On the Evolutionary Optimisation of Many Objectives


Evolutionary algorithms based on the concept of Pareto dominance are popular and sucessful techniques for multi-objective optimisation. Such algorithms are required to obtain a good approximation (in terms of proximity and diversity) to trade-off surfaces of interest to the decision-maker. To achieve good performance, the design of an evolutionary optimiser should be component-based and application-focused. Multi-objective algorithms tend to be analysed as composites using empirical approaches that lack statistical rigour. This severely limits the availability of process knowledge, itself essential for component-level design. The thesis addresses this issues with a new, structured, framework for algorithm development and assessment. A rigurous, nonparametric statistical methodology is applied to appropriate performance indicators for proximity and spread. Applications of the framework can be found throughout the thesis. Evolutionary multi-objective theory mainly considers bi-objective problems. However, multi-objective applications often consider many more objectives. Thus, more theoretical work into the optimisation of many objectives is required. A platform for such research is established via consideration of the relationships that may exist between pairs of objectives. Three relationships---conflict, harmony, and independence--- are identified, and issues and current research are discussed for each relationship. The effect of many conflicting objectives on a class of popular evolutionary multi-objective optimisers is considered in a detailed, exploratory investigation. The study reveals that conclusions drawn from bi-objective analysis cannot be generalised to higher numbers of conflicting objectives. The probable mechanisms that underpin the observed variation in behaviour are also identified and discussed. The thesis also demonstrates, using the rigorous empirical framework, that if independence exists in a multi-objective problem, then identification and exploitation of this relationship can produce improved optimiser results. An innovative method for identifying suitable objective-space decomposition on-line is subsequently developed, based on concepts from parallel evolutionary topologies and nonparametric statistics. Excellent results are obtained in a proof-of-principle assessment.