Learning of SOR network employing soft-max adaptation rule of neural gas network Takanori Koga * , Keiichi Horio, Takeshi Yamakawa Graduate School of Life Science and Systems Engineering, Kyushu Institute of Technology, 2-4 Hibikino, Wakamatsu, Kitakyushu, 8080196, Japan Abstract. A Self-Organizing Relationship Network (SORN) can approximate the desirable input/ output (I/O) relationship of a target system from not only good examples but also bad ones. The learning of SORN is achieved with employing the soft-max adaptation rule of Self-Organizing Maps (SOM). In this paper, we simplify the learning law by employing the soft-max adaptation rule of Neural Gas Network. This modification improves the approximation performances and lightens burdens imposed on a network designer in the design process of SORN. D 2006 Elsevier B.V. All rights reserved. Keywords: Self-Organizing Relationship (SOR) Network; Vector quantization; Soft-max adaptation rule; Neural Gas Network 1. Introduction Self-Organizing Relationship Network (SORN) is a neural network, which can approximate the desirable I/O relationship of a target system from not only good examples but bad ones [1]. The learning and the execution process of SORN are shown in Fig. 1. The learning of SORN (Fig. 1(a)) is achieved in a manner similar to Self- Organizing Maps (SOM) [2]. In the learning of SORN, in this regard, learning data are evaluated by the designer’s subjective criteria or the objective ones before the learning. Then, the learning is attractively or repulsively achieved according to the evaluation. In the execution mode, SORN works as a fuzzy inference engine (Fig. 1(b)). When SORN is used as a nonlinear controller, the accuracy of the vector quantization in the learning should be improved without needing the knowledge layer. This is because SORN decides 0531-5131/ D 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.ics.2006.01.052 * Corresponding author. Tel.: +81 93 695 6123; fax: +81 93 695 6133. E-mail address: koga-takanori@edu.brain.kyutech.ac.jp (T. Koga). International Congress Series 1291 (2006) 165 – 168 www.ics-elsevier.com