A greedy approach to identification of piecewise affine models

This paper addresses the problem of identification of piecewise affine (PWA) models. This problem involves the estimation from data of both the parameters of the affine submodels and the partition of the PWA map. The procedure that we propose for PWA identification exploits a greedy strategy for partitioning an infeasible system of linear inequalities into a minimum number of feasible subsystems: this provides an initial clustering of the datapoints. Then a refinement procedure is applied repeatedly to the estimated clusters in order to improve both the data classification and the parameter estimation. The partition of the PWA map is finally estimated by considering pairwise the clusters of regression vectors, and by finding a separating hyperplane for each of such pairs. We show that our procedure does not require to fix a priori the number of affine submodels, which is instead automatically estimated from the data.