Design of a life-long learning system based on multiple probabilistic fuzzy models

Abstract

Nowadays, increasing attention is being paid to learning capability, especially in robotics area, as a research direction for enhancing machine intelligence toward human intelligence. Note that approaches of analysis and design for learning systems are diverse, depending on application domains with various notions on learning. In general, a learning system structure and its learning algorithm are selected based on analysis of the learning target in consideration of the target\\`s environment. However, determination of a learning system becomes painstaking when human behavior is the subject of the learning target due to its complex characteristics (i.e., high dimensionality, nonlinear-coupling of attributes, subjectivity, apparent inconsistency, susceptibility to environmental noise and disturbances, and time-variance as well as situation-dependency). In this thesis, in particular, a life-long learning system is presented based on multiple probabilistic fuzzy models, to handle inconsistent/time-varying data effectively such as human behavioral pattern. First, a new iterative fuzzy clustering algorithm with a supervisory scheme has been proposed to derive a probabilistic fuzzy rule base from numerical data with inconsistency, in view of both classification and probabilistic interpretation. Note that the concept of probability incorporated into the fuzzy logic provides useful information about the certainty factor in deciding output values. The learning process starts in a fully unsupervised manner using a fuzzy clustering algorithm and a cluster validity criterion. Then it gradually constructs meaningful fuzzy partitions over input space and obtains corresponding fuzzy rules with probabilities through an iterative learning process of selective clustering with supervisory guidance based on cluster-pureness and class-separability. Secondly, an effective life-long learning system has been proposed to deal with time-varying and possibly periodic/repeated da...