IOE 899 Seminar Presents: Enlu Zhou, University of Illinois - Oct 27, 2010
Oct 27, 2010 | 4-5 PM | 1680 IOE
"Solving Continuous-State Partially Observable Markov Decision Processes (POMDPs) via Density Projection"
Partially observable Markov decision processes (POMDPs) provide a powerful paradigm for modeling discrete-time optimal decision making under uncertainty and partial observation. POMDPs suffer from the well-known curse of dimensionality, and continuous-state POMDPs are especially hard to solve, due to the infinite dimensionality of the belief state. Based on the idea of density projection, I have developed a theoretically sound and computationally viable method, which effectively reduces the dimension of the belief state and has the flexibility to represent arbitrary belief states, such as multimodal or heavy tailed distributions. The density projection idea is also incorporated into particle filtering to provide a filtering scheme for online decision making. I have proved rigorous convergence results and error bounds, and obtained good numerical results on an inventory control example and a class of financial investment problems under stochastic volatility.
Audience-Based Site-Wide Navigation:back to top