Created by W.Langdon from gp-bibliography.bib Revision:1.8098
The investigations in this thesis are motivated by this requirement for generalization, and need for dynamic optimisation techniques.
We obtain a number of significant results which contribute to the body of knowledge concerning generalization and dynamic optimization using genetic programming, as summarised below.
We first illustrate for a small number of symbolic regression target functions, that overfitting is an issue even in the absence of noise in the training data, when GP is expected to extrapolate. We implement sophisticated early stopping criteria and our results show that these criteria can identify the optimal point at which to stop training.
We investigate the implementation of a landscape-based dynamic benchmark for GP, but discover that it is not possible to both consistently set fitness values for trees which are identical upon evaluation, while simultaneously tuning the shape of our fitness landscape. We investigate the establishment of a symbolic regression-based dynamic benchmark for GP: using continuous populations, our experimental results show that it is possible to create a benchmark by varying the semantic distance between successive target functions in order to tune the magnitude of change.
Finally, we look at two real-world economic datasets of interest, and produce a GP system which statistically significantly out performs a naive benchmark. Our results show that our system performance is not particularly sensitive to the length of historical training data provided. Over multiple time windows, when we compare using continuous populations against using re-initialized populations, our experimental results show that both approaches perform approximately equivalently on these dynamic problems.",
Supervisors: Michael O'Neill and Anthony Brabazon",
Genetic Programming entries for Cliodhna Tuite