[joss-reviews#2974] Software paper about framework HOT 12 CLOSED

Thanks for the advice, the paper is updated on the PR #76

from framework.

The paper was already updated with @sjvrijn (but not the pdf I think) :

The first version of MOEA/D and its most famous variants [@moead_de; @moead_dra] are implemented in recent multi-objective optimization software such as pymoo [@pymoo], pygmo [@pygmo] and jMetal [@jMetal]. These implementations offer many state-of-the-art algorithms, visualization tools or parallelization abstraction, but they do not enable detailed testing and analysis of the various algorithm's components' behavior.

from framework.

These implementations offer many state-of-the-art algorithms, visualization tools or parallelization abstraction, but they do not enable detailed testing and analysis of the various algorithm's components' behavior.

Since this is the comparative advantage of MOEA/D I think that you should show it in the paper.

from framework.

I suggest adding a sentence:

These implementations offer many state-of-the-art algorithms, visualization tools or parallelization abstraction, but they do not enable detailed testing and analysis of the various algorithm's components' behavior.
In these libraries, the algorithms are seen as components that do not have sub-components, which makes their parameterization and analysis more difficult, we cannot easily modify their behavior.

from framework.

In these libraries, the algorithms are seen as components that do not have sub-components, which makes their parameterization and analysis more difficult, we cannot easily modify their behavior.

I think that this is probably related to #69. Apart from the new text, can you please provide a concrete example of an idea that cannot be implemented with a commonly-used package but can easily be implemented with MOEA/D?

from framework.

With pygmo and Jmetal for example, users are very free to implement new algorithms because they are not restricted by subcomponents, but this also makes it more difficult to implement new algorithms based on MOEA/D. With these frameworks, it is necessary to reimplement a complete algorithm (with code duplication) to implement variants of MOEA/D whereas with our framework, it is only sufficient to implement the component or the function that we want to modify or study. Each of the examples proposed in the documentation cannot therefore be implemented more easily than with our framework.

from framework.

With pygmo and Jmetal for example, users are very free to implement new algorithms because they are not restricted by subcomponents, but this also makes it more difficult to implement new algorithms based on MOEA/D. With these frameworks, it is necessary to reimplement a complete algorithm (with code duplication) to implement variants of MOEA/D whereas with our framework, it is only sufficient to implement the component or the function that we want to modify or study. Each of the examples proposed in the documentation cannot therefore be implemented more easily than with our framework.

Here you are trying to explain the modularity that the package offers. I understand.

Although, in the paper pymoo, pygmo and jMetal are mentioned followed by the sentence "...but they do not enable to test and analyze in detail the behavior of components implemented in each algorithm"

Either you have to show somehow the aforementioned phrase or you should rephrase.

from framework.

The paper is updated in #76

These implementations offer many state-of-the-art algorithms, visualization tools or parallelization abstraction, but they are not modular enough to test easily all MOEA/D components.

from framework.

but they are not modular enough to test easily all MOEA/D components.

Should we show somehow this claim?

from framework.

I don't know, I don't see how to demonstrate otherwise that if the other libraries are not modular then we can't test specific components easily without completely reimplementing the algorithm ?

from framework.

I don't know, I don't see how to demonstrate otherwise that if the other libraries are not modular then we can't test specific components easily without completely reimplementing the algorithm ?

Could you make a comparison table of moead-framwork with SOTA libraries? Like this you can show a direct comparison without implementing a complete example.

from framework.

Great

from framework.