Automatic synthesis of neural networks using learning automata

A learning procedure for the structure identification of feedforward neural networks is proposed. The presented algorithm is based on a hierarchical system of learning automata. From a learning set, a given set of network structures and a pre-specified performance criterion, the learning procedure leads to the suitable neural network structure. The evaluation of the different network structures uses the back-propagation algorithm to estimate an index performance for each selected structure. The reinforcement scheme increments the probability of occurrence of good structures. The learning procedure stops when a probability converges to one. Simulation results are presented to show the effectiveness and the easiness of the studied approach.