如何在"悖论"中设置特定值?

Question

有没有办法在 R 包paradox中设置参数的特定值？假设我为随机森林方法进行超参数调优，我想测试mtry = c(2, 3, 7, 8)和min.node.size = c(2, 5, 7)，即值之间距离不相等的 4 x 3 网格。

目前，我必须进行大型 7 x 6 网格搜索以包含这些值，测试我不感兴趣的组合：

tuner_params = ParamSet$new(list(
ParamInt$new("mtry", lower = 2, upper = 7),
ParamInt$new("min.node.size", lower = 2, upper = 6)
))
generate_design_grid(tuner_params, param_resolutions = c(mtry = 7, min.node.size = 5))

Answer 1

克服这个问题的一种方法是不使用网格搜索，而是使用TunerDesignPoints。

请参阅示例：

library(paradox)
library(mlr3)
library(mlr3tuning)
library(mlr3learners)
library(data.table)
tuner_params = ParamSet$new(list(
ParamInt$new("mtry", lower = 2, upper = 8),
ParamInt$new("min.node.size", lower = 2, upper = 7)
))

指定自定义设计点：

design = data.table(expand.grid(mtry = c(2, 3, 7, 8),
min.node.size = c(2, 5, 7)))
tuner = tnr("design_points", design = design)
sonar_task = tsk("sonar")
r_lrn  = lrn("classif.ranger", predict_type = "prob")
instance = TuningInstance$new(
task = sonar_task,
learner =  r_lrn,
resampling = rsmp("cv", folds = 3),
measures = msr("classif.acc"),
param_set = tuner_params,
terminator = term("none")) #no terminator since you want all design points evaluated

tuner$tune(instance)
instance$archive()

#output

nr batch_nr  resample_result task_id     learner_id resampling_id iters params tune_x warnings errors classif.acc
1:  1        1 <ResampleResult>   sonar classif.ranger            cv     3 <list> <list>        0      0   0.8462388
2:  2        2 <ResampleResult>   sonar classif.ranger            cv     3 <list> <list>        0      0   0.8366460
3:  3        3 <ResampleResult>   sonar classif.ranger            cv     3 <list> <list>        0      0   0.8317460
4:  4        4 <ResampleResult>   sonar classif.ranger            cv     3 <list> <list>        0      0   0.8269151
5:  5        5 <ResampleResult>   sonar classif.ranger            cv     3 <list> <list>        0      0   0.8366460
6:  6        6 <ResampleResult>   sonar classif.ranger            cv     3 <list> <list>        0      0   0.8173913
7:  7        7 <ResampleResult>   sonar classif.ranger            cv     3 <list> <list>        0      0   0.8221532
8:  8        8 <ResampleResult>   sonar classif.ranger            cv     3 <list> <list>        0      0   0.8124914
9:  9        9 <ResampleResult>   sonar classif.ranger            cv     3 <list> <list>        0      0   0.8415459
10: 10       10 <ResampleResult>   sonar classif.ranger            cv     3 <list> <list>        0      0   0.8173223
11: 11       11 <ResampleResult>   sonar classif.ranger            cv     3 <list> <list>        0      0   0.8221532
12: 12       12 <ResampleResult>   sonar classif.ranger            cv     3 <list> <list>        0      0   0.8221532

12 个点的评估就像我们在设计网格中指定的那样。