A Review Of Mf-Artmap Toward An Improvement Classification Accuracy Using Simulated Annealing

This paper deals with an MF ARTMAP neural network. We study its behavior while training with different data sets and using different parameters. It gives us better knowledge of its strong and weak points. Subsequently, we focus on alleviation of weak points and improvement of strong points like the utilization of a one-shot learning, an incremental ability of the network without forgetting the already obtained knowledge or post-processing of information stored in the form of the transparent internal structure of identified clusters and classification classes. We have shown the incrementality of this neural network. As for the weak part of the MF ARTMAP algorithm, we try to increase the generalization ability by adopting Simulated Annealing method to find the best shape of membership functions with the best possible ratio between generalization of the neural network and its classification performance. Using simulated annealing algorithm, we optimize network's parameters namely the membership function's shapes of fuzzy clusters in the feature space. Subsequently, we compare classification accuracy of MF ARTMAP with and without parameters optimization, as well. Moreover, we compare against the classification precision of the Multi-Layer Perceptron (MLP) using benchmark data sets, with the aim to get a relevant image of the overall MF ARTMAP efficiency beside the well-known and frequently-used algorithm, like the MLP.