Abstract: The ACOR algorithm is an ant colony optimization(ACO) algorithm for real-valued optimization.The ACOR can be used for training neural network.Unlike most of the conventional ACO algorithms,ACOR does not consider heuristic information when training neural networks.So in this work,a new algorithm named h-ACOR that incorporates the heuristic information into the framework of ACOR was proposed for neural network training.The heuristic information in h-ACOR is a gradient vector,which is obtained by computing the partial derivative of error term of the neural network with respect to weight vector.Using 10-fold cross-validation method,h-ACOR is applied to train neural networks for pattern classification problems of zoo,iris and tic-tac-toe in UCI datasets.Compared with ACOR,h-ACOR can reduce classification errors while speeding up the convergence process,with the average training generations of h-ACOR being nearly 1/2 of that of ACOR.After completely training by h-ACOR,the classification accuracy about zoo,iris and tic-tac-toe are respectively 91.1%,93.3% and 95.6%,which have better performance than that of ACOR’s 83.1%,88.7% and 91.9%.

Key words: ACO,Heuristic information,Artificial neural network,Neural network training