Open AccessKnot-optimizing spline networks (KOSNETS) for nonparametric regression
Author(s) -
Song Wang,
Quanxi Shao,
Xian Zhou
Publication year - 2008
Publication title -
journal of industrial and management optimization
Language(s) - English
Resource type - Journals
eISSN - 1553-166X
pISSN - 1547-5816
DOI - 10.3934/jimo.2008.4.33
Subject(s) - knot (papermaking) , spline (mechanical) , nonlinear system , computer science , mathematical optimization , series (stratigraphy) , multivariate adaptive regression splines , nonlinear programming , mathematics , algorithm , nonparametric regression , regression analysis , machine learning , paleontology , physics , structural engineering , quantum mechanics , chemical engineering , engineering , biology
In this paper we present a novel method for short term forecast of time series based on Knot-Optimizing Spline Networks (KOSNETS). The time series is first approximated by a nonlinear recurrent system. The resulting recurrent system is then approximated by feedforward B-spline networks, yielding a nonlinear optimization problem. In this optimization problem, both the knot points and the coefficients of the B-splines are decision variables so that the solution to the problem has both optimal coefficients and partition points. To demonstrate the usefulness and accuracy of the method, numerical simulations and tests using various model and real time series are performed. The numerical simulation results are compared with those from a well-known regression method, MARS. The comparison shows that our method outperforms MARS for nonlinear problems.20 page(s