統計的相関に基づく動作パターンのリアルタイム教師なし分節化と原始シンボルの自律的獲得

Abstract

Mimesis is the hypothesis that human intelligence originated in the interactive communication of motion recognition and generation through imitation. This is attractive for artificial intelligence. We have developed a mimesis system using Hidden Markov Models (HMMs) and their parameter sets are defined as the proto symbols. In conventional systems, designers have to segment a motion pattern in sequencial motion data to embed the motion pattern in an HMM. However it is necessary to have the ability of motion pattern segmentation in order to autonomously learn and develop through imitation. In this paper, we propose a motion segmentation method that consists of three phases. In the first phase short sequences of motions are encoded. In the second phase the correlation matrix of the encoded sequences are computed. In the third phase motion patterns are segmented based on error between the encoded sequences observed and predicted from the correlation matrix. Moreover we show that it is possible to acquire the proto symbols by providing the mimesis system with the segmented motion patterns.