BBOB data: strange costs

[berndbischl]

• Why do we have 0 costs for some steps?
• Are some of the costs really as big as 3686000 ?

[kerschke]

Yes. Some steps cost nothing because they only use the initial design (and those costs are already listed in the corresponding step).
And yes, some features were veeeery expensive - as I stated within my thesis. Maybe I should give you the thesis back again ;-)

[berndbischl]

Yes. Some steps cost nothing because they only use the initial design
(and those costs are already listed in the corresponding step).

I got confused here. The FEATURES need the initial design step so there are no FEATURES without costs.

And yes, some features were veeeery expensive - as I stated within my thesis.
Maybe I should give you the thesis back again ;-)

Please give me a set of steps which are likely to work well for prediction but are somewhat cheap.
Like Marius did for SAT. So we can use them in the experiments.

[kerschke]

Actually, I am not sure how well the models would work without the two expensive feature steps (local search and curvature). However, as those two groups are soooo expensive, I suggest to drop them.

[berndbischl]

Please dont close when we dont have a final answer....

[berndbischl]

I still need to decide the step set

[kerschke]

Why? We talked about it last Friday... :-)

[kerschke]

@berndbischl Can this be closed now?

[berndbischl]

I will use this later to finally discuss bbob. whether it is completely ok.

[berndbischl]

We need to do 2 things:

• check data set
• decide default feature steps

[kerschke]

I thought, we've already decided about the default feature steps:

• initial_design
• meta
• convexity
• y_distr
• levelset

[berndbischl]

Well, maybe you proposed these, but - because I had no time - I never ran on BBOB again.

It is also good that they are now written down here.

[kerschke]

They are also mentioned in the config-file :-)