Performance comparison of various MLPs for material recognition based on sonar data

Abstract

Sonar data from the UCI Machine Learning Repository database has large input features. It is known that too many input features have high tendency for redundant data and difficult to be handled by Multilayer Perceptron (MLP). This paper proposes the integration between MLP and circle-segments method for material detection based on sonar data. Circle-segments is a data visualization methods useful for feature selection to the reduce number of inputs but yet closely maintain the integrity of original data. The proposed method has been compared with MLP without feature selection. The results show that the MLP trained without feature selection obtains higher percentage of correct classification compared to MLP trained with the circle-segments feature selection data.