Discovery and learning of models with predictive state representations for dynamical systems without reset

Modeling dynamical systems is a common problem in science and engineering. After a system has been modeled, the system can be controlled and predicted. Predictive state representations (PSRs) is a recently proposed method of modeling controlled dynamical systems. One central problem in the PSRs literature is concerned with discovery and learning of PSRs. This paper presents a new algorithm for discovery and learning of PSRs by using only a continuous trace of actions and observations as the training data, in which the history at any time step in the training data can be identified, and then the prediction of test at a history and the PSR model of the system can be obtained. We empirically evaluate and compare our algorithm on a standard set of POMDP test problems and the empirical results show that our algorithm is competitive and outperforms the suffix-history algorithm.