Matthew Taylor mtaylor@cs.utexas.edu Postdoctoral Research Associate Department of Computer Science The University of Southern California Los Angeles, CA 90089

Research

I am currently working with Milind Tambe as part of the TEAMCORE research group.
I am a former member of the Learning Agents Research Group, directed by Professor Peter Stone.

I gave a talk at AGI-08 (artifical general intelligence). The talk gives a brief introduction to, and motivation for, transfer learning: http://video.google.com/videoplay?docid=1984013763155542745&hl=en

Reinforcement learning (RL) methods have become popular in recent years because of their ability to solve complex tasks with minimal feedback. While these methods have had experimental successes and have been shown to exhibit some desirable properties in theory, the basic learning algorithms have often been found slow in practice. Much of the current RL research focuses on speeding up learning by taking advantage of domain knowledge or better utilizing agents' experience. The ambitious goal of transfer learning, when applied to RL tasks, is to accelerate learning on some target task after training on a different, but related, source task. In my thesis I utilize transfer learning methods to speed up learning in RL tasks via experience from previously learned tasks, increasing RL's applicability to difficult tasks and allowing agents to better generalize their experience.

My work concentrates on methods for performing transfer between source and target tasks in which agents may take different actions and/or use different state variables to describe their world view. These methods enable transfer by relying on domain knowledge to describe similarities between the two tasks. However, there may be situations in which domain knowledge is unavailable, or insufficient, to describe how two given tasks are related. I therefore am also interested in how little domain knowledge is required to transfer knowledge successfully and introduce algorithms that can learn the similarities between pairs of tasks automatically.