A "recruiting neural-gas" for function approximation
Michael J Marie Aupetit, Pierre Couturier, Pierre Massotte · 2000
An algorithm for function approximation with an artificial neural network is presented. It is based on neural-gas networks which combine self-organization of the neurons in the input space and supervised learning of the output values according to the function to approximate. The original learning rule of the input weights is modified to take into account the output error. The neurons with a greater error tend to "recruit" their neighbors to help them in their approximation task. The resulting network called a "recruiting neural-gas", organizes the neurons in the input space respecting the input data distribution and also the output error density. This algorithm gives very promising results and perspectives.