General pattern learning and recognition using genetically-optimized training of a biased ARTMAP ensemble voting system

Sayeed, M. S. and Liew, W. S. and Loo, C. K. (2012) General pattern learning and recognition using genetically-optimized training of a biased ARTMAP ensemble voting system. International Journal of Innovative Computing, Information and Control (IJICIC), 8 (11). pp. 7543-7560. ISSN 1349-418X

[img] Text
11.pdf
Restricted to Repository staff only

Download (489kB)
Official URL: http://www.ijicic.org/vol-8%2811%29.htm

Abstract

This paper is an attempt to create a method and system for generating an optimal machine-based pattern recognition system. If successful, this allows any given classifier to improve its classification accuracy by introducing a genetic optimization pre-processing method combined with a probabilistic ensemble voting system. Several problems were identified, and the solutions proposed were based on literature review on similar fields of research. The proposed system utilizes a Fuzzy ARTMAP variant, Biased ARTMAP, as the core pattern learning and classification method for extracted features due to its ability for incremental learning and a biasing parameter which improves its online learning capability over the traditional Fuzzy ARTMAP. One weakness in the ARTMAP system is the effect of the training data sequence on the ARTMAPs learning processes, and consequently, its classification accuracy. A genetic permutation method is proposed to solve this problem by optimizing the training data sequence over several generations of genetic mating and mutation operations. The best training sequences are selected to train multiple Biased ARTMAPs and combined in a probabilistic voting system to determine the final class prediction. Classification performance of the voting system can be improved by implementing a reliability threshold to filter unreliable predictions from the final results. Genetic optimization of the training process combined with the probabilistic voting system improved the Biased ARTMAPs classification accuracy to 75% - 87%, up from 67% using only the Biased ARTMAP system.

Item Type: Article
Subjects: T Technology > T Technology (General)
Divisions: Faculty of Information Science and Technology (FIST)
Depositing User: Ms Rosnani Abd Wahab
Date Deposited: 10 Jan 2014 02:29
Last Modified: 10 Jan 2014 02:29
URI: http://shdl.mmu.edu.my/id/eprint/4778

Actions (login required)

View Item View Item