The adaptive resonance theory networks (ART-nets) are typical unsupervised learning systems having feature extraction function of objective environments. In order to improve the algorithm performance, several variants have been studied: hypersphere ART-nets, ellipsoid ART-nets, radial basis ART-nets, etc. The adaptive resonance theory maps (ART-maps) are supervised learning systems based on the ART-nets and have flexible classification functions. Applications of the ART-nets and ART-maps are many, including image-processing, adaptive anomaly detection and combinatorial optimization problems. This paper presents a radial basis ART-map (RB-ARTM) and its application to classification problems. The RB-ARTM is based on the standard ART-maps and the radial basis ART-nets. In order to classify input data with class information, the RB-ARTM generates categories characterized by center, radius and class. The learning algorithm includes two key parameters: vigilance parameter ρ and distance parameter k. These parameters can control the number and size of categories. Using basic three benchmarks, we have confirmed that our algorithm is competitive at classification function with Fuzzy-ARTMAP and support vector machine (SVM). And we investigate the algorithm performance in basic numerical experiments for two example problems: classification of typical benchmarks, and classification of abdomen shapes of expectant mothers.

114 工学部 (旧) ・工学研究科 ＞ 学位論文 ＞ 修士論文