search

numbbo / coco

Numerical Black-Box Optimization Benchmarking Framework
https://numbbo.github.io/coco
Other
261 stars 87 forks source link

Choice of the region of interest in the (mixed-integer) multiobjective case #1793

Closed brockho closed 11 months ago

brockho commented 6 years ago

The next step towards answering this question is to see whether the solutions in our Pareto set approximations, we know are lying outside of [-5,5]^n, are actually non-dominated (are they close to the Pareto set?).

We can do this by optimizing (e.g. with CMA-ES) the COCO hypervolume with the Pareto set approximation's solution as reference point.

ttusar commented 5 years ago

As per today's discussion, we should set the ROI of the continuous variables for the mixed-integer bi-objective suite to [-5, 5]^n.

nikohansen commented 5 years ago

Isn't that a little bit too optimistic that we can go from [-100,100] right back to [-5,5]?

ttusar commented 5 years ago

I believe the reasoning was that since the smallest problem dimension is 5, this should be reasonably safe to do...

nikohansen commented 5 years ago

Yes, I also understand that with integers there is less of a chance to get good solutions outside the bounds, because integers themselves cannot take values outside the bounds at all. Yet, I don't see the reasoning how/why [-5,5] is now "safe" in >=5-D knowing that before it was set to [-100,100] to make it safe in 2-D.

EDIT: the reasoning is that in higher dimensions (>=3?) we never observed Pareto optimal points outside of [-5, 5].

nikohansen commented 11 months ago

Currently, the lower_bounds and upper_bounds property of a cocoex.interface.Problem is set to -100 and 100 for continuous variables.

import cocoex

suite = cocoex.Suite('bbob-biobj-mixint', '', '')
p = suite[1]
p.lower_bounds, p.upper_bounds
(array([   0.,    0.,    0.,    0., -100.]),
 array([  1.,   3.,   7.,  15., 100.]))
brockho commented 11 months ago

Looks like, we can close the issue then.