Advanced p-Metric Based Many-Objective Evolutionary Algorithm

Abstract

Evolutionary many objective based optimization has been gaining a lot of attention from the evolutionary computation researchers and computational intelligence community. Many of the state-of-the-art multi-objective and many-objective optimization problems (MOPs, MaOPs) are inefficient in maintaining the convergence and diversity performances as the number of objectives increases in the modern-day real-world applications. This phenomenon is obvious indeed as Pareto-dominance based EAs employ non-dominated sorting which fails considerably in providing enough convergent pressure towards the Pareto front (PF). Researchers invested much more time and effort in addressing this issue by improving the scalability in MaOPs and they have come up with non-Pareto-dominance-based EAs such as decomposition-based, indicator-based and reference-based approaches. In addition to that, the algorithm has to account for the additional computational budget. This thesis proposes an advanced polar-metric (p-metric) based Many-objective EA (in short APMOEA) for tackling both MOPs and MaOPs. p-metric, a recently proposed performance based visualization metric, employs an array of uniformly, distributed direction vectors. In APMOEA, a two-phase selection scheme is employed which combines both non-dominated sorting and p-metric. Moreover, this thesis also proposes a modified P-metric methodology in order to adjust the direction vectors dynamically. In the experiments, we compare APMOEA with four state-of-the-art Many-objective EAs under, three performance indicators. According to the empirical results, APMOEA shows much improved performances on most of the test problems, involving both MOPs and MaOPs.