TOWARD GENERATING NEURAL-NETWORK STRUCTURES FOR FUNCTION APPROXIMATION

This paper proposes an algorithm based on the back propagation procedure that dynamically configures the structure of feedforward multilayered neural networks and demonstrates its potential for control applications. The algorithm applies an intelligent generate-and-test procedure that evaluates the learning performance of the structure used and modify it accordingly by exploring different alternatives and selecting the most promising one. The algorithm modifies the structure of the neural networks by adding or deleting neurons/layers. The efficiency of the algorithm is demonstrated using several case studies with very promising results.