Investigations in multi-objective evolutionary algorithms for design optimization workflows

In the past decade, complex engineering systems have seen increased development towards computer simulations and visualization. This trend is driven by the growing capabilities and availability of information technology resources. As these systems developed, it is now possible to solve difficult problems that could not be solved in the past.

When approaching problems using virtual engineering design, the optimization stage is important step as most of the times the modelling stage still relies on costly black-box simulations. Real-world applications often involve multiple objectives that must be optimized simultaneously, and these objectives are frequently in conflict with one another. It is important to identify which algorithm performs best for a given set of problem features, while considering the constraint of a limited evaluation budget. Another important question is which optimization software is better to be used by practitioners when attempting to solve these type of optimization problems.

In order to address these questions, it is first essential to discuss the algorithms available for solving multi-objective optimization problems. A new optimization software named Liger is used in this work, this software shows promising capabilities and can be used easily by nonexperts in optimization to obtain a good approximation set of optimal design on the true Pareto front. This work focusses on the surrogate based optimization algorithms in Liger and a benchmarking procedure for testing these algorithms. To simulate real world problem features a set of test problems based on Walking Fish Group (WFG) framework are used. The user can adjust the uncertainty in either the radial or perpendicular direction, or simultaneously in both directions. This thesis also demonstrates how an indicator-based multi-objective optimization algorithm from the literature can be easily implemented within the Tigon library. The results obtained using Liger are validated against the ones published in the literature.

The benchmarking framework established in this research was designed to compare stochastic surrogate-based optimization algorithms — specifically, the sParEGO algorithm — with simpler variants of the ParEGO algorithm for robust multi-objective optimization problems. Although sParEGO was specifically designed as a hybrid algorithm to handle uncertainties and provide optimal robust solutions, it will be shown that in some cases, depending on the features of the stochastic problem, the Monte Carlo (MC) version of ParEGO can achieve faster convergence toward the Pareto front and deliver competitive performance in terms of robust solutions for the decision maker.