search

jMetal / jMetalPy

A framework for single/multi-objective optimization with metaheuristics
https://jmetal.github.io/jMetalPy/index.html
MIT License
498 stars 150 forks source link

Why execution time increases? #93

Closed Antonio-Nappi closed 3 years ago

Antonio-Nappi commented 3 years ago

Hi everybody, I'm working on OMOPSO algorithm and something strange happens: at the initial time there are a lot of it/s but when the execution time increases there is a decreases of the it/s. I'm working on a very power cpu with 32 core and I've set the multiprocessevaluator but it takes too long to finish. Is it the right behaviour? How cold I increase the execution speed?

ajnebro commented 3 years ago

Which problem are you solving and how have you configured OMOPSO?.

Just have into account that OMOPSO uses an external archive to store the non-dominated solutions found during the search. When that archive becomes full, a procedure to remove the worst solution (in the case of OMOPSO, an EpsilonDominanceComparator is used) is applied. If the archive size is large, this can take time; furthermore, this is a sequential step, so adding more cores will not help.

Antonio-Nappi commented 3 years ago

I'm trying to estimate parameters for an epidemiologica compartment model. I've used the configuration present in the documentation, just changing swarm size( in order to keep it between 5 times and 10 times the dimensionality of my problem) and changing the number of iteration. The epsilon value suggested is epsilon=0.0075, how can I change this value in order to keep the archive small?

ajnebro commented 3 years ago

Have you checked the size of the archive using that epsilon value?

Antonio-Nappi commented 3 years ago

algorithm = OMOPSO( problem=problem, swarm_size=swarm_size, epsilon=0.0075, uniform_mutation=UniformMutation(probability=mutation_probability, perturbation=0.5), non_uniform_mutation=NonUniformMutation(mutation_probability, perturbation=0.5, max_iterations=int(max_evaluations / swarm_size)), leaders=CrowdingDistanceArchive(100), termination_criterion=StoppingByEvaluations(max=max_evaluations) ) This is the code that I run, how can I check the size of the archive?

ajnebro commented 3 years ago

The size of the leaders archive is 100, so it seems ok. Have you tried SMPSO instead of OMOPSO?

Antonio-Nappi commented 3 years ago

no, which are the main differences?

ajnebro commented 3 years ago

SMPSO is based on OMOPSO, but it does not use epsilon dominance (https://ieeexplore.ieee.org/document/4938830). It can produce more accurate fronts than OMOPSO when dealing with multi-frontal problems (as, for example, ZDT4).

Antonio-Nappi commented 3 years ago

Ok, thanks. I'll try it.