When evaluating the performance of a stochastic optimizer it is sometimes desirable to express performance in terms of the quality attained in a certain fraction of sample runs. For example, the sample median quality is the best estimator of what one would expect to achieve in 50% of runs, and similarly for other quantiles. In multiobjective optimization, the notion still applies but the outcome of a run is measured not as a scalar (i.e. the cost of the best solution), but as an attainment surface in k-dimensional space (where k is the number of objectives). In this paper we report an algorithm that can be conveniently used to plot summary attainment surfaces in any number of dimensions (though it is particularly suited for three). A summary attainment surface is defined as the union of all tightest goals that have been attained (independently) in precisely s of the runs of a sample of n runs, for any s/spl isin/1..n, and for any k. We also discuss the computational complexity of the algorithm and give some examples of its use. C code for the algorithm is available from the author.
[1]
Jonathan E. Fieldsend,et al.
Dominance measures for multi-objective simulated annealing
,
2004,
Proceedings of the 2004 Congress on Evolutionary Computation (IEEE Cat. No.04TH8753).
[2]
Peter J. Fleming,et al.
On the Performance Assessment and Comparison of Stochastic Multiobjective Optimizers
,
1996,
PPSN.
[3]
Joshua D. Knowles,et al.
ParEGO: a hybrid algorithm with on-line landscape approximation for expensive multiobjective optimization problems
,
2006,
IEEE Transactions on Evolutionary Computation.
[4]
Marco Laumanns,et al.
Performance assessment of multiobjective optimizers: an analysis and review
,
2003,
IEEE Trans. Evol. Comput..