Plausible Inference: A Knowledge-Intensive Approach to Induction (An Application of PI to the Problem of Hidden Homology Modeling)

We adopt the perspective of learning as a user-controlled tool that assists in complex domain modeling. Using this perspective we develop Plausible Inference (PI)--a framework based on the tradeoffs suggested by this learning-as-a-tool perspective. Four aspects of the approach distinguish it from other learning approaches, and are responsible for its improved learning performance: (1) PI uses problem-decomposition information. (2) It accepts local bias information on a domain-term by domain-term basis. (3) It allows the user to fashion a task-specific learning system by composing algorithms from a library of learning algorithms, and (4) it allows the user to make efficient use of the fixed learning resource by allowing the user to interleave directly encoded and induced portions of the domain model being learned. Increased performance from these aspects of PI is measured in empirical comparisons with two well known learning algorithms: FOIL and FOCL. We use PI to draw a parallel between induction and inference, and use this parallel to illustrate how each discipline can benefit from the important ideas in the other field. Assumptions made by our tool approach are tested by encoding and using knowledge from a complex modeling task. The last third of the dissertation is devoted to the prediction of Protein Structure. Available expertise on the protein-folding process is encoded for use by PI and is tested and refined using our framework.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2015

Author dc:creator

• Oblinger, Daniel A.

Contributors dc:contributor

• Gerald DeJong

• Artificial Intelligence

Language dc:language

eng

Identifier

(MiAaPQ)AAI9904552

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/81918