A survey of decomposition approaches in multiobjective evolutionary algorithms

Jia Wang,Yuchao Su,Qiuzhen Lin,Lijia Ma,Dunwei Gong,Jianqiang Li,Zhong Ming

DOI: https://doi.org/10.1016/j.neucom.2020.01.114

IF: 6

2020-09-01

Neurocomputing

Abstract:<p>Since the multiobjective evolutionary algorithm based on decomposition (MOEA/D) was proposed by Zhang and Li in 2007, this interesting framework has attracted a considerable attention from researchers. In MOEA/D, a multiobjective optimization problem is decomposed into a series of aggregated subproblems, which are optimized simultaneously in a collaborative way by using the information from their neighboring subproblems. The decomposition approach has significant impact on MOEA/D as it directs the evolutionary search. Many improved MOEA/D variants proposed various kinds of decomposition approaches and have shown promising performance for different kinds of problems. In this paper, we give a survey of decomposition approaches, which are classified into five categories, i.e., the tradition decomposition, the modified Tchebycheff decomposition, the modified penalty-based boundary intersection decomposition, the constrained decomposition, and other special cases of decomposition. Moreover, discussions are further given in this paper to analyze the performance of different decomposition approaches. One clarifies the difference between Tchebycheff decomposition and Pareto-based domination. The other one compares the performance of various decomposition approaches on different benchmark problems. Experiments results have demonstrated that the Tchebycheff decomposition and its varieties are robust on solving most problems while some specific decomposition approaches are very effective for some problems with special features.</p>

computer science, artificial intelligence

What problem does this paper attempt to address?

This paper aims to solve the problems of classification, analysis and evaluation of decomposition - based methods in multi - objective evolutionary algorithms (MOEAs). Specifically, the paper focuses on how to optimize multi - objective optimization problems (MOPs) through different decomposition methods, especially in the framework of multi - objective evolutionary algorithms, how to achieve effective optimization by decomposing complex problems into a series of simpler sub - problems.

The paper mentions that since Zhang and Li proposed the decomposition - based multi - objective evolutionary algorithm (MOEA / D) in 2007, this framework has attracted wide attention from researchers. In MOEA / D, an MOP is decomposed into a series of aggregated sub - problems, and these sub - problems are simultaneously optimized in a collaborative manner by using the information of their neighboring sub - problems. The decomposition method has a great impact on MOEA / D because it guides the direction of evolutionary search. Many improved MOEA / D variants have proposed various types of decomposition methods and have shown satisfactory performance on different types of problems.

The main contributions of this paper are as follows:
1. Classify a series of decomposition methods into five categories, and visualize and discuss the improvement region (IR) where the solution can be updated using weight vectors and decomposition methods.
2. Explain the differences between Tchebycheff Decomposition and traditional Pareto - based Dominance.
3. Test the performance of several typical decomposition methods on six benchmark problems with different characteristics to better understand these benchmark problems and the tested decomposition methods.

Through these works, the paper not only provides a comprehensive review of existing decomposition methods, but also provides a valuable reference for further research and development.