Fuzzy neural networks with genetic algorithm-based learning method
This thesis is on the reasoning of artificial neural networks based on granules for both crisp and uncertain data. However, understanding the data in this way is difficult when the data is so complex. Reducing the complexity of the problems that these networks are attempting to learn as well as decr...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | English |
| Published: |
2013
|
| Subjects: | |
| Online Access: | http://eprints.utm.my/id/eprint/35849/5/MRezaMashinchiPFSKSM2013.pdf http://eprints.utm.my/id/eprint/35849/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | This thesis is on the reasoning of artificial neural networks based on granules for both crisp and uncertain data. However, understanding the data in this way is difficult when the data is so complex. Reducing the complexity of the problems that these networks are attempting to learn as well as decreasing the cost of the learning processes are desired for a better prediction. A suitable prediction in artificial neural networks depends on an in-depth understanding of data and fine tracking of relations between data points. Inaccuracies of the prediction are caused by complexity of data set and the complexity is caused by uncertainty and quantity of data. Uncertainties can be represented in granules, and the reasoning based on granules is known as granular computing. This thesis proposed an improvement of granular neural networks to reach an outcome from uncertain and crisp data. Two methods based on genetic algorithms (GAs) are proposed. Firstly, GA-based fuzzy granular neural networks are improved by GA-based fuzzy artificial neural networks. They consist of two parts: granulation using fuzzy c-mean clustering (FCM), and reasoning by GAbased fuzzy artificial neural networks. In order to extract granular rules, a granulation method is proposed. The method has three stages: construction of all possible granular rules, pruning the repetition, and crossing out granular rules. Secondly, the two-phase GA-based fuzzy artificial neural networks are improved by GA-based fuzzy artificial neural networks. They are designed in two phases. In this case, the improvement is based on alpha cuts of fuzzy weight in the network connections. In the first phase, the optimal values of alpha cuts zero and one are obtained to define the place of a fuzzy weight for a network connection. Then, in the second phase, the optimal values of middle alpha cuts are obtained to define the shape of a fuzzy weight. The experiments for the two improved networks are performed in terms of generated error and execution time. The results tested were based on available rule/data sets in University of California Irvine (UCI) machine learning repository. Data sets were used for GA-based fuzzy granular neural networks, and rule sets were used for GA-based fuzzy artificial neural networks. The rule sets used were customer satisfaction, uranium, and the datasets used were wine, iris, servo, concrete compressive strength, and uranium. The results for the two-phase networks revealed the improvements of these methods over the conventional onephase networks. The two-phase GA-based fuzzy artificial neural networks improved 35% and 98% for execution time, and 27% and 26% for the generated error. The results for GA-based granular neural networks were revealed in comparison with GA-based crisp artificial neural networks. The comparison with other related granular computing methods were done using the iris benchmark data set. The results for these networks showed an average performance of 82.1%. The results from the proposed methods were analyzed in terms of statistical measurements for rule strengths and classifier performance using benchmark medical datasets. Therefore, this thesis has shown GA-based fuzzy granular neural networks, and GA-based fuzzy artificial neural networks are capable of reasoning based on granules for both crisp and uncertain data in artificial neural networks. |
|---|