The Pareto Archived Evolution Strategy (PAES)
Introduction
The Pareto Archived Evolution Strategy (PAES) is a multiobjective optimizer which uses a simple (1+1)
local search evolution strategy. Nonetheless, it is capable of finding diverse solutions in the
Pareto optimal set because it maintains an archive of nondominated solutions which it exploits to
estimate accurately the quality of new candidate solutions.
Three versions, (1+1), (1+lambda) and (mu+lambda)-PAES have been developed. Each of these versions has been
tested against two well known multiobjective evolutionary algorithms - the Niched Pareto Genetic
Algorithm (NPGA) and a nondominated sorting GA (NSGA). Tests were carried out using five test functions
(f2-f6) and results have been processed using statistical techniques introduced by Fonseca and
Fleming. C code for each of the test functions and the statistical techniques are available below.
Please drop me an email if you use any of these resources.
PAES is described in detail in two of the publications listed below. The submission to Evolutionary Computation is the
largest paper and includes use of Fonseca and Fleming's statistical techniques to measure the performance of
PAES with respect to five test functions.
In August 1999, PAES was tested on six different test functions. These functions, designed by Kalyanmoy Deb, each to
capture and isolate a specific problem feature known to present difficulties to multiobjective optimizers, were first
used in a paper by Zitzler, Deb and Thiele (Technical Report 70, TIK)
in which eight multiobjective evolutionary algorithms were compared. We
use the raw results from the study by Zitzler et al. to compare the performance of PAES with the eight algorithms tested
in that study. The results showed that PAES exhibits at least comparable performance to the best algorithm from the Zitzler et al. study,
the Strength Pareto Evolutionary Algorithm (SPEA), on four
of the six Deb test functions. The raw results from the Zitzler et al. study are available here. The third paper below describes this work in
more detail.
For the first time, we are making a skeleton version of our PAES code available for download here. See the resources section below. If you do download this code, please drop me an e-mail to let me know.
|