Author:  Ling, Saiho 
Title:  Realcoded genetic algorithm based variable feedforward neural networks and their applications 
Degree:  Ph.D. 
Year:  2007 
Subject:  Hong Kong Polytechnic University  Dissertations. Neural networks (Computer science) Evolutionary programming (Computer science) Genetic algorithms. 
Department:  Dept. of Electronic and Information Engineering 
Pages:  xxxiii, 261 leaves : ill. ; 31 cm. 
Language:  English 
InnoPac Record:  http://library.polyu.edu.hk/record=b2093882 
URI:  http://theses.lib.polyu.edu.hk/handle/200/2304 
Abstract:  This thesis focuses on the realcoded genetic algorithm and different topologies of feedforward neural networks. Results in the following areas will be reported: (1) a realcoded genetic algorithm with new crossover and mutation operations, and its applications; (2) three different topologies of variable feedforward neural networks, and their applications to shortterm electric load forecasting in Hong Kong and handwritten graffiti recognition. The realcoded genetic algorithm (RCGA) is one evolutionary computation technique that can tackle complex optimization problems. In this thesis, RCGA with new genetic operations called the averagebound crossover (ABX) and wavelet mutation (WM) will be presented. On realizing the ABX operation, the offspring spreads over the domain so that a higher chance of reaching the global optimum can be obtained. Taking advantage of the wavelet theory, the performance of the mutation operation in terms of the cost function value, solution stability (standard deviation of solutions) and convergence rate is improved. A suite of benchmark test functions are used to evaluate the performance of the proposed algorithm. Also, the sensitivity of the parameters in WM and the sensitivity of the genes' initial range to the searching performance of the proposed RCGA will be discussed. Application examples on economic load dispatch and tuning the parameters of neural networks are used to show the merits of the proposed RCGA. The three proposed topologies of variable feedforward network networks are: (1) the variablestructure neural network (VSNN), (2) the variableparameter neural network (VPNN), and (3) the variablenodetonodelink neural network (VN2NN). By taking advantage of these networks' structures, the learning and generalization abilities of the networks can be increased. All the network parameters are tuned by the proposed RCGA with ABX and WM. The proposed VSNN consists of a neural network with link switches (NNLS) and a network switch controller (NSC). In the NNLS, switches are introduced in the links between the hidden and output layers. By using the NSC to control the onoff states of the switches in the NNLS, the proposed neural network can model different input patterns with different network structures. In other words, it operates with a variable structure to handle different sets of input patterns. Consequently, the VSNN offers better results and an enhanced learning ability. In the proposed VPNN, the parameters will adaptively tackle different sets of input data sparsely located in a large domain. Each input data set is effectively handled by its corresponding set of network parameters. The network operates as if there is a corresponding neural network for each set of input data. This makes the VPNN exhibit a better learning and generalization ability than the traditional neural networks. In the proposed VN2NN, the parameters of the activation functions in the hidden layer change with respect to the network inputs. Nodetonode links are introduced in the hidden layer and the forms of the links are tuned by the RCGA. The introduction of the nodetonode links increases the degree of freedom of the network's modelling ability, enabling the network to provide better learning and generalization abilities. Application examples on shortterm electric load forecasting and handwritten graffiti recognition are given to illustrate the merits of the proposed neural networks. Shortterm electric load forecasting is important to power systems for their economic, reliable and secure operation. The nonlinearities and largescale characteristics of the problem require a welldesigned neural network to tackle it. Handwritten graffiti recognition is another problem demanding the superior learning and generalization abilities of the proposed neural networks to do classifications with respect to a large number of input data sets. Based on the results from the two application examples, the performance characteristics of the three proposed networks will be investigated and explained. Comparisons among the networks will be conducted in order to let the user have an idea on how to choose which network to use for different kinds of problems. 
Files  Size  Format 

b20938822.pdf  29.45Mb 


As a bona fide Library user, I declare that:  


By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms. 