Files in this item
|(no description provided)|
|Title:||Learning flexible concepts from examples: Employing the ideas of two-tiered concept representation|
|Doctoral Committee Chair(s):||Michalski, R.S.|
|Department / Program:||Computer Science|
|Degree Granting Institution:||University of Illinois at Urbana-Champaign|
|Abstract:||This thesis describes an exploration of methods involved in learning flexible concepts that is an important and less explored area in machine learning. The two approaches described in this thesis are based on the idea of two-tiered (TT) concept representations. In the TT representation, the first tier, called the Base Concept Representation (BCR), contains an explicit description of core concept properties, and the second tier, called the Inferential Concept Interpretation (ICI), defines allowable modifications of the explicit meaning and its dependence on the context of discourse.
The first approach generates a BCR of a concept by first creating a complete and consistent (CC) description from supplied training examples, and then optimizing the description according to a general description quality (GDQ) criterion. The CC description is obtained by applying the AQ inductive learning methodology. The optimization process is done by a double-level best-first search, using the SG-TRUNC description reduction process. The ICI consists of a user-specified procedure of flexible matching and a set of inference rules. This approach has been implemented in the POSEIDON learning system, and experimentally tested on two different real-world problems.
In the second approach, a concept is presented as a set of extended complexes and a set of exemplars. Each extended complex is a generalized description and consists of a base complex (equivalent to a disjunct), a similarity measure, a threshold, and a set of weights. Accordingly, the task of learning a concept is to create a set of extended complexes and a set of exemplars. An extended complex is generated in two phases: base complex generation and extended complex optimization. When generating an extended complex, two evaluation functions, quality evaluation function and potential quality evaluation function are applied. This approach has been implemented in FCLS that is tested in various domains including artificial and real world domains.
|Rights Information:||Copyright 1990 Zhang, Jianping|
|Date Available in IDEALS:||2011-05-07|
|Identifier in Online Catalog:||AAI9114481|