Proposed Multi Agents Classification System using Neural Networks

Multiagents technology took a significant role in the field of decision making and machine learning for solving complex problems in the real world. They simulate human ability to decision making where they have ability to reasoning and behave autonomously to solve problems or to support human user.In this thesis, a classification system using hierarchal multi agent's technology based on neural network and fuzzy logic is introduced,where each agent is implemented as a neural network (trained using back propagation learning algorithm). The system classifies a collection of datasets with some degree of generalization.
The system consists of two layers of agents. The top layer contains one agent working as control agent. Its responsibility is to select the right agent from the agents in the bottom layer to classify the related
pattern depending on features of data. If the selected pattern do not recognized, then it is declared as unknown pattern. The developed system was tested using different standard datasets obtained from the University of California, Irvine (UCI) these are User Knowledge Level, Iris, and Banknote Authentication datasets. Earlier stopping criterion and regularization techniques were used to estimate the generalization of the agents. The final results indicated that the best generalization technique for user knowledge dataset is regularization with cross validation selection mode, while for both iris dataset and banknote authentication is the earlier stopping criterion,also accuracy of testing each classification agent as fallowing ,agent1 is 97.53% , agent2 and agent3 100%, agent4 69.04%.