A new method based on a recursive stochastic algorithm is presented fo
r neural networks synthesis. The cost function of the difference betwe
en the output of the net and the output of the process to be modelled
by the net has nonunique stationary points. The common optimization te
chniques lead to local optima. It is shown that the solution of this o
ptimization problem is connected with the construction of a convex env
elope, characterized by local extrema of the initial problem. Then a r
ecursive stochastic random search algorithm is derived for finding the
optimum using realizations of a random variable associated with the f
unction to be minimized. The application of this method is illustrated
by an experimental example concerning neural networks synthesis for a
n industrial calcinator.