OptimInstance$eval_batch ignores .objective_multiplicator

[mb706]

OptimInstance$objective_function multiplies objectives that should be maximized with -1, but $eval_batch does not.

(Also, I hope you guys thought this through with $objective_function returning a value that differs from what will be in the archive)

[mb706]

P.S. This line would probably fail if a param has more than one tag (e.g. a tag that is neither "minimize" nor "maximize")

https://github.com/mlr-org/bbotk/blob/166e3e510ca28eb6ce4e0580de28f43b2a7ad911/R/OptimInstance.R#L55

[jakob-r]

OptimInstance$objective_function is just an interface so that you can use the OptimInstance from external optimizers like optim(). So we do not want to log the (eventually) negated output. In all other parts of bbotk we never negate y for the output. Do you have a suggestion how we can make this more clear / Do you still think there is something wrong?

PS: I will fix the tag issue.

[jakob-r]

Tag issue is fixed

[mb706]

Some thoughts on this:

• I found it confusing that objective_function behaves differently from eval_batch, I would prefer if they behave in the same way
• Maybe the best solution is to have an argument that switches between actual values and values multiplied to minimization, for all things that get objective values (objective_function, eval_batch, things with which the archive is accessed, etc)
• objective_multiplicator should probably be publicly accessible

[berndbischl]

• objective_multiplicator should probably be publicly accessible

this we can do

for the rest you would need to argue @mb706 why you really need this. atm interface and semantics seem clear to me. which is:

• you get "unchanged" values everywhere.
• OptimInstance$objective_function is convenience sugar - which is always minimized

seems pretty clear and good to me?

[be-marc]

objective_multiplicator should probably be publicly accessible

Done by https://github.com/mlr-org/bbotk/commit/e142962c55c00e11b62e796166c0886782094d11.