摘要：AbstractWe introduce a Bayesian nonparametric approach to identification of piecewise affine ARX systems. The clustering properties of the Dirichlet process are used to construct a prior over the partition of the regressor space as well as the parameters of each local model. This enables us to probabilistically reason about and to identify the number of modes, the partition of the regressor space, and the linear dynamics of each local model from data. By appropriate choices of base measure and likelihood function, we give explicit expressions for how to perform both inference and prediction. Simulations and experiments on real data from a pick and place machine are used to illustrate the capabilities of the new approach.
