An integrated immune-inspired framework for many-objective optimization

Abstract

Optimization problems are common in the real world. Many problems arising from engineering, science or business require the simultaneous considerations of several objectives. These types of problems are termed as multi-objective optimization problems. Pareto-based Multi-Objective Evolutionary Algorithms perform well in bi-objective are tri-objective problems. However, these algorithms demonstrate difficulties in many-objective problems with four or more objectives. The failure in accurately rank the current population given rise to the poor performance. Many-objective problems occur frequently in engineering and industrial applications. It is desirable to develop effective and efficient approaches for handling many-objective problems. This research therefore aims to provide an integrated approach for a posteriori decision-making for many-objective optimization.

Immune system is a defense system against external invasion. From an engineering point of view, the collective characteristics of self-organization, self-learning and self-adaptability have offered real potential to streamline the evolution dynamics in optimization. Artificial immune systems (AIS), the engineering abstraction of the human immune system, have been applied with promising results in single-objective and multi-objective problems; however the application to many-objective domain is rather limited. Hence the dual Region-defined Immune-inspired evolution strategy for Many-objective Optimization Algorithm (RIMA) is developed in this study for many-objective optimization which is developed based on the key mechanisms of AIS. Novel features including the region-based selection facilitated by the region-dominance evaluation, the territory-centered evolution facilitated by the region decomposition and the immune network theory inspired regulation mechanism are developed to regenerate the selection pressure, to manage the evolution dynamics, to enhance the exploitation and exploration process and to minimize the overheads in the evolution process. The proposed RIMA is benchmarked with seven selected optimizers in the experimental studies. The results reveal that RIMA is very competitive when compared with the state-of-art algorithms.

Besides the development in the algorithm domain, the possible reduction in the number of objectives is also under investigation. Region-dependent Objective Clustering Scheme (ROC) is proposed to group the harmonious objectives into clusters. ROC utilizes the region-based concept to employ the region-based decision in determining whether two objectives are in conflict or in harmony. As for any objective reduction or cluster schemes, it is difficult to control the number of objectives after the reduction process in which the number of objectives may still be uncontrollable by common MOEAs. To tackle this problem, the ROC scheme is incorporated with the partial dominance scheme (ROCPD). The problem is first being reduced by the ROC scheme. The partial dominance scheme then converts the objective vector into a series of objective subsets in which each subset contain only a manageable number of objectives. Very promising result is reported from the experimental study. ROC scheme is able to identify the redundant objectives in most cases in the simulation study.