Analysis Workflow | Tunny docs

This page walks through a recommended analysis workflow. Each step names the widgets to use and links to the theory pages that explain the methods behind them.

If you are analyzing optimization results for the first time, steps 1, 2, and 6 are often enough. Steps 3, 4, and 5 provide information for deeper analysis.

flowchart TD
    s1["1. Load data into the Dashboard"] --> s2["2. Get the big picture"]
    s2 --> s3["3. Understand distributions and relationships"]
    s3 --> s4["4. Find the parameters that matter"]
    s4 --> s5["5. See the shape of the results"]
    s5 --> s6["6. Pick good trials"]
    s6 --> s7["7. Validate the chosen design"]
    s7 --> s8["8. Explore further (optional)"]
    s8 -. "feed new candidates into the next optimization" .-> s1
    s2 -. "when you need results first" .-> s6

1. Load data into the Dashboard

Open the optimization result you want to analyze. The dashboard reads Optuna storage files directly (Journal / SQLite / PostgreSQL / MySQL), so no conversion is needed.

Details: Supported storage formats and Session files (what view state is saved and restored)

2. Get the big picture

Before any detailed analysis, check whether the optimization itself went well.

For single-objective optimization

Skim the raw results in the Trial Table and Optimization History to check the number of trials and the improvement trend of the objective function.
For example, check whether the objective function improves as the number of trials increases. If it does not improve, the optimization is not going well.

For multi-objective optimization

Check the convergence status with the Convergence Indicators chart, which tracks Hypervolume, IGD+, ε-indicator, and R2 on a shared scale so multiple studies can be compared. See that page for how to choose and define each indicator.
If these charts are still climbing (or decreasing) and have not flattened out, it indicates that the optimization still has room for improvement.

3. Understand distributions and relationships

Survey how parameters and objectives are distributed and how they relate to each other.

One variable at a time: Histogram (skew, multimodality, outliers), Box Plot (side-by-side comparison)
All pairs at once: Correlation Matrix for a heatmap overview, then the Scatter Matrix to inspect the actual shapes behind interesting correlations

4. Find the parameters that matter

Use the Importance Chart to rank parameters by their influence on each objective. A practical two-stage approach:

First, check the results of Spearman or Ridge.
- Both are fast and signed (they show the direction of the effect), but they do not account for interactions or nonlinearities and assume a linear relationship.
To check interactions or nonlinearities, use a more computationally intensive method such as Sobol (interactions), SHAP (explainability), or Permutation (stability).

The selection guide compares all nine methods and when to use each.

5. See the shape of the results

Once you know which parameters matter, look at how they affect the objectives.

One parameter: 1D PDP or the Slice Chart
Two parameters: 2D PDP on a surrogate model, or the Observed Contour if you want to stay on observed values only (no model, no extrapolation)
Local landscape around a candidate: Response Surface 3D

The choice of surrogate model (Ridge / Random Forest / GP variants) involves a trade-off between speed and expressiveness. For details, see the model comparison.

6. Pick good trials

Narrow the trials down to the designs you would actually adopt.

See the full trade-off: Pareto Scatter 2D (Pareto Scatter 3D for three objectives)
Rank trials with a single overall score using MCDM methods:
- TOPSIS (fast, intuitive), VIKOR (balance + worst case), PROMETHEE (pairwise ordering).
  - Weights can be set manually or derived objectively with the Entropy Weight Method.
Group similar trials to make selection easier:
- Clustering methods (k-means, hierarchical, PCA Biplot, SOM Map) group trials with similar behavior so you can pick a representative from each cluster, making the selection much easier. See the clustering overview.
Compare the short-listed candidates side by side: Radar Comparison for the shape and Comparison Table for exact numbers.

7. Validate the chosen design

A design that is optimal in computation may perform differently in practice due to manufacturing tolerances and environmental variations.

Check stability under input noise with the Robustness widget (Monte Carlo noise propagation on a surrogate)
If the design space has distinct regions, cluster the trials (k-means, hierarchical, PCA Biplot, SOM Map) to confirm your candidate is not an isolated outlier. See the clustering overview.

8. Explore further (optional)

If no trial is good enough, use the Surrogate Optimizer to search the fitted model for promising new candidates. Running the optimizer on the surrogate lets you look for regions that improve on the trials you already have, without spending new evaluations.