New publication available

 

Martin Schlueter, Masaharu Munetomo:

MIDACO parallelization scalability on 200 MINLP benchmarks  [ PDF ]

Journal of Artificial Intelligence and Soft Computing (JAISCR), De Gruyter, accepted (Nov 2016)

 


This article investigates the scalability behavior of MIDACO in regard to parallel executed problem function calls. On a set of 200 MINLP benchmarks the algorithmic impact (measured in performed blocks of evaluation) of a varying parallelization factor P from 1 to 300 is investigated. It is demonstrated  that for a maximal parallelization factor of P=300 the MIDACO algorithm can efficiently reduce the number of problem function calls by a factor of over 150 times (see Section 3, Figure 3):

This study is relevant for CPU-time expensive optimization problems, where a single function evaluation of the objectives and/or constraints might take significant time. By performing several such function evaluation in parallel, the total time of the optimization process can be drastically reduced. According to above results, applying a parallelization factor of P=300 can make the required overall optimization time over 150 times faster.