%0 Thesis %1 talko:mastersthesis %A Talko, Bret %D 1999 %K algorithms, genetic programming %T A Rule-Based Approach for Constructing Neural Networks Using Genetic Programming %U http://citeseer.ist.psu.edu/312140.html %X This thesis presents a novel use of Genetic Programming (GP) to evolve recurrent, weightless neural networks. The approach taken uses neural network construction rules as the data structures that undergo adaptation by the GP algorithm. These rules can be used to construct a neural network by adding neurons and connections to an initial basic network configuration. In addition to evolving the architectures of networks, the system evolves the formulae for the activation function of each neuron in the networks and the number of processing cycles for the networks. The system has been applied to a number of Boolean functions and it is shown that solution networks were able to be found for each. Some variations in the system design were investigated on the Boolean functions to identify possible improvements that could be made to the system which would result in better performance. One variation to the system design which resulted in a significantly large increase in the system performance was made by changing the construction rules that are used by the system. A number of characteristics of the produced networks were noted. Among them is the generation of network construction rules that are similar to each other. A system variation was made which succeeded in making the rules more diverse but does not generally result in better performance. Another characteristics of the networks is that their construction rules often contain unused and redundant rules. The construction rules were designed to allow efficient specification of networks which contain multiple instances of the same sub-network. The system uses this when discovering solution networks for Boolean functions which can be decomposed into two identical Boolean functions. Importantly, the system achieved significantly better results than a modified version of the system in which the features enabling efficient network specification were not present. This suggests that incorporating a modular construction process for building networks is useful for obtaining solution networks to decomposable problems.

@mastersthesis{talko:mastersthesis, abstract = {This thesis presents a novel use of Genetic Programming (GP) to evolve recurrent, weightless neural networks. The approach taken uses neural network construction rules as the data structures that undergo adaptation by the GP algorithm. These rules can be used to construct a neural network by adding neurons and connections to an initial basic network configuration. In addition to evolving the architectures of networks, the system evolves the formulae for the activation function of each neuron in the networks and the number of processing cycles for the networks. The system has been applied to a number of Boolean functions and it is shown that solution networks were able to be found for each. Some variations in the system design were investigated on the Boolean functions to identify possible improvements that could be made to the system which would result in better performance. One variation to the system design which resulted in a significantly large increase in the system performance was made by changing the construction rules that are used by the system. A number of characteristics of the produced networks were noted. Among them is the generation of network construction rules that are similar to each other. A system variation was made which succeeded in making the rules more diverse but does not generally result in better performance. Another characteristics of the networks is that their construction rules often contain unused and redundant rules. The construction rules were designed to allow efficient specification of networks which contain multiple instances of the same sub-network. The system uses this when discovering solution networks for Boolean functions which can be decomposed into two identical Boolean functions. Importantly, the system achieved significantly better results than a modified version of the system in which the features enabling efficient network specification were not present. This suggests that incorporating a modular construction process for building networks is useful for obtaining solution networks to decomposable problems.}, added-at = {2008-06-19T17:46:40.000+0200}, author = {Talko, Bret}, biburl = {https://www.bibsonomy.org/bibtex/268135367527855ddfcb9376fead9ae4e/brazovayeye}, broken = {msc_thesis.ps.gz}, interhash = {97699e6f6948cdd5803e4098d3c3585a}, intrahash = {68135367527855ddfcb9376fead9ae4e}, keywords = {algorithms, genetic programming}, month = {March}, notes = {p129 XOR {"}Therefore Gruau's result is significantly better than the GPNN result{"}}, school = {University of Melbourne}, size = {166 pages}, timestamp = {2008-06-19T17:52:34.000+0200}, title = {A Rule-Based Approach for Constructing Neural Networks Using Genetic Programming}, url = {http://citeseer.ist.psu.edu/312140.html}, year = 1999 }