A central problem in dynamical system modeling is state discovery--that is, finding a compact summary of the past that captures the information needed to predict the future. Predictive State Representations (PSRs) enable clever spectral methods for state discovery; however, while consistent in the limit of infinite data, these methods often suffer from poor performance in the low data regime. In this paper we develop a novel algorithm for incorporating domain knowledge, in the form of an imperfect state representation, as side information to speed spectral learning for PSRs. We prove theoretical results characterizing the relevance of a user-provided state representation, and design spectral algorithms that can take advantage of a relevant representation. Our algorithm utilizes principal angles to extract the relevant components of the representation, and is robust to misspecification. Empirical evaluation on synthetic HMMs, an aircraft identification domain, and a gene splice dataset shows that, even with weak domain knowledge, the algorithm can significantly outperform standard PSR learning.

SUMMARY: This paper proposes a method to incorporate prior knowledge into the spectral learning algorithm for predictive state representations (PSR). The prior knowledge consists of an imperfect/incomplete state representation which is'refined' and'completed' by the learning algorithm. This contribution addresses one of the main caveats of spectral methods: while these methods are fast and consistent, they tend to perform worse than local methods (e.g. By leveraging domain specific knowledge, the proposed algorithm overcomes this issue. The proposed extension, PSR-f, is relatively straightforward: the belief vector at each time step is the concatenation of the user-specified state representation f with a learned state representation b; the parameters of b are learned in the same fashion as for the classical method by solving linear regression problems constrained to the row space of the concatenation of some Hankel/system matrices (e.g. now mapping [P(T h); P(h)f(h)] to P(oT h) for each B_o).

A central problem in dynamical system modeling is state discovery--that is, finding a compact summary of the past that captures the information needed to predict the future. Predictive State Representations (PSRs) enable clever spectral methods for state discovery; however, while consistent in the limit of infinite data, these methods often suffer from poor performance in the low data regime. In this paper we develop a novel algorithm for incorporating domain knowledge, in the form of an imperfect state representation, as side information to speed spectral learning for PSRs. We prove theoretical results characterizing the relevance of a user-provided state representation, and design spectral algorithms that can take advantage of a relevant representation. Our algorithm utilizes principal angles to extract the relevant components of the representation, and is robust to misspecification.