使用 mlr3 进行超参数调整的贝叶斯优化

Question

Are there any hyperparameter tuners using Bayesian optimization within the mlr3 ecosystem? In particular as an argument in the wrapper function

tuner = tnr("grid_search", resolution = 10)

?

Otherwise, what would be a recommended tuning method supported by mlr3 ?

UPDATE: I tried following @Sebastian's mlr3mbo link -- in particular the "Simple Tuning Example Using Defaults" section.

However, when I tried to use it in tune_nested

rr = tune_nested(
    method = "mbo", 
    task,
    learner=lrn("surv.cv_glmnet",  nfolds=3, type.measure = "C", alpha = to_tune(0.01, 1)) ,
    inner_resampling = rsmp("cv", folds = 10), 
    outer_resampling = rsmp("cv", folds = 10), 
    measure = msr("surv.cindex"), 
    term_evals = 20
    )

I get the error:

Error in auto_tuner(method, learner, inner_resampling, measure, term_evals,  : 
  Assertion on 'method' failed: Must be element of set {'cmaes','design_points','gensa','grid_search','irace','nloptr','random_search'}, but is 'mbo'.
Calls: tune_nested ... auto_tuner -> assert_choice -> makeAssertion -> mstop

UPDATE2 : after using devtools::install_github("mlr-org/mlr3mbo@add_tuner") as @Sebastian helpfully suggested, I get instead the following error message:

Error in .__ParamSet__search_space(self = self, private = private, super = super,  : 
  Assertion on 'names(values)' failed: Must have names.

Does alpha = to_tune(0.01, 1) within lrn() not sufficiently specify the required name and value range of the HP alpha ?

Is there any way around this? Thanks!

SOLUTION : Thanks to @Sebastian who fixed this -- in his comment:

manually define the search_space like search_space = ps(alpha = p_dbl(0.01, 1)) and then pass it as the search_space argument to tune_nested

Super!

Answer 1

Just checking in to add a few comments and answer your latest question @Alex: A minor note on the error:

Error in .__ParamSet__search_space(self = self, private = private, super = super,  : 
  Assertion on 'names(values)' failed: Must have names.

This was rather caused due to old bbotk / paradox versions that have been fixed upstream and you most likely should no longer encounter any errors anymore. The manual fix is no longer needed.

The BO Tuner will live in mlr3mbo and a PR will be merged soon.

In general mlr3mbo is slowly but steady getting to a point where it can be considered stable and will be pushed to CRAN. I do agree that some documentation is still missing. Also, one or two gallery posts about the inner workings of mlr3mbo will be provided in the near future.

Regarding:

One last question: is there a reference for how the optimiser works, eg the maths behind it? I searched and only found mlr3mbo.mlr-org.com/reference/mlr_optimizers_mbo.html which doesn't have many details.

All current implementations are somewhat standard/basic BO stuff, therefore we currently did not included detailed descriptions of the algorithms (but we may do this in the future).

(sequential) (single-criteria) BO (as eg, implemented in bayesopt_ego ) works by fitting a Surrogate model to your evaluated points, learning the relationship between inputs and outputs. Based on an acquisition function AcqFunction (relying on this surrogate) that encodes for each potential next candidate to evaluate the gain in evaluation (ie, how much can I expect to improve my solution if I evaluate this candidate) you find the next point to query, evaluate it and update your archive of evaluated points and start over again with refitting the surrogate.

I highlighted the respective classes/functions and most of them include references to standard papers where you can learn more about the maths/technical details.

Also, if you find mlr3mbo useful or have any further comments/questions, feel free to also communicate with us via GitHub https://github.com/mlr-org/mlr3mbo . Issues/open questions etc. are welcome!