Determining Multi-layer Perceptron Structure Using Clustering Techniques
Mohamed Lafif Tej, Ştefan Holban · 2019
This paper presents a new method to determine the optimal Multi-layer Perceptron (MLP_structure using pattern recognition and data mining. Using Clustering techniques on the dataset used to train the neural network and based on a given criterion, we can define a number of clusters. The results obtained from clustering of training dataset can be used as an indicator to determine the level of complexity of the problem to be solved. Based on the number of clusters obtained we can determine the number of hidden layers for an MLP. Different types of training datasets are taken into consideration such as interval, ordinal, categorical or mixture of different types of variables, which affect the distance measures to recognize clusters in a dataset. The experience from using these different types gives generalization capabilities to the proposed method. Using this method avoids making a complicated neural network structure by setting a more than required number of hidden layers. On the contrary, using few numbers of hidden layers makes the network unable to achieve a satisfactory performance, and this happens in the case of miscalculation of the level of complexity of the problem. Unsupervised training is considered in relation with the proposed method.