Agent based solution for artificial neural network optimisation

Abstract

Artificial neural networks are highly used in the areas of pattern recognition, feature extraction, function approximation, scientific classification, control systems, noise reduction and prediction. Feed-forward and back-propagation neural networks are the most commonly used artificial neural networks. Many researchers face difficulties when selecting a proper ANN architecture and training parameters. The manual ANN training process is not the best practical solution because it is a much time consuming task. Also most of the people conduct the manual process in an ad-hoc manner without having a proper knowledge about artificial neural networks. At the end of this research project a multi-agent system: MASAnnt (Multi Agent System for Artificial Neural Network Training) was developed to automate the neural network training for feed-forward and back-propagation neural network. Interaction among agents enables emergence of quality training sessions which cannot be archived by an ad-hoc training sessions conducted by humans. It is straight forward to recognize training parameters such as number of hidden layers, number of neurons in each hidden layer, momentum, learning rate, Emax (Error goal) and activate function of an ANN as a set of agents. Inherent features of agents including coordination, communication and negotiation are able to mimic the ANN optimizing and training process by manipulating theses parameters. Our experiments show that the more rational results can be obtained from the system with both simple data sets like XOR as well as with real life data sets. We can conclude that the neural network optimization and training tasks are successfully accomplished by the agent based approach by analysing the results of the evaluation.