Optimizer using the Hyperband (HB) algorithm.
HB runs the Successive Halving Algorithm (SHA) with different numbers of stating configurations.
The algorithm is initialized with the same parameters as Successive Halving but without n.
Each run of Successive Halving is called a bracket and starts with a different budget r_0.
A smaller starting budget means that more configurations can be tried out.
The most explorative bracket allocated the minimum budget r_min.
The next bracket increases the starting budget by a factor of eta.
In each bracket, the starting budget increases further until the last bracket s = 0 essentially performs a random search with the full budget r_max.
The number of brackets s_max + 1 is calculated with s_max = log(r_min / r_max)(eta).
Under the condition that r_0 increases by eta with each bracket, r_min sometimes has to be adjusted slightly in order not to use more than r_max resources in the last bracket.
The number of configurations in the base stages is calculated so that each bracket uses approximately the same amount of budget.
The following table shows a full run of HB with eta = 2, r_min = 1 and r_max = 8.
s | 3 | 2 | 1 | 0 | ||||||||
i | n_i | r_i | n_i | r_i | n_i | r_i | n_i | r_i | ||||
| 0 | 8 | 1 | 6 | 2 | 4 | 4 | 8 | 4 | ||||
| 1 | 4 | 2 | 3 | 4 | 2 | 8 | ||||||
| 2 | 2 | 4 | 1 | 8 | ||||||||
| 3 | 1 | 8 |
s is the bracket number, i is the stage number, n_i is the number of configurations and r_i is the budget allocated to a single configuration.
The budget hyperparameter must be tagged with "budget" in the search space.
The minimum budget (r_min) which is allocated in the base stage of the most explorative bracket, is set by the lower bound of the budget parameter.
The upper bound defines the maximum budget (r_max) which is allocated to the candidates in the last stages.
Source
Li L, Jamieson K, DeSalvo G, Rostamizadeh A, Talwalkar A (2018). “Hyperband: A Novel Bandit-Based Approach to Hyperparameter Optimization.” Journal of Machine Learning Research, 18(185), 1-52. https://jmlr.org/papers/v18/16-558.html.
Dictionary
This Tuner can be instantiated via the dictionary
mlr_tuners or with the associated sugar function tnr():
TunerHyperband$new()
mlr_tuners$get("hyperband")
tnr("hyperband")Subsample Budget
If the learner lacks a natural budget parameter, mlr3pipelines::PipeOpSubsample can be applied to use the subsampling rate as budget parameter. The resulting mlr3pipelines::GraphLearner is fitted on small proportions of the mlr3::Task in the first stage, and on the complete task in last stage.
Custom Sampler
Hyperband supports custom paradox::Sampler object for initial configurations in each bracket. A custom sampler may look like this (the full example is given in the examples section):
# - beta distribution with alpha = 2 and beta = 5
# - categorical distribution with custom probabilities
sampler = SamplerJointIndep$new(list(
Sampler1DRfun$new(params[[2]], function(n) rbeta(n, 2, 5)),
Sampler1DCateg$new(params[[3]], prob = c(0.2, 0.3, 0.5))
))Progress Bars
$optimize() supports progress bars via the package progressr
combined with a Terminator. Simply wrap the function in
progressr::with_progress() to enable them. We recommend to use package
progress as backend; enable with progressr::handlers("progress").
Parallelization
This hyperband implementation evaluates hyperparameter configurations of equal budget across brackets in one batch.
For example, all configurations in stage 1 of bracket 3 and stage 0 of bracket 2 in one batch.
To select a parallel backend, use the plan() function of the future package.
Logging
Hyperband uses a logger (as implemented in lgr) from package
bbotk.
Use lgr::get_logger("bbotk") to access and control the logger.
Resources
The gallery features a collection of case studies and demos about optimization.
Tune the hyperparameters of XGBoost with Hyperband.
Use data subsampling and Hyperband to optimize a support vector machine.
Parameters
etanumeric(1)
With every stage, the budget is increased by a factor ofetaand only the best1 / etapoints are promoted to the next stage. Non-integer values are supported, butetais not allowed to be less or equal to 1.samplerparadox::Sampler
Object defining how the samples of the parameter space should be drawn in the base stage of each bracket. The default is uniform sampling.repetitionsinteger(1)
If1(default), optimization is stopped once all brackets are evaluated. Otherwise, optimization is stopped afterrepetitionsruns of HB. The bbotk::Terminator might stop the optimization before all repetitions are executed.
Archive
The bbotk::Archive holds the following additional columns that are specific to HB:
bracket(integer(1))
The bracket index. Counts down to 0.stage(integer(1))
The stages of each bracket. Starts counting at 0.repetition(integer(1))
Repetition index. Start counting at 1.
Super classes
mlr3tuning::Tuner -> mlr3tuning::TunerFromOptimizer -> TunerHyperband
Examples
if(requireNamespace("xgboost")) {
library(mlr3learners)
# define hyperparameter and budget parameter
search_space = ps(
nrounds = p_int(lower = 1, upper = 16, tags = "budget"),
eta = p_dbl(lower = 0, upper = 1),
booster = p_fct(levels = c("gbtree", "gblinear", "dart"))
)
# \donttest{
# hyperparameter tuning on the pima indians diabetes data set
instance = tune(
tnr("hyperband"),
task = tsk("pima"),
learner = lrn("classif.xgboost", eval_metric = "logloss"),
resampling = rsmp("cv", folds = 3),
measures = msr("classif.ce"),
search_space = search_space,
term_evals = 100
)
# best performing hyperparameter configuration
instance$result
# }
}
#> Loading required namespace: xgboost
#> nrounds eta booster learner_param_vals x_domain classif.ce
#> 1: 8 0.899514 gbtree <list[7]> <list[3]> 0.2460938