Related Experiment Videos
Connectionist-based Dempster-Shafer evidential reasoning for data fusion.
Otman Basir1, Fakhri Karray, Hongwei Zhu
1Machine Intelligence Research Group, Department of Electrical and Computer Engineering, University of Waterloo, Waterloo, ON N2L 3G1, Canada. obasir@uwaterloo.ca
IEEE Transactions on Neural Networks
|December 14, 2005
Summary
This study introduces a novel neural network approach, DSETNN, to enhance Dempster-Shafer evidence theory (DSET) for pattern classification. DSETNN effectively addresses uncertainty and source dependence, improving accuracy and learning speed.
Area of Science:
- Artificial Intelligence
- Machine Learning
- Pattern Recognition
Background:
- Dempster-Shafer Evidence Theory (DSET) offers a robust framework for managing uncertainty and imprecision.
- Real-world applications of DSET are limited by challenges with basic probability assignments (masses) and information source dependence.
- Neural networks present a potential solution for overcoming these limitations in evidential reasoning.
Purpose of the Study:
- To develop a novel neural network framework, DSETNN, that integrates with Dempster-Shafer Evidence Theory.
- To address the issues of basic probability assignments and source dependence within the DSET framework.
- To improve the performance of evidential reasoning in pattern classification tasks.
Main Methods:
- A multilayer perceptron neural network was used to compute posterior probabilities for each class from information sources.
- A novel evidence structure construction scheme was developed to translate posterior probabilities into DSET masses and focal elements.
- A DSET-based neural network (DSETNN) was designed and optimized via supervised learning to handle evidence structures and source dependence.
Main Results:
- The proposed DSETNN demonstrated superior performance compared to traditional DSET in pattern classification tasks.
- Experiments on benchmark problems showed significant improvements in classification accuracy.
- DSETNN exhibited faster learning convergence rates compared to existing methods.
Conclusions:
- The DSETNN framework effectively addresses key limitations of Dempster-Shafer Evidence Theory, particularly concerning basic probability assignments and source dependence.
- The integration of neural networks with DSET offers a promising approach for enhancing evidential reasoning in complex applications.
- The proposed method provides a powerful tool for pattern classification, delivering improved accuracy and efficiency.