Bounding the population size of IPOP-CMA-ES on the noiseless BBOB testbed
Tianjun Liao, Thomas Stützle · 2013
A variant of CMA-ES that uses occasional restarts coupled with an increasing population size, which is called IPOP-CMA-ES, has shown to be a top performing algorithm on the BBOB benchmark set. In this paper, we test a mechanism that bounds the maximum size that the population may reach in IPOP-CMA-ES, and we experimentally explore the impact of a maximum population size on the BBOB benchmark set. In the proposed bounding mechanism, we use a maximum population size of 10 × D2 where D is problem dimension. Once the maximum population size is reached or surpassed, the population size is reset to its default starting value λ, which is defined by the λ = 4 + 3 ln(D). Our experimental results show that our scheme for the population size update can lead to improved performances on separable and weakly structured multi-modal functions.