jphanna [at] cs.utexas.edu
Department of Computer Science
The University of Texas at Austin
I am a fourth year PhD student in the Learning Agents Research Group in the Computer Science Department at the University of Texas at Austin. My advisor is Peter Stone.
I study a branch of machine learning called reinforcement learning (RL). The goal of RL is that autonomous agents can acquire useful skills from interaction with their environment. A main goal of my research is applying RL algorithms to allow robots to acquire new capabilities..
I graduated from the University of Kentucky with a B.S. in computer science and mathematics.
As an undergraduate I worked with Judy Goldsmith on algorithms for multi-objective and factored Markov Decision Processes.
I also spent a summer working with Patrice Perny and Paul Weng at LIP6 in Paris, France working on approximating the set of Pareto optimal solutions for multi-objective MDPs.
Many RL algorithms are able to solve complex control tasks in simulated environments. However, the amount of experience required to find a good solution often prevents their application to physical robotic systems. Towards removing this barrier, I'm investigating methods for transferring behaviors learned in simulation to physical robots. A second goal of my research is to evaluate new robot behaviors before they are deployed on the physical robot.
I am a part of UT Austin Villa - UT's robot soccer team. I've contributed to both the standard platform (SPL) and 3D simulation league. In SPL I work on motion control (e.g. walking) and high level behavior (e.g. keeper behavior, kicking strategy). For the 3D Sim team I've used policy search reinforcement learning to learn kicks for different length kicks.
I am also part of a project investigating adaptive tolling in traffic networks. Assuming that travelers desire to minimize travel time and tolls paid, adjusting tolls can improve the efficiency of the system as traffic patterns change over a day. While existing system might adapt tolls on a few links in a network, we consider potentially tolling every link to maximize performance gains.