Title: Using Expert Traces to Reduce Training Time in Reinforcement Learning

Speaker: Joel Goldfinger

Start Date: Monday, April 21, 2008, 11:15AM

Location: 325b ITE

Abstract: Reinforcement learning algorithms are typically used to solve Markov decision problems. Using reinforcement learning to find an optimal or sub-optimal solution is a slow process. The use of prior knowledge can shorten the length of this task. In this thesis, we explore the use of expert traces to speed up learning. An expert trace is a record of the states visited and actions taken by an expert in a given domain, which is then used to reduce the training time of reinforcement learning. We demonstrate in a variety of domains significant improvements in learning when complementing reinforcement learning with expert traces.

Host: Tim Oates