Extraction of Embedded Hierarchical Structure of Knowledge from Trained Boltzmann Machine by Genetic algorithm
拓 奥野, Yukinori Kakazu · Institutional Repositories DataBase (IRDB) · 1994
This study focuses on knowledge representa-tion as based on the distributed representation by the Boltzmann machine. In this type of framework, concepts are represented uniformly without any information on their structural relationship. To utilize such information, this study explores how the hierarchical structure of concepts can be extracted from the weights of a trained network. The hierarchy dealt with here is a taxonomical one which is represented by a tree of microfeatures of a concept. Here, the tree is extracted by using various kinds of local criteria relating to weight distribution. To satisfy them simultaneously, a Genetic Algorithm is adopted.