NSGA-III - Results
A Free and Open Source Java Framework for Multiobjective Optimization
Brought to you by:
dhadka
Menu
▾
▴
View and moderate all "feature-requests Discussion" comments posted by this user
Mark all as spam, and block user from posting to "Feature Requests"
is it working??
View and moderate all "feature-requests Discussion" comments posted by this user
Mark all as spam, and block user from posting to "Feature Requests"
Ran fine on 6-objective DTLZ2. Are you experiencing an error when trying more than 4 objectives?
View and moderate all "feature-requests Discussion" comments posted by this user
Mark all as spam, and block user from posting to "Feature Requests"
Without erros, but the values (results) are so "lousy". Do you have some example?
I saw here: https://www.researchgate.net/post/Is_there_a_fully_functional_NSGA-III_implementation#view=55d459c05dbbbddaea8b45b1
The people are talking about the NSGA-III and they said that there is no some implementation (NSGA-III) working fully :/
View and moderate all "feature-requests Discussion" comments posted by this user
Mark all as spam, and block user from posting to "Feature Requests"
The implementation within the MOEA Framework is based on the reading of the NSGA-III paper and comparing against other implementations (such as the C++ version). I'm not exactly sure why they are saying on the researchgate forum that the implementation is not working. Initial tests showed it producing similar figures to those shown in the paper (due to stochastic effects, it can never be exact). No one has yet pointed out a flaw in the implementation.
Of course, as they rightly point out, there is no official publicly-available implementation from Kalyanmoy Deb. As such, we can't be 100% sure any implementation is exactly as Deb intends, but it matches the description provided in the published paper.
You can use the Diagnostic Tool via the "Demo Application" download to test NSGA-III against other algorithms on many, many problems.
Could you also clarify what you mean by "lousy"? Are the results worse than what you expect? Are they worse compared to another algorithm? I'd be happy to investigate further if you have an example where it performs worse than what one expects.
View and moderate all "feature-requests Discussion" comments posted by this user
Mark all as spam, and block user from posting to "Feature Requests"
Since this is a potentially important algorithm, I'm going to add some better unit testing in the next release (essentially replicating what the nsga3cpp author did).
One thing to be cautious of when comparing results from the MOEA Framework to Deb's results is Deb redefines the inverted generational distance (IGD) metric so that the reference set only contains the Pareto optimal points nearest to their reference points. I will also provide a means to replicate their IGD calculations in the next release.