Dissertations and Theses @ UNI

Availability

Open Access Thesis

Keywords

Algorithms; Machine learning;

Abstract

This study compares the effectiveness of inductive and analytical learning techniques for a specific type of problem. The set of problems has a large number of parameter combinations as inputs, and a small set of possible output solutions. The card game, Euchre, was the basis for research. Four algorithms were used for comparison. Two were variants of induction using decision tree learning, and the other two used variants of explanation-based learning, an analytical approach. The example data were gathered from human players, along with one set of random, "noisy" data. These data were then used as training sets for the four learners. Each of the training sets contained 90 examples, with an additional 30 used for testing the learning system. In addition to self-testing, each of the four systems was pitted against each other in simulated games to see which was the most effective in playing. The results demonstrate that the analytical approaches benefit greatly from knowing the domain rules previous to learning. This enabled the learner to focus only on learning the strategy behind each play. The inductive approaches proved unable to learn these domain rules. They were unable to identify the difference between a game rule and a strategic one. Therefore, the rules induced were a meld of the two, and the rules generated caused illegal choices, resulting in losing hands. The inductive learners were able to overcome this limitation by being forced to follow the rules. When not allowed to play an illegal card, the inductive learners performed comparably to the analytic ones. This indicates that the overhead, and increased processing required by the analytical approaches, offers no benefit in problems of this type.

Year of Submission

2001

Degree Name

Master of Science

Department

Department of Computer Science

First Advisor

Eugene Wallingford

Second Advisor

Mark Fienup

Third Advisor

Jack Yates

Comments

If you are the rightful copyright holder of this thesis and wish to have it removed from the Open Access Collection, please submit a request to scholarworks@uni.edu and include clear identification of the work, preferably with URL.

Date Original

2001

Object Description

1 PDF file (42 leaves)

Language

en

File Format

application/pdf

Share

COinS