Created by W.Langdon from gp-bibliography.bib Revision:1.7175
In common with most Machine Learning approaches, GP has been used to solve many trivial problems. When applied to larger and more complex problems, however, several difficulties become apparent. When focusing on the basic features of GP, this thesis highlights the immense size of the GP search space, and describes an approach to measure this space. A stupendously flexible but frustratingly useless representation, Anarchically Automatically Defined Functions, is described. Some difficulties associated with the normal use of the GP operator Crossover (perhaps the most common method of combining GP trees to produce new trees) are demonstrated in the simple MAX problem. Crossover can lead to irreversible sub-optimal GP performance when used in combination with a restriction on tree size. There is a brief study of tournament selection which is a common method of selecting fit individuals from a GP population to act as parents in the construction of the next generation.
The main contributions of this thesis however are two approaches for avoiding the fitness evaluation bottleneck resulting from the use of SL in GP. To establish the capability of a GP individual using SL, it must be tested or evaluated against each example in the set of training examples. Given that there can be a large set of training examples, a large population of individuals, and a large number of generations, before good solutions emerge, a very large number of evaluations must be carried out, often many tens of millions. This is by far the most time-consuming stage of the GP algorithm. Limited Error Fitness (LEF) and Dynamic Subset Selection (DSS) both reduce the number of evaluations needed by GP to successfully produce good solutions, adaptively using the capabilities of the current generation of individuals to guide the evaluation of the next generation. LEF curtails the fitness evaluation of an individual after it exceeds an error limit, whereas DSS picks out a subset of examples from the training set for each generation.
Whilst LEF allows GP to solve the comparatively small but difficult Boolean Even N parity problem for large N without the use of a more powerful representation such as Automatically Defined Functions, DSS in particular has been successful in improving the performance of GP across two large classification problems, allowing the use of smaller population sizes, many fewer and faster evaluations, and has more reliably produced as good or better solutions than GP on its own.
The thesis ends with an assertion that smaller populations evolving over many generations can perform more consistently and produce better results than the `established' approach of using large populations over few generations.
Genetic Programming entries for Chris Gathercole