Study Of Artificial Bee Colony Algorithm And Its Application In Artificial Neural Network

Artificial neural network (ANN) is one of the important branches in the artificialintelligence fields. ANNs are the nature-based computing techniques which have been appliedwidely in tasks such as controls, prediction, optimization, system identification, signalprocessing and pattern recognition, etc. In view of the shortcomings of traditional neuralnetwork training methods, many researchers begin to apply the heuristic algorithm such asgenetic algorithm and particle swarm algorithm into the structure design of neural networkand the optimization of the weights and thresholds so as to solve questions they faced. In thispaper the artificial bee colony algorithm and artificial neural network will be combinedtogether. The artificial bee colony algorithm is used to optimize the mean square errorfunction of the artificial neural network and get the best weights and thresholds of thenetwork.Artificial bee colony algorithm is a new meta-heuristic algorithm, which simulates theprocess of gathering honey by bees in the nature to solve optimization problems. It has severaladvantages such as less control parameters, simple calculation and easy to implement, it alsohas strong global search capability, so it attracts more and more attention since it was raised.In this paper the algorithm is used to optimize the weights and thresholds of the neuralnetwork. We have done some research and made three improvements for this algorithm. First,the bees searching behavior has been improved, including restrain the poor food sources asneighbor bees and accelerate movement if it finds high quality food sources. Second, the poorfood sources are replaced for every certain period in the space which is dynamic adjusted bythe evolutionary process, so that it maybe improves the sources’ quality by resetting. Last, theway of scout bees appearing has been improved. If a food source’s retention time exceeds thepreset number, it was abandoned and reset. When it was reset, it will be distinguished for itsposition in the population. If it is the global optimal, it will be reset by the opposition baseslearning algorithm, otherwise it will be crossed with the food source which is better thanitself.In the practical application of classification, the structure of single neural network whichis built to classify the samples is very complicated and the training time of the network also will increase dramatically when the problem has a lot of categories. If the output nodes aretoo many, it will increase the coupling of the output and reduce the classification accuracy. Sothe article proposes to build multiple sub-networks instead of single neural network with thehelp of modular. The number of the sub-networks equals to the categories of the samples, andeach sub-network responses for a two classification problems, that is to say it has only oneoutput node, so the structure of the sub-networks is definitely simpler than the single neuralnetwork’s structure. The sub-networks are parallel running when they are used, so the trainingtime of the whole modular neural network is less. At last, the outputs of all sub-networks aresynthesized and the class of one sample is determined, which will improve the networktraining accuracy and testing accuracy.