Created by W.Langdon from gp-bibliography.bib Revision:1.6717
The primary objective of this thesis is to propose a novel paradigm of EDA named Probabilistic Model Building Genetic Network Programming (PMBGNP), where the directed graph structure of a novel graph-based EA called Genetic Network Programming (GNP) is used to represent its individuals. Different from most of the current EDAs proposed in string structure based Genetic Algorithm (GA) and tree structure based Genetic Programming (GP), the distinguished graph structure allows PMBGNP to ensure higher expression ability. As a result, a sort of problems can be explored and solved efficiently and effectively comparing with the conventional research in EDA literature.
To achieve this objective, contributions of this thesis are presented on the following two aspects: algorithm part and application part.
From the perspective of algorithm part, first, the thesis proposes the high-level PMBGNP to use Maximum Likelihood Estimation (MLE) to model the probability distribution of the promising individuals. PMBGNP is empirically studied to show the capability of speeding up the evolution efficiency by the estimation of probability distribution. Second, the thesis addresses the issue of population diversity loss by theoretical comparison with classical EDAs, and proposes a hybrid algorithm to maintain the population diversity of PMBGNP. Third, the integration of PMBGNP and Reinforcement Learning (RL) is studied. Inspired by behaviourist psychology, RL concerns with reinforcing the growth of the individuals by learning their experiences. The learning knowledge formulated by Q values can be approximated and incorporated into the probabilistic modelling of PMBGNP to improve the performance by constructing a more accurate model. Finally, PMBGNP is extended from discrete optimisation problems to continuous optimization problems.
From the viewpoint of application part, most of the current studies in EDA are carried out in the benchmark problems of GA and GP, such as function optimisation and symbolic regression. Therefore, to accomplish one of the essential challenges of EDA for novel applications, the thesis applies PMBGNP to two novel applications of EDA, including data mining and the problems of controlling the agents' behaviour. By comparing with the other state-of-the-art algorithms, PMBGNP is testified to be capable of achieving better performances.",
Genetic Programming entries for Xianneng Li