ANN Models Optimized using Swarm Intelligence Algorithms

Artificial Neural Network (ANN) has found widespread application in the field of classification. Many domains have benefited with the use of ANN based models over traditional statistical models for their classification and prediction needs. Many techniques have been proposed to arrive at optimal values for parameters of the ANN model to improve its prediction accuracy. This paper compares the improvement in prediction accuracy of ANN when it is trained using warm intelligence algorithms. Swarm intelligence algorithms are inspired by the natural social behaviour of a group of biological organisms. Models have been formulated for evaluating the various ANN-Swarm Intelligence combinations. Fault prediction in Object oriented systems through the use of OO metrics has been considered as the objective function for the models. The swarm intelligence algorithms considered in this paper are Particle Swarm Optimization, Ant Colony Optimization, Artificial Bee Colony Optimization and Firefly. The object oriented metrics and fault information for the analysis have been taken from NASA public dataset. The models are compared for their convergence speed and improvement in prediction accuracy over traditional ANN models. The results indicate that Swarm Intelligence Algorithms bring improvement over ANN models trained with gradient descent. Key-Words: Artificial Neural Network, Swarm Intelligence, Particle Swarm Optimization, Ant Colony Optimization, Artificial Bee Colony Optimization, Firefly