Matt Rudary

My primary research has been in the area of predictive state models for stochastic dynamical systems. Predictive state models use predictions about future outcomes as their state, rather than distributions over latent variables, as state is commonly represented in traditional models. These models have the advantage that observational data provides samples of the values of state, whereas the latent variables of traditional models are never observed. The original paper on PSRs established that a PSR can represent any system that a POMDP can and, moreover, can do it using no more space than a POMDP.

More specifically, my current research is on the predictive linear Gaussian model (PLG), which models dynamical systems with continuous actions and observations. It has equivalent representational power to the linear dynamical system (on which the widely used Kalman filter is based). I have designed a consistent parameter estimation algorithm for this model, and am exploring improvements to the model itself and to estimating its parameters.

The Autominder project is directed by Martha Pollack. Autominder is a cognitive orthotic designed to help people with mild cognitive impairment carry out routine tasks while limiting users' dependence on it. My adviser (Satinder Singh) and I worked with Professor Pollack to study the effectiveness of using reinforcement learning to improve Autominder's user interaction.

At HRL Laboratories, I worked on a sensor resource management problem in which a sensor with several modes must trade off among qualities like accuracy, focus area, and energy cost to maximize an informational measure. I adapted Kearns, Mansour and Ng's sparse sampling algorithm to this problem, which has a mixture of continuous and discrete elements.

Publications

Conference Papers

M. Rudary and S. Singh (2006). Predictive linear-Gaussian models of controlled stochastic dynamical systems. In Cohen, W. & Moore, A. (Eds.), Proceedings of the 23rd International Conference on Machine Learning, pp. 777–784. pdf

M. Rudary, D. Khosla, J. Guillochon, P. A. Dow and B. J. Blyth (2006). A sparse sampling planner for sensor resource management. In Kadar, I. (Ed.), Proceedings of the SPIE Vol. 6235: Signal Processing, Sensor Fusion, and Target Recognition XV, pp. 62350A-1–62350A-9. pdf

M. Rudary, S. Singh and D. Wingate (2005). Predictive linear-Gaussian models of stochastic dynamical systems. In Bacchus, F. & Jaakkola, T. (Eds.) Uncertainty in Artificial Intelligence 21, pp. 501–508. pdf

S. Singh, M. R. James and M. Rudary (2004). Predictive state representations: A new theory for modeling dynamical systems. In Chickering, M. & Halpern, J. (Eds.), Uncertainty in Artificial Intelligence 20, pp. 512–519. pdf

M. Rudary, S. Singh and M. Pollack (2004). Adaptive cognitive orthotics: Combining reinforcement learning and constraint-based temporal reasoning. In Greiner, R. & Schuurmans, D. (Eds.), Proceedings of the 21st International Conference on Machine Learning, pp. 719–726. pdf

C. Kiekintveld, M. P. Wellman, S. Singh, J. Estelle, Y. Vorobeychik, V. Soni and M. Rudary (2004). Distributed feedback control for decision making on supply chains. In Zilberstein, S., Koehler, J., & Koenig, S. (Eds.), Proceedings of the 14th International Conference on Automated Planning and Scheduling, pp. 384–392. pdf

M. Rudary and S. Singh (2004). A nonlinear predictive state representation. In Thrun, S., Saul, L. K., & Schölkopf, B. (Eds.), Advances in Neural Information Processing Systems 16, pp. 855–862. pdf

Workshop Papers

M. Rudary, S. Singh and M. Pollack. Reinforcement learning for adaptive cognitive orthotics. In Supervisory Control of Learning and Adaptive Systems: Papers from the AAAI Workshop, 2004. pdf

Matt Rudary

Research Interests

Publications

Conference Papers

Workshop Papers