A Hybrid ACOR Algorithm for Pattern Classification Neural Network Training

Abstract

The ACOR algorithm is an Ant Colony Optimization (ACO) extended to continuous domains, and has been used for training neural network. However, when training neural networks, ACOR does not allow for heuristic information like most conventional ACO algorithms do. So in this work we propose a hybrid ACOR algorithm, named h-ACOR, which incorporates the heuristic information into the framework of ACOR for neural network training. The heuristic information in h-ACOR is a gradient vector obtained by computing the partial derivative of error term of the neural network with respect to weight vector. The h-ACOR is tested on training neural networks for pattern classification problems with UCI datasets: zoo, iris and tic-tac-toe. The experiments were carried out using 10-fold cross-validation method, and the results show that: h-ACOR has better performance than ACOR with almost half of convergence generations; and after completely training by h-ACOR, the average classification accuracy of datasets zoo, iris and tic-tac-toe is 92.6% while that of ACOR is 86.6%.