A bayesian approach to learning and planning for partially observable dynamical systems

Abstract

In this paper, we propose a learning and planning algorithm for a partially observable dynamic system. The Gaussian process state-space model (GPSSM) is adopted to learn the latent dynamics model from the partially observable measurements. GPSSM is a probabilistic dynamical system that represents unknown transition and/or measurement models as the Gaussian Process (GP), and enables the learning of a robust system model from a small number of partially observable time series data. GPSSM is integrated with a variant of Differential Dynamic Programing (DDP) called iterative Linear Quadratic Regulator (iLQR). The proposed method can generate a robust control policy for control/planning. Numerical examples are presented to demonstrate the applicability of the proposed method.