Enhancement of Model Generalisation in Multiobjective Genetic Programming

Multiobjective genetic programming (MOGP) is a powerful evolutionary algorithm that requires no human pre-fixed model sets to handle regression and classification problems in the machine learning area. We aim to improve the model generalisation of MOGP in both regression and classification tasks. The work in this thesis has three main contributions.
First, we propose replacing the division operator used in genetic programming with an analytic quotient (AQ) operator in regression to systematically achieve lower mean squared error due principally to removing the discontinuities or singularities caused by conventional protected or unprotected division. Further, this AQ operator is differentiable.
Second, we propose using Tikhonov regularisation, in conjunction with node count (using an extension of Pareto comparison from vectors to tuples) as a general complexity measure in MOGP. We demonstrate that employing this general complexity yields mean squared test error measures over a range of regression problems which are typically superior to those from conventional node count. We further analysed the reason why our new method outperforms the conventional complexity measure and conclude that it forms a decision mechanism which balances both syntactic and semantic information.
Third, we propose using a loss measure complementary to Vapnik's statistical learning theory, which can effectively stabilise classifiers trained by MOGP. We demonstrate that this loss measure has a number of attractive properties and has a better correlation with generalisation error compared to 0/1 loss, so that better generalisation performance is achievable.