Learning non-linear dynamical systems by alignment of local linear models

Learning dynamical systems is one of the important problems in many fields. In this paper, we present an algorithm for learning non-linear dynamical systems which works by aligning local linear models, based on a probabilistic formulation of subspace identification. Because the procedure for constructing a state sequence in subspace identification can be interpreted as the CCA between past and future observation sequences, we can derive a latent variable representation for this problem. Therefore, as in a similar manner to the recent works on learning a mixture of probabilistic models, we obtain a framework for constructing a state space by aligning local linear coordinates. This leads to a prominent algorithm for learning non-linear dynamical systems. Finally, we apply our method to motion capture data and show how our algorithm works well.