Is our archiving reliable? Multiobjective archiving methods on “simple” artificial input sequences

Research output: Contribution to journalArticlepeer-review

56 Downloads (Pure)


In evolutionary multiobjective optimisation (EMO), archiving is a common component that maintains an (external or internal) set during the search process, typically with a fixed size, in order to provide a good representation of high-quality solutions produced. Such an archive set can be used solely to store the final results shown to the decision maker, but in many cases may participate in the process of producing solutions (e.g., as a solution pool where the parental solutions are selected). Over the last three decades, archiving stands as an important issue in EMO, leading to the emergence of various methods such as those based on Pareto, indicator, or decomposition criteria. Such methods have demonstrated their effectiveness in literature and have been believed to be good options to many problems, particularly those having a regular Pareto front shape, e.g., a simplex shape.

In this article, we challenge this belief. We do this through artificially constructing several sequences with extremely simple shapes, i.e., 1D/2D simplex Pareto front. We show the struggle of predominantly used archiving methods which have been deemed to well handle such shapes. This reveals that the order of solutions entering the archive matters, and that current EMO algorithms may not be fully capable of maintaining a representative population on problems with linear Pareto fronts even in the case that all of their optimal solutions can be found.
Original languageEnglish
Article number9
Pages (from-to)1-19
Number of pages19
JournalACM Transactions on Evolutionary Learning and Optimization
Issue number3
Early online date18 Aug 2021
Publication statusPublished - Sept 2021


  • Multiobjective optimisation
  • archiving
  • elitism
  • population maintenance
  • deterioration


Dive into the research topics of 'Is our archiving reliable? Multiobjective archiving methods on “simple” artificial input sequences'. Together they form a unique fingerprint.

Cite this