Inverse mapping of continuous functions using feedforward neural networks

In this paper we present a methodology for solving inverse mapping of continuous functions modeled by multilayer feedforward neural networks. The methodology is based on an iterative update of the input vector towards a solution, which escapes local minima of the error function. The update rule is able to detect local minima through a phenomenon called "update explosion". The input vector is then relocated to a new position based on a probability density function (PDF) gradually constructed over the input vector space. The PDF is built using local minima detected during the search history. Simulation results demonstrate the effectiveness of the proposed method in solving the inverse mapping problem for a number of cases.