Abstract
The backpropagation neural network algorithm is a popular machine learning methodology. One difficulty with using the algorithm is that the network architecture must be designed before learning can occur. This includes deciding the number of layers, the number of nodes in each layer, and the number of connections. Traditionally this problem is solved by heuristics gained by an expert through experience.
This paper presents an algorithm for using ID3 decision trees to design the network architecture. Previous approaches rely on binary decision trees. The proposed technique utilizes n-ary trees. These trees are easier to understand, are better suited for continuous-valued attributes, and may be more generally applicable than their binary tree counterparts.
The proposed method was tested on both artificial and real domains. The results were compared to traditional networks designed by an expert and to the original ID3 trees from which they were designed. The ID3-designed neural networks demonstrated learning accuracies comparable to the traditional networks and consistently superior to the original ID3 trees.
Recommended Citation
M., B. Van Horn and St. Clair, D. C., "Design of BackPropagation Neural Network Architectures using a Decision Tree Classifier" (1992). Computer Science Technical Reports. 140.
https://scholarsmine.mst.edu/comsci_techreports/140
Department(s)
Computer Science
Second Department
Mathematics and Statistics
Report Number
CSc-92-11
Document Type
Technical Report
Document Version
Final Version
File Type
text
Language(s)
English
Rights
© 1992 University of Missouri - Rolla, All rights reserved
Publication Date
1 May, 1992
Comments
The first Author is a Graduate Student.
This report is substantially the M.S. thesis of the first author, completed May, 1992.
This thesis was prepared in accordance with the format required for publication in the Communications of the Association for Computing Machinery (ACM).