An association of simulated annealing and electrical simulator SPICE-PAC for learning of analog neural networks

Simulated annealing adapted to continuous variables is used to determine the synaptic coefficients of an analog multilayer neural network, approximating any continuous function of one or several variables. The open electrical simulator SPICE-PAC driven by simulated annealing produces a globally optimal set of synaptic weights, in a reasonable time and without requiring heavy and inaccurate gradient computations. The authors illustrate and improve the weights-tuning strategy through two simple examples.<<ETX>>