Quantifying the effects of objective space dimension in evolutionary multiobjective optimization

  • Authors:

  • Joshua Knowles;David Corne

  • Affiliations:

  • School of Computer Science, University of Manchester, Manchester, UK;School of Mathematical and Computer Sciences, Heriot-Watt University, UK

  • Venue:
  • EMO'07 Proceedings of the 4th international conference on Evolutionary multi-criterion optimization
  • Year:
  • 2007

Quantified Score

Hi-index 0.00

Visualization

Abstract

The scalability of EMO algorithms is an issue of significant concern for both algorithm developers and users. A key aspect of the issue is scalability to objective space dimension, other things being equal. Here, we make some observations about the efficiency of search in discrete spaces as a function of the number of objectives, considering both uncorrelated and correlated objective values. Efficiency is expressed in terms of a cardinality-based (scaling-independent) performance indicator. Considering random sampling of the search space, we measure, empirically, the fraction of the true PF covered after p iterations, as the number of objectives grows, and for different correlations. A general analytical expression for the expected performance of random search is derived, and is shown to agree with the empirical results. We postulate that for even moderately large numbers of objectives, random search will be competitive with an EMO algorithm and show that this is the case empirically: on a function where each objective is relatively easy for an EA to optimize (an NK-landscape with K=2), random search compares favourably to a well-known EMO algorithm when objective space dimension is ten, for a range of inter-objective correlation values. The analytical methods presented here may be useful for benchmarking of other EMO algorithms.