Brute Force Tree Plot

Abstract

optuna.samplers.BruteForceSampler explores a define-by-run search space as a tree: each internal node is a parameter, and a path from the root to a leaf is one trial’s complete params. This package renders that tree as an interactive icicle chart, so you can see which branches have been sampled, how each trial’s parameters relate to one another, and how much of a finite (grid-like) search space is still unexplored.

Leaves are colored by trial state (complete/pruned/failed/running); hovering over a complete trial’s leaf shows its objective value(s). A synthetic ... unexplored leaf is added under every branching point that still has unvisited candidate values.

APIs

plot_brute_force_tree(study)

• study (optuna.Study): A study, ideally optimized with BruteForceSampler. Any study works, but the “unexplored” counts are only meaningful for finite, grid-like search spaces (e.g. IntDistribution or FloatDistribution with step set, or CategoricalDistribution).

Returns a plotly.graph_objects.Figure.

Example

import optuna
import optunahub


module = optunahub.load_module("visualization/plot_brute_force_tree")


def objective(trial: optuna.Trial) -> float:
    c = trial.suggest_categorical("c", ["float", "int"])
    if c == "float":
        return trial.suggest_float("x", 1, 3, step=0.5)
    else:
        a = trial.suggest_int("a", 1, 3)
        b = trial.suggest_int("b", a, 3)
        return a + b


study = optuna.create_study(sampler=optuna.samplers.BruteForceSampler(seed=42))
study.optimize(objective, n_trials=30)

fig = module.plot_brute_force_tree(study)
fig.show()

Others

This package only reads public optuna.trial.FrozenTrial fields (params, distributions, state, values), so it does not depend on BruteForceSampler’s private tree implementation and works across Optuna versions.