A Restart CMA Evolution Strategy With Increasing Population Size
Anne Auger and Nikolaus Hansen
In this paper we introduce a restart-CMA-evolution strategy, where
the population size is increased for each restart (IPOP). By
increasing the population size the search characteristic becomes
more global after each restart. The IPOP-CMA-ES is evaluated on
the test suit of 25 functions designed for the special session on
real-parameter optimization of CEC 2005. Its performance is
compared to a local restart strategy with constant small
population size. On unimodal functions the performance is similar.
On multi-modal functions the local restart strategy significantly
outperforms IPOP in 4 test cases whereas IPOP performs
significantly better in 29 out of 60 tested cases.
ERRATA:
End of Section 2, stopping criterion "noeffectcoord":
"Stop if adding 0.2-standard deviation in each coordinate does
change ..."
should be
"Stop if adding 0.2-standard deviation in any coordinate does
not change ..."
The following has been pointed out by Xiao-Min Hu from the Sun
Yat-sen University:
1) On function F4 the initial search point was chosen by
mistake as Xup + (Xup-Xlow)*rand instead of Xlow +
(Xup-Xlow)*rand The corrected initialization will
presumably give better result.
2) On some function the boundary setting was omitted, but
this has most probably no impact on the results.