%0 Thesis %1 luke:dissertation %A Luke, Sean %C A. V. Williams Building, University of Maryland, College Park, MD 20742 USA %D 2000 %K algorithms, genetic programming %T Issues in Scaling Genetic Programming: Breeding Strategies, Tree Generation, and Code Bloat %U http://www.cs.gmu.edu/~sean/papers/thesis2p.ps.gz %X Genetic Programming is an evolutionary computation technique which searches for those computer programs that best solve a given problem. As genetic programming is applied to increasingly difficult problems, its effectiveness is hampered by the tendency of candidate program solutions to grow in size independent of any corresponding increases in quality. This bloat in solutions slows the search process, interferes with genetic programming's searching, and ultimately consumes all available memory. The challenge for scaling up genetic programming is to find the best solutions possible before bloat puts a stop to evolution. This can be tackled either by finding better solutions more rapidly, or by taking measures to delay bloat as long as possible. This thesis discusses issues both in speeding the search process and in delaying bloat in order to scale genetic programming to tackle harder problems. It describes evolutionary computation and genetic programming, and details the application of genetic programming to cooperative robot soccer and to language induction. The thesis then compares genetic programming breeding strategies, showing the conditions under which each strategy produces better individuals with less bloating. It then analyzes the tree growth properties of the standard tree generation algorithms used, and proposes new, fast algorithms which give the user better control over tree size. Lastly, it presents evidence which directly contradicts existing bloat theories, and gives a more general theory of code growth, showing that the issue is more complicated than it first appears.

@phdthesis{luke:dissertation, abstract = {Genetic Programming is an evolutionary computation technique which searches for those computer programs that best solve a given problem. As genetic programming is applied to increasingly difficult problems, its effectiveness is hampered by the tendency of candidate program solutions to grow in size independent of any corresponding increases in quality. This bloat in solutions slows the search process, interferes with genetic programming's searching, and ultimately consumes all available memory. The challenge for scaling up genetic programming is to find the best solutions possible before bloat puts a stop to evolution. This can be tackled either by finding better solutions more rapidly, or by taking measures to delay bloat as long as possible. This thesis discusses issues both in speeding the search process and in delaying bloat in order to scale genetic programming to tackle harder problems. It describes evolutionary computation and genetic programming, and details the application of genetic programming to cooperative robot soccer and to language induction. The thesis then compares genetic programming breeding strategies, showing the conditions under which each strategy produces better individuals with less bloating. It then analyzes the tree growth properties of the standard tree generation algorithms used, and proposes new, fast algorithms which give the user better control over tree size. Lastly, it presents evidence which directly contradicts existing bloat theories, and gives a more general theory of code growth, showing that the issue is more complicated than it first appears.}, added-at = {2008-06-19T17:35:00.000+0200}, address = {A. V. Williams Building, University of Maryland, College Park, MD 20742 USA}, author = {Luke, Sean}, biburl = {https://www.bibsonomy.org/bibtex/2a023c9b72a247ab465be67766a15d13e/brazovayeye}, interhash = {5a51d847aa1e8b0ee336e683ad739520}, intrahash = {a023c9b72a247ab465be67766a15d13e}, keywords = {algorithms, genetic programming}, notes = {errata 1. In Algorithm 2 (p. 6), the line P<-P\\[q] should read P<-P\\[s]. 2. Figures 5.2 through 5.5 (p. 38-39) are not in proper evolutionary-time order. The proper order is 5.4, 5.5, 5.2, 5.3.}, school = {Department of Computer Science, University of Maryland}, size = {178 pages}, timestamp = {2008-06-19T17:45:58.000+0200}, title = {Issues in Scaling Genetic Programming: Breeding Strategies, Tree Generation, and Code Bloat}, url = {http://www.cs.gmu.edu/~sean/papers/thesis2p.ps.gz}, year = 2000 }