Piecewise Neural Networks for Function Approximation, Cast in a Form Suitable for Parallel Computation

Abstract

Abstract. We present a technique for function approximation in a partitioned domain. In each of the partitions a form containing a Neural Network is utilized with parameterized boundary conditions. This parameterization renders feasible the parallelization of the computation. Conditions of continuity across the partitions are studied for the function itself and for a number of its derivatives. A comparison is made with traditional methods and the results are reported.