The adaptive resonance theory networks (ART-Net) is unsupervised learning system and can form a categories to classify according to the input data on a feature space. The adaptive resonance theory Map (ART-Map) are supervised learning systems based on the ART-nets. This network have flexible classification functions. In this paper, consider the basic classification ability of Radial Basis ART (RBART) using two benchmark problems. In particular, we analyze the classification capability when the input sequence changes. Next, this paper a novel ART-Map with hierarchical structure and considers its classification function. Although existing ART-Maps have two-layer structure, the novel ART-Map has three-layer structure for more efficient and flexible classification function. The ART-Map can construct more various shapes of categories than that by existing ART-Maps. Performing numerical experiments for a typical benchmark, the algorithm efficiency is investigated.

114 工学部 (旧) ・工学研究科 ＞ 学位論文 ＞ 修士論文