Establishing Mechanisms for Self-Adaptation in Genetic Programming

It has long been a desire of computer scientists to develop a computer system that is able to learn and improve without being explicitly programmed to do so. The idea of software that is able to analyse, update and alter itself has been discussed. The thesis is structured as follows: Firstly, we refine and improve the Tartarus problem, proposing it as a benchmark problem for use in GP. Secondly, we establish a mechanism for incorporating self-adaptation into a GP system in order to increase the performance of candidate solutions. Finally, we explore the impact of a fitness bias, inspired by the Dunning-Kruger effect, on the robustness of a GP system. The on-the-fly adaptation of parameter values at runtime can lead to improvements in performance.Self-adaptation aims at biasing the distribution of individuals in a population towards more appropriate and effective areas of the search space. Therefore, we propose, outline and evaluate a novel self-adaptive mechanism favouring a continuous opportunity for modifications to be made during an execution, as-and-when they are deemed to be appropriate. This creates a more flexible parameter modification approach, leading to an increase in solution performance: leading to an approximate 15% and a 10% increase for the Tartarus and Santa-Fe problems respectively. Robustness is often referred to as a characteristic of a candidate solution whose performance is not diminished despite perturbations in environmental parameters or constraints. A solution that does not lose utility or performance quality under these changes is said to be robust. The Dunning-Kruger Effect (DK) is a form of cognitive bias observed in populations, first described by psychologists Dunning and Kruger in 1999: individuals with a low level of ability mistakenly over-estimate their performance and conversely, individuals with a high level of ability will often under-estimate their performance. We propose that the introduction of a DK style bias into the fitness distribution of the population will enable a system to maintain a higher level of population diversity over time