Online Knowledge-based Evolutionary Multi-objective Optimisation

Abstract

Knowledge-based optimization is a recent direction in evolutionary optimization research which aims at understanding the optimization process, discovering relationships between decision variables and performance parameters, and using discovered knowledge to improve the optimization process, using machine learning techniques. This thesis makes two major contributions in the existing body of knowledge in the area of evolutionary multi-objective optimization. First, in addition to the well-researched objective space, it highlights the need for focusing on decision space performance analysis for benchmarking multi-objective evolutionary algorithms in general, and more specifically the knowledge-based class of these algorithms. In this respect, the thesis proposes a new method to generate multi-objective optimization test problems with clustered Pareto sets in hyper-rectangular defined areas of the decision space, which mimics knowledge representation in propositional logic. Further, a new metric is introduced for performance measurement in terms of their coverage of the optimal decision sub-space. The proposed test problems and metrics are used to benchmark multi-objective evolutionary algorithms in both objective and decision spaces. Second, this thesis introduces a novel evolutionary optimization framework that incorporates a knowledge-based representation to search for Pareto optimal patterns in decision space replacing the conventional point-based representation. Compared to the extant approaches, which process the post-optimization Pareto sets for knowledge discovery using statistical or machine learning methods, the framework facilitates online discovery of knowledge during the optimization process in the form of interpretable rules. The core contributing idea is that the multi-objective evolutionary process is applied on a population of bounding hypervolumes, or rules, instead of evolving individual point-based solutions. The framework is generic in the sense that existing algorithms can be adapted to evaluate the quality of rules based on sampled solutions from the bounded space. Two algorithmic instantiations of the framework are presented in this thesis for both the multi and many objective optimizations respectively. The results and analysis of the experimentation with standard and proposed test benchmarks demonstrate the capabilities of the proposed optimization algorithm in comparison to the state-of-the-art multi-objective evolutionary algorithms.