Enhanced Evolutionary Algorithms for Solving Computationally-Intensive Multi-Objective Optimization Problems

Over the past decade, specialized evolutionary algorithms have emerged as one of the best methods for solving multi-objective optimization problems (MOOPs). One of the characteristics of evolutionary algorithms is that a large number of evaluations of the objective function need to be performed during their execution. This inherent requirement is extremely problematic in the case of industrial optimization contexts where evaluating objective performance and constraint satisfaction is very computationally intensive. We are particularly concerned with the need to improve optimization performance on computationally-intensive multi-objective optimization problems (CIMOOPs) from the field of electrical drive design. Such design problems are fairly complicated as they arise from the need to simultaneously increase the efficiency, reduce the costs and improve the fault tolerance and operating characteristics of new electrical machines. Furthermore, (several) computationally-intensive finite element (FE) simulations are usually required in order to evaluate the performance of a single design. In this thesis we present the results of research that was aimed at improving the performance of multi-objective evolutionary algorithms (MOEAs) when applied on CIMOOPs. Initial focus falls on designing and applying on-the-fly surrogate modeling in order to reduce the dependency of the MOEAs on FE simulations. By surrogate modeling we understand the creation of fast-to-evaluate linear and non-linear regression models that can accurately approximate FE results. Next, we investigate the best way of distributing MOEA computations over a high-throughput computing environment, when considering a master-slave architecture. Finally, we propose two new MOEAs (based on coevolution) that are both highly competitive when compared to state-of-the-art approaches and quite robust with regard to their own parameterization settings.