mlr3基准与嵌套重采样：从benchmark对象提取调优模型计算特征重要性

我在使用mlr3中的benchmark()函数对比多个机器学习算法，其中一个是经过超参数调优的XGB模型。我采用了嵌套重采样方案：外层重采样使用holdout样本评估整体性能，内层重采样使用5折交叉验证完成超参数调优。
除了获取所有机器学习算法的准确率估值外，我还希望查看调优后XGB的特征重要性。要实现这一点需要访问benchmark对象内部的调优模型，但我不知道该如何操作：benchmark()返回的对象是深度嵌套的列表，我不理解它的结构。
已有的公开方案都无法满足我的需求：

• 部分方案使用流水线中的学习器作为示例，和我使用benchmark对象的场景不符
• 部分方案仅展示了如何一次性提取基准测试的所有信息，没有说明如何单独提取基准中某一个学习器的调优后模型

我用于执行嵌套重采样的代码如下，完成基准测试后我需要获取调优后的XGB模型来计算特征重要性：

require(mlr3verse)

### 参数设置

## 调优相关参数
n_folds = 5
grid_search_resolution = 2
measure = msr("classif.acc")
task = tsk("iris")

# 开启mlr3调试信息输出
options("mlr3.debug" = TRUE)

### 超参数调优设置
# 内层重采样用的AutoTuner

## 内层重采样设计
inner_resampling = rsmp("cv", folds = n_folds)
terminator = trm("none")
 
## 未调优的XGB学习器
xgb_no_tuning = lrn("classif.xgboost", eval_metric = "mlogloss")
set_threads(xgb_no_tuning, n = 6)

## XGB自动调优器配置
xgb_learner_tuning = lrn("classif.xgboost", eval_metric = "mlogloss")
xgb_search_space = ps(nrounds = p_int(lower = 100, upper= 500),
                      max_depth = p_int(lower = 3, upper= 10),
                      colsample_bytree = p_dbl(lower = 0.6, upper = 1)
                  )
xgb_tuner = tnr("grid_search", resolution = grid_search_resolution)

# 开启隐式并行
set_threads(xgb_learner_tuning, n = 6)

xgb_tuned = AutoTuner$new(xgb_learner_tuning, inner_resampling, measure, terminator, xgb_tuner, xgb_search_space, store_tuning_instance = TRUE)

## 外层重采样：holdout
outer_resampling = rsmp("holdout")
outer_resampling$instantiate(task)

bm_design = benchmark_grid(
  tasks = task,
  learners = c(lrn("classif.featureless"), 
               xgb_no_tuning,
               xgb_tuned 
  ),
  resamplings = outer_resampling
)

begin_time = Sys.time()
bmr = benchmark(bm_design, store_models = TRUE)
duration = Sys.time() - begin_time

print(duration)

## 基准测试结果
benchmark_results = bmr$aggregate(measure)
print(benchmark_results)


## 概览
mlr3misc::map(as.data.table(bmr)$learner, "model")

## 详细结果
# 输出学习器配置
print(bmr$learners$learner)

require(mlr3verse)
require(mlr3tuning)
require(mlr3misc)

### 参数设置

## 调优相关参数
n_folds = 5
grid_search_resolution = 2
measure = msr("classif.acc")
task = tsk("iris")

# 开启mlr3调试信息输出
options("mlr3.debug" = TRUE)

### 超参数调优设置
# 内层重采样用的AutoTuner

## 内层重采样设计
inner_resampling = rsmp("cv", folds = n_folds)
terminator = trm("none")
 
## 未调优的XGB学习器
xgb_no_tuning = lrn("classif.xgboost", eval_metric = "mlogloss")
set_threads(xgb_no_tuning, n = 6)

## XGB自动调优器配置
xgb_learner_tuning = lrn("classif.xgboost", eval_metric = "mlogloss")
xgb_search_space = ps(nrounds = p_int(lower = 100, upper= 500),
                      max_depth = p_int(lower = 3, upper= 10),
                      colsample_bytree = p_dbl(lower = 0.6, upper = 1)
                  )
xgb_tuner = tnr("grid_search", resolution = grid_search_resolution)

# 开启隐式并行
set_threads(xgb_learner_tuning, n = 6)

xgb_tuned = AutoTuner$new(xgb_learner_tuning, inner_resampling, measure, terminator, xgb_tuner, xgb_search_space, store_tuning_instance = TRUE)

## 外层重采样：holdout
outer_resampling = rsmp("holdout")
outer_resampling$instantiate(task)

bm_design = benchmark_grid(
  tasks = task,
  learners = c(lrn("classif.featureless"), 
               xgb_no_tuning,
               xgb_tuned 
  ),
  resamplings = outer_resampling
)

begin_time = Sys.time()
bmr = benchmark(bm_design, store_models = TRUE)
duration = Sys.time() - begin_time

print(duration)

## 基准测试结果
benchmark_results = bmr$aggregate(measure)
print(benchmark_results)


## 概览
mlr3misc::map(as.data.table(bmr)$learner, "model")

## 详细结果
# 输出学习器配置
print(bmr$learners$learner)

## 特征重要性计算
# 从外层采样结果中提取模型
data = as.data.table(bmr)
outer_learners = map(data$learner, "learner")

# 本示例中调优XGB是第三个学习器，因此取索引为3的元素
xgb_tuned_model = outer_learners[[3]]

print(xgb_tuned_model)

# 输出特征重要性（默认计算指标为增益）
print(xgb_tuned_model$importance())

内容的提问来源于stack exchange，提问作者AW2