Tune Hyperparameters

Hyperparameter tuning finds the optimal settings for your model. MLOps Desktop integrates Optuna for intelligent hyperparameter search with three strategies: Bayesian (TPE), Random, and Grid.

Time to complete: ~15 minutes

Prerequisites

• Completed the Quickstart tutorial
• Python packages: pip install optuna

Enable Hyperparameter Tuning

1. Open your pipeline

   Load a pipeline with DataLoader → DataSplit → Trainer → Evaluator.

2. Select Tune mode

   Click the Trainer node. At the top of the node, click the Tune button (next to Train and Load).

3. Configure model and target

   • Model Type: Select any model except Linear Regression (no tunable params)
   • Target Column: Your target variable

4. Open Tuning Configuration

   Click the Configure Tuning button to open the TuningPanel.

5. Set tuning parameters

   Setting	Recommended	Range
   Search Strategy	Bayesian (TPE)	TPE, Random, Grid
   Number of Trials	50	1-1000
   CV Folds	3	2-10
   Scoring Metric	accuracy (classification) or r2 (regression)	varies

6. Run the pipeline

   Click Run. Watch the Trials tab for real-time results.

Search Strategies

Bayesian (TPE)

Tree-structured Parzen Estimator — Default and recommended.

• Learns from previous trials to focus on promising regions
• Most efficient for complex search spaces
• Better results with fewer trials

Trial 1-10: Exploration (random-ish)
Trial 11+: Exploitation (focuses on good regions)

Random

Random Search — Simple uniform sampling.

• Each trial is independent
• Good baseline for comparison
• Works well with large search spaces

Every trial: Random sample from full space

Grid

Grid Search — Exhaustive enumeration.

• Tests every combination
• Only practical for small, discrete spaces
• Warning shown if >10,000 combinations

All combinations of:
n_estimators: [50, 100, 150]
max_depth: [5, 10, 15]
= 9 total trials

Search Spaces by Model

Each model has predefined search ranges optimized for common use cases:

Random Forest

Parameter	Range	Type
n_estimators	50-300 (step 50)	Integer
max_depth	[None, 10, 15, 20, 30]	Categorical
min_samples_split	2-10 (step 2)	Integer
min_samples_leaf	1-4 (step 1)	Integer

Gradient Boosting

Parameter	Range	Type
n_estimators	50-300 (step 50)	Integer
learning_rate	0.01-0.3	Log-uniform
max_depth	3-8 (step 1)	Integer
subsample	0.7-1.0	Uniform

SVM (SVC/SVR)

Parameter	Range	Type
C	0.1-100	Log-uniform
kernel	[rbf, linear, poly]	Categorical
gamma	[scale, auto]	Categorical

KNN

Parameter	Range	Type
n_neighbors	3-21 (step 2)	Integer
weights	[uniform, distance]	Categorical
metric	[euclidean, manhattan, minkowski]	Categorical

MLP Neural Network

Parameter	Range	Type
hidden_layer_sizes	[(50,), (100,), (100,50), (100,100)]	Categorical
alpha	0.0001-0.1	Log-uniform
learning_rate_init	0.0001-0.1	Log-uniform
max_iter	200-1000 (step 100)	Integer

Logistic Regression

Parameter	Range	Type
C	0.01-100	Log-uniform
max_iter	500-2000 (step 500)	Integer

Note

Linear Regression cannot be tuned — it has no hyperparameters. Select a different model for tuning.

Scoring Metrics

Choose the metric to optimize:

Classification:

Metric	Best For
accuracy	Balanced classes (default)
f1	Imbalanced classes
precision	Minimize false positives
recall	Minimize false negatives
roc_auc	Ranking quality

Regression:

Metric	Best For
r2	Overall fit (default)
neg_mse	Penalize large errors
neg_mae	Robust to outliers
neg_rmse	Same units as target

Trials Tab

During tuning, the Trials tab shows real-time results:

Column	Description
#	Trial number
Score	Cross-validation score (higher = better)
Parameters	Hyperparameter values tried
Duration	Time for this trial
Status	Complete, Pruned, or Failed

The best trial is highlighted with a star icon.

Example Output

Trial 1:  0.823  n_estimators=150, max_depth=10     2.1s
Trial 2:  0.845  n_estimators=200, max_depth=15     2.8s
Trial 3:  0.831  n_estimators=100, max_depth=None   1.9s
...
★ Trial 17: 0.867  n_estimators=250, max_depth=20   3.2s  ← Best

Cross-Validation

Tuning uses k-fold cross-validation (not a simple train/test split):

1. Training data split into k folds
2. Model trained k times, each time validating on a different fold
3. Scores averaged for final trial score

This provides more reliable scoring than a single split.

CV Folds	Trade-off
2	Fast, high variance
3	Good balance (default)
5	More reliable, slower
10	Most reliable, slowest

After Tuning

When tuning completes:

1. Best parameters are displayed in the Trials tab
2. Final model is trained on full training data with best params
3. Evaluator receives the tuned model for metrics

The best configuration is automatically applied—no manual copying needed.

Tips for Better Results

Start Small

Begin with 20-30 trials to understand the landscape:

Trial 1-30: Quick exploration
→ Review results
→ Narrow search space if needed
→ Run 50-100 more trials

Use TPE for Most Cases

Bayesian (TPE) outperforms Random search in almost all scenarios. Only use Grid when:

• You have a very small search space
• You need exhaustive coverage

Increase CV Folds for Small Data

With fewer than 1,000 samples, use 5-fold CV to get reliable scores.

Check for Overfitting

If tuning score is much higher than evaluation score, your model may be overfitting to the validation folds. Try:

• Simpler model (fewer estimators, shallower trees)
• More regularization (C for SVM, alpha for MLP)

Common Issues

”Optuna not installed"

pip install optuna

"Linear Regression cannot be tuned”

Linear Regression has no hyperparameters. Use Ridge, Lasso, or another model.

Tuning takes too long

• Reduce number of trials (20-30 for quick tests)
• Use Random instead of Grid search
• Reduce CV folds to 2
• Use a simpler model (Logistic Regression vs MLP)

Grid search warning

If Grid search shows “>10,000 combinations”, the search space is too large. Switch to TPE or Random.

Next Steps

Explain Predictions

Understand why your tuned model makes predictions

Track Experiments

Compare tuning runs across experiments