tutorial information on Reinforcement Learning is the RL Archive at Michigan State University.

References


OSU Real-Time Skill Acquisition Group