| Summary: | 碩士 === 國立中山大學 === 電機工程學系研究所 === 101 === Multi-valued Neuron with Periodic activation function (MVN-P) was proposed by Aizenberg for solving classification problems. The boundaries between two distinct categories are crisply specified in MVN-P, which may result in slow convergence or being unable to converge at all in the learning process. In this paper, we propose two revised models of MVN-P based on the idea of un-sharp boundaries. In the first revised model, a crisp buffer is provided around a boundary between two distinct categories, allowing incorrect assignments in the buffer to be tolerated in the training phase. In the second revised model, a fuzzy buffer is provided instead and an incorrect assignment with membership degree less than a Threshold can be tolerated. Genetic algorithms are applied to derive optimal values for the parameters involved in different models, alleviating the burden of setting them manually by the user. Besides, MVN-P has difficulties solving the classification problems having a large number of categories. A tree structure is developed to overcome these difficulties. Simulations have been done and the results are presented to demonstrate the effectiveness of our proposed ideas.
|
|---|
