Annealing based dynamic learning in second-order neural networks

An algorithm that simultaneously determines an appropriate number of neurons and their interaction parameters in a single hidden layer feedforward neural network classification model is proposed. First, a large pool of candidate hidden units with second-order inputs interaction is constructed. Next, the hidden layer is designed by selecting appropriate units from the pool. This is achieved through global hidden layer optimization by a simulated annealing technique that adds and deletes hidden units as needed. Experimental results using the proposed model show improved generalization and reduced complexity as compared to previous constructive learning algorithms based on greedy design techniques.