MARTIAN: A concept learning system.

Abstract

This thesis introduces an empirical concept learning system, MARTIAN. MARTIAN learns concepts incrementally. An arbitrary number of concept representations can be formed. MARTIAN is designed to be a robust concept learning system. It can cope with missing data, incorrect data, and concepts that change their constituencies over time. Instances of concepts are described using attributes that take on discrete values. The description language is hierarchical in the sense that values of attributes can contain an embodiment of attributes and values. Such a description language has several benefits. These benefits include the ability to acquire concepts faster and the ability to transfer learned knowledge to new situations. Different concepts can be described by sets of attributes that fully overlap, partially overlap, or are completely disjoint. In order to represent concepts, MARTIAN uses both a numeric and a symbolic representation. The numeric representation is based upon conditional probabilities. The symbolic representation introduces a new concept learning data structure, the hypothesis space. The hypothesis space is approximately equivalent in representational power to a Boolean description language (negation can only be represented indirectly) but can converge upon accurate concept representations using many fewer training examples. Concepts are represented using a lazy approach: the symbolic representation is only invoked if the numeric representation seems to be insufficient. MARTIAN is evaluated using theoretical analysis, empirical evidence, and through comparison with existing concept learning systems (STAGGER and ID3).