Model Tuning Patterns

Overview

Comprehensive hyperparameter optimization using the tune package. Covers grid search, iterative search, racing methods, and Bayesian optimization.

Tunable Parameters

Marking Parameters for Tuning

library(tidymodels)

# Use tune() placeholder in model specification
rf_spec <- rand_forest(
  mtry = tune(),
  trees = 1000,
  min_n = tune()
) |>
  set_engine("ranger") |>
  set_mode("classification")

# Tune recipe steps
rec <- recipe(outcome ~ ., data = train) |>
  step_pca(all_numeric_predictors(), num_comp = tune()) |>
  step_normalize(all_numeric_predictors())

Parameter Objects (dials)

library(dials)

# View parameter information
mtry()
min_n()
trees()
learn_rate()
penalty()

# Customize parameter ranges
mtry(range = c(2, 20))
min_n(range = c(5, 50))
learn_rate(range = c(-3, -1), trans = log10_trans())

# Update based on data
mtry_final <- finalize(mtry(), train_data)

Grid Search

Regular Grid

# Evenly spaced grid
regular_grid <- grid_regular(
  mtry(range = c(2, 10)),
  min_n(range = c(2, 20)),
  levels = 5  # 5 levels per parameter = 25 combinations
)

# Different levels per parameter
regular_grid <- grid_regular(
  mtry(range = c(2, 10)),
  min_n(range = c(2, 20)),
  levels = c(mtry = 5, min_n = 3)
)

Random Grid

# Random sampling of parameter space
random_grid <- grid_random(
  mtry(range = c(2, 10)),
  min_n(range = c(2, 20)),
  size = 50
)

Space-Filling Designs

# Latin hypercube (better coverage than random)
lhs_grid <- grid_latin_hypercube(
  mtry(range = c(2, 10)),
  min_n(range = c(2, 20)),
  size = 30
)

# Maximum entropy grid
maxent_grid <- grid_max_entropy(
  mtry(range = c(2, 10)),
  min_n(range = c(2, 20)),
  size = 30
)

Running Grid Search

# Basic grid search
tune_results <- workflow |>
  tune_grid(
    resamples = cv_folds,
    grid = 20,  # auto-generates grid
    metrics = metric_set(roc_auc, accuracy),
    control = control_grid(verbose = TRUE)
  )

# With explicit grid
tune_results <- workflow |>
  tune_grid(
    resamples = cv_folds,
    grid = my_grid,
    metrics = metric_set(roc_auc, accuracy)
  )

Bayesian Optimization

Basic Bayesian Tuning

# Iterative Bayesian search
tune_results <- workflow |>
  tune_bayes(
    resamples = cv_folds,
    iter = 50,           # maximum iterations
    initial = 10,        # initial grid points
    metrics = metric_set(roc_auc),
    control = control_bayes(
      verbose = TRUE,
      no_improve = 20    # stop if no improvement for 20 iters
    )
  )

Bayesian with Custom Initial Points

# Start with grid results
initial_grid <- workflow |>
  tune_grid(
    resamples = cv_folds,
    grid = 10,
    metrics = metric_set(roc_auc)
  )

# Continue with Bayesian
bayes_results <- workflow |>
  tune_bayes(
    resamples = cv_folds,
    iter = 40,
    initial = initial_grid,  # use grid results as initial points
    metrics = metric_set(roc_auc)
  )

Acquisition Functions

control_bayes(
  # Expected improvement (default)
  objective = exp_improve(),

  # Probability of improvement
  objective = prob_improve(),

  # Confidence bound
  objective = conf_bound(kappa = 2)
)

Racing Methods (finetune)

ANOVA Racing

library(finetune)

# Eliminate poor configurations early
race_results <- workflow |>
  tune_race_anova(
    resamples = cv_folds,
    grid = 50,
    metrics = metric_set(roc_auc),
    control = control_race(
      verbose = TRUE,
      burn_in = 3  # evaluate all for first 3 resamples
    )
  )

# Plot racing results
plot_race(race_results)

Win/Loss Racing

# Statistical comparison of configurations
race_results <- workflow |>
  tune_race_win_loss(
    resamples = cv_folds,
    grid = 50,
    metrics = metric_set(roc_auc),
    control = control_race()
  )

Simulated Annealing

library(finetune)

# Simulated annealing search
sa_results <- workflow |>
  tune_sim_anneal(
    resamples = cv_folds,
    iter = 50,
    initial = 5,
    metrics = metric_set(roc_auc),
    control = control_sim_anneal(
      verbose = TRUE,
      cooling_coef = 0.1
    )
  )

Analyzing Tuning Results

Best Parameters

# Best by primary metric
best_params <- select_best(tune_results, metric = "roc_auc")

# Best within one SE of optimal (more parsimonious)
best_params <- select_by_one_std_err(
  tune_results,
  metric = "roc_auc",
  mtry, min_n  # parameters to minimize
)

# Best by percent loss from optimal
best_params <- select_by_pct_loss(
  tune_results,
  metric = "roc_auc",
  limit = 2  # within 2% of best
)

Visualizing Results

# Performance across parameters
autoplot(tune_results)

# Specific parameter effects
autoplot(tune_results, type = "marginals")

# Performance vs parameters
autoplot(tune_results, type = "parameters")

# Collect all metrics
metrics_df <- collect_metrics(tune_results)

Finalizing Model

# Finalize workflow with best parameters
final_wf <- workflow |>
  finalize_workflow(best_params)

# Final fit on all training data
final_fit <- final_wf |>
  last_fit(data_split)

# Extract metrics on test set
collect_metrics(final_fit)

Parallel Processing

library(doParallel)

# Register parallel backend
cl <- makePSOCKcluster(parallel::detectCores() - 1)
registerDoParallel(cl)

# Tuning will automatically use parallel
tune_results <- workflow |>
  tune_grid(
    resamples = cv_folds,
    grid = 100
  )

# Stop cluster
stopCluster(cl)

Using future

library(future)

# Set up parallel plan
plan(multisession, workers = 4)

# With finetune racing
race_results <- workflow |>
  tune_race_anova(
    resamples = cv_folds,
    grid = 100,
    control = control_race(parallel_over = "everything")
  )

Tuning XGBoost Example

# XGBoost with many tunable parameters
xgb_spec <- boost_tree(
  trees = tune(),
  tree_depth = tune(),
  min_n = tune(),
  loss_reduction = tune(),
  sample_size = tune(),
  mtry = tune(),
  learn_rate = tune()
) |>
  set_engine("xgboost") |>
  set_mode("classification")

# Define parameter grid
xgb_grid <- grid_latin_hypercube(
  trees(range = c(100, 1500)),
  tree_depth(range = c(3, 15)),
  min_n(range = c(2, 30)),
  loss_reduction(),
  sample_prop(range = c(0.5, 1)),
  finalize(mtry(), train_data),
  learn_rate(range = c(-3, -1)),
  size = 50
)

# Use racing for efficiency
xgb_results <- workflow(rec, xgb_spec) |>
  tune_race_anova(
    resamples = cv_folds,
    grid = xgb_grid,
    metrics = metric_set(roc_auc)
  )

Control Options Summary

Best Practices

1. Start with grid search to understand parameter space
2. Use racing methods for large grids (>50 combinations)
3. Use Bayesian optimization for expensive models
4. Always set a seed for reproducibility
5. Use stratified resampling for imbalanced outcomes
6. Consider select_by_one_std_err for simpler models
7. Monitor for overfitting during iterative search