LEARNING OF NEURAL NETWORKS APPROXIMATING CONTINUOUS-FUNCTIONS THROUGH CIRCUIT SIMULATOR SPICE-PAC DRIVEN BY SIMULATED ANNEALING

Simulated annealing (SA) adapted to continuous variables is used to de termine the synaptic coefficients of an analogue multilayer neural net work, approximating any continuous function of one or several variable s. The 'open' electrical simulator SPICE-PAC driven by SA produces a g lobally optimal set of synaptic weights, in a reasonable time and with out requiring heavy and inaccurate gradient computations. We illustrat e and improve our weights tuning strategy through simple examples.