Scalable Reinforcement Learning Systems

Background and Motivation

Reinforcement learning (RL) represents the most promising framework for designing intelligent machines. As a biologically-inspired paradigm, RL offers the potential to build sub-optimal decision making systems that operate under partial and imperfect observations.

Research Approach

Our research focuses on the use of function approximation techniques in adaptive dynamic programming frameworks, to design systems that scale in both size and speed. In particular, we study effective solutions for solving Partially-Observable Markov Decision Processes (POMDPs), which correspond to a broad range of real-life stochastic control problems. MIL Wiki