Piecewise linear regression networks and its application to time series prediction

This paper presents a new approach of function approximation based on piecewise linear regression technique, referred to as the piecewise linear regression network (PLRN). The PLRN is designed for three purposes: 1) to alleviate the difficulty due to high dimensional settings of the given data; 2) to eliminate the necessity of forming ordered topological maps used in the conventional techniques of piecewise linear approximation; and 3) to achieve fast learning without being stuck to the local minima of an error surface. To show the effectiveness of our approach, the PLRN is applied to the prediction of Mackey-Glass chaotic time series and compared to other approaches.