Research On The RBFNN For Complex Classifications

Radial Basis Function Neural Network (RBFNN) can be viewed as a three-layerfeed-forward neural network, which has drawn much attention due to its goodgeneralization ability and fast convergence. And it has been successfully used in manyfields, such as system identification and data mining. Aiming at improving the abilityof the RBFNN for complex classification, some researches based on RBFNN havebeen studied in this thesis substantially as follows:A three-phase RBFNN learning algorithm is proposed in this thesis, whichcombines the decayed radius selection clustering (DRSC) method with the sumsquared error (SSE) rule, i.e., an extra tuning process. Firstly, the initial hiddenstructure of the network is determined by DRSC, and the minimum decayed radius forclustering is adjusted simultaneously rather than being kept fixed, which improves theself-adapting ability of the minimum radius and prevents many trials before getting tothe optimum network, and avoids the over-learning as the radius decreases infinitely.Then the positions of the hidden layer centers are adjusting slightly with the SSE ruleby considering the effect when a certain samples are moved from one class to another.The within-cluster and between-cluster distances are here employed to calculate thecorresponding radius widths. The between-cluster distances are taken into accountespecially in this thesis, which are able to shrink the region of overlap. Finally thepseudo-inverse algorithm is utilized to train the weights between the hidden layer andthe output layer. The experiments are implemented on Iris, Wines and Glass datasets,which show that the proposed RBFNN training algorithm has a better classificationability compared with the conventional methods.An adaptive learning algorithm, e.g. GA-RBFNN, is presented, to build aRBFNN model by employing the Genetic Algorithm (GA)'s parallel-search abilitywith the aim at improving the classification accuracy of the RBFNN. Firstly, theinitial network hidden structure of a RBFNN model is determined by the three-phraselearning algorithm. Then the hidden centers of a RBFNN are modified by a speciallydesigned GA, which is based on the matrix-form mixed encoding scheme with acontrol vector for regulating the structure of a RBFNN, and the new genetic operatorsare presented correspondingly. The pseudo-inverse algorithm is adopted to train theweights. Finally, experiments are implemented on datasets as Iris, Wines, and Glass,which show that the proposed algorithm has performed better compared with theconventional methods.