Nonparametric Adaptive Bayesian Optimal Control

Speaker:

Tao Chen, Department of Mathematics, University of Michigan, Ann Arbor

Description:

In this work, we use a nonparametric Bayesian approach to address the issue of Knightian uncertainty in optimal control problems. To handle model misspecification and estimation error in learning the unknown underlying model dynamics, we propose to use the Dirichlet process to model the unknown distribution of the underlying model in a nonparametric manner. Such framework integrates the optimization and online learning of the model without restricting it to any specific family of distributions. We will also present some numerical results to show the comparison between our approach and other classical control methods.

Mathematical Finance, Stochastic Analysis, and Machine Learning Seminar