A Generalized Surrogate-Assisted Evolutionary Algorithm for Expensive Multi-Objective Optimization

A number of real-world problems involve optimization of multiple conflicting criteria assessed using expensive evaluations. In practice, such expensive multi/many-objective optimization problems (EMOPs) need to be solved using small number of design evaluations. Surrogate-assisted evolutionary algorithms (SAEAs) are commonly used to solve EMOPs and broadly follow either a generational or a steady-state form, both having their characteristic advantages. The generational forms involve evaluation of multiple solutions in each generation, and are more suitable for scenarios where the parallelization of the evaluations is possible. The steady-state forms evaluate one solution at a time, therefore incorporating the additional information more frequently, but are more suited to non-parallelizable scenarios. Although it is possible to run the generational algorithms in a steady-state manner through certain parameter settings, we show in this study that such settings have a negative impact on their search performance. This leads to an inference that existing methods are not able to seamlessly switch between generational and steady-state forms based on the application at hand. In this study, we address this gap by extending a recently proposed steady-state framework (SASSEA) to a generalized form (GSAEA) which can be run in both steady-state or generational form by simply prescribing the number of true evaluations per generation. Numerical experiments are conducted using low evaluation budgets on a range of MOPs with up to 7 objectives and diverse types of Pareto-optimal fronts. The results indicate that the proposed framework shows competitive or better performance relative to the state-of-the-art methods, both in generational and steady-state setting.