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

• I will be giving part of a one day tutorial on reinforcement learning.
• I am co-chairing / on the organizing committee for the Adaptive and Learning Agents (ALA) workshop and the Agents Learning Interactively from Human Teachers (ALIHT) workshop.
• I'll be presenting our full paper: When Should There be a "Me" in "Team"? Distributed Multi-Agent Optimization Under Uncertainty by Taylor, Tambe, Jain, Jin, and Yooko
• I also contributed to the following accepted workshop papers at ALIHT, DCR, and ALA (respectively):
  • Integrating Human Demonstration and Reinforcement Learning: Initial Results in Human-Agent Transfer by Taylor and Chernova.
  • Towards a Theoretic Understanding of DCEE by Alfeld, Taylor, Tandon, and Tambe.
  • Transfer Learning for Reinforcement Learning on a Physical Robot by Barrett, Taylor, and Stone.

Research

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

My research focuses on agents, physical or virtual entities that interact with their environments. My main goals are to enable individual agents, and teams of agents, to

1. learn tasks in real world environments that are not fully known when the agents are designed;
2. perform multiple tasks, rather than just a single task; and
3. allow agents to robustly coordinate with, and reason about, other agents.

		Transfer Learning My dissertation focused on leveraging knowledge from a previous task to speed up learning in a novel task, focusing on reinforcement learning domains. I gave a talk at AGI-08 that gives a brief introduction to, and motivation for, transfer learning. Representative Publication: Transfer Learning via Inter-Task Mappings for Temporal Difference Learning (JMLR-07) Full list of relevant publications
		Reinforcement Learning Much of my graduate work centered on reinforcement learning (RL) tasks, where agents learn to perform (initially) unknown tasks by optimizing a scalar reward. RL is well suited to allowing both virtual and physical agents to learn when humans are unable (or unwilling) to design optimal solutions themselves. Representative Publication Critical Factors in the Empirical Performance of Temporal Difference and Evolutionary Methods for Reinforcement Learning (JAAMAS-09) Full list of relevant publications
		Multi-agent Exploration and Optimization Since coming to USC, one of the most exciting projects we have worked on is a version of Distributed Constraint Optimization Problem (DCOP) where the agents have unknown rewards. This may also be thought of as a multi-agent, multi-armed bandit. This problem is relevant for tasks that require coordination under uncertainty, such as in wireless sensor networks. Representative Publication DCOPs Meet the Real World: Exploring Unknown Reward Matrices with Applications to Mobile Sensor Networks (IJCAI-09) Full list of relevant publications
		Use-Based Machine Learning I am also interested in applying machine learning techniques to problems not explicitly designed by machine learning specialists. Put differently, I want to improve learning algorithms so that they are more useful in real-world situations and usable by non-specialists. Representative Publication Feature Selection and Policy Optimization for Distributed Instruction Placement Using Reinforcement Learning (PACT-08) Full list of relevant publications
		Multi-Agent Planning Under Uncertainty Distributed Partially Observable Markov Decision Processes (DEC-POMDPs) are a powerful formalism for allowing multiple agent/robots to plan in real world domains. However, the computational complexity of current techniques must be significantly reduced in order for algorithms to be sufficiently fast for real-world applications. Representative Publication Exploiting Coordination Locales in Distributed POMDPs via Social Model Shaping (ICAPS-09) Full list of relevant publications
		Game Theory for Security Applications Game theory has been an important component of deployed security applications, such as the ARMOR project at the Los Angeles International Airport. I have worked to evaluate one such project, helping to verify that game theoretic principals still apply in the deployed, real world, application. Representative Publication A Framework for Evaluating Deployed Security Systems: Is There a Chink in your ARMOR? (Informatica-10) Full list of relevant publications
		Crowd Evacuation Simulation Agent-based modeling of large-scale crowds has the potential to help plan building layouts, assist with security and disaster response training, and estimate the impact of different disasters. In collaboration with Massive Software, the LAX police, and the USC School of Cinema, we aim to create psychologically and physically plausable simulations of hundreds of agents with movie-quality visualization. Representative Publication Agent-based Evacuation Modeling: Simulating the Los Angeles International Airport (EMWS-09)

• I gave a tutorial with Alessandro Lazaric on transfer learning in RL at AAMAS-09.
• I was a co-chair of the ALA workshop at AAMAS-09.
• I helped organize the 2009 RL Competition.
• Selected code from my dissertation can be found in the on-line appendix.
• I organized an AAAI-08 workshop: Transfer Learning for Complex Tasks.
• In the Fall of 2007 and the Spring of 2008 I taught cs 108, Introduction to Linux.