« Back to top page

DisCoMBO: Expert-in-the-Loop Black-Box Optimization via Distributional Conformance

DisCoMBO samples candidates from a probabilistic circuit conditioned on good objective values, selects them by their distributional conformance score, and lets experts inject beliefs whose conformance with the observations it reports.

Abstract

DisCoMBO is the sampler of the NeurIPS 2026 paper DisCoMBO: Steering Expert-in-the-Loop Black-Box Optimization via Distributional Conformance. Instead of a Gaussian process or an ensemble, it uses a probabilistic circuit (a sum-product network) as a generative surrogate of the hyperparameters and the objective value. Candidates are sampled from the circuit conditioned on good objective values and selected by their Distributional Conformance score (DisCo): a bounded, [0, 1]-normalized measure of how much a candidate surprises the circuit. DisCo recovers properties comparable to kernel-based uncertainty known from Gaussian processes while maintaining linear-time inference, and DisCoMBO cycles between exploring surprising and exploiting conforming candidates. The paper proves that DisCoMBO is a zero-regret algorithm.

Expert knowledge is injected directly as interventions, i.e. beliefs about good parameter values, via conditional sampling from the circuit. For each intervention, DisCoMBO reports its conformance with the observations, so that misleading beliefs can be detected; the influence of an intervention decays over the following iterations.

The sampler handles mixed search spaces: continuous, integer (also with log scale or steps) and categorical parameters.

APIs

  • DisCoMBOSampler(search_space: dict[str, BaseDistribution] | None = None, *, seed: int | None = None, n_startup_trials: int = 20, batch_size: int = 10, **optimizer_kwargs)
    • search_space: A dictionary of Optuna distributions. If not provided, the search space is inferred from the completed trials; parameters outside of it (e.g. conditional parameters) are sampled at random.
    • seed: Seed for the random number generators.
    • n_startup_trials: Number of random trials before DisCoMBO takes over.
    • batch_size: Number of configurations DisCoMBO proposes per iteration, i.e. per learned circuit. The proposals are assigned to the following trials, also to trials running in parallel.
    • optimizer_kwargs: Further arguments of discombo.DisCoOptimizer, e.g. stratified_pool_size (number of candidates sampled per iteration) or stratified_cycle_length (iterations from exploration to exploitation).
  • DisCoMBOSampler.intervene(beliefs: dict[str, Any], trial_number: int): Injects expert beliefs into the batch DisCoMBO proposes at or after trial trial_number. beliefs maps parameter names to
    • a value (a point intervention), given like the values of the parameter, e.g. "adam" or 1e-3, or
    • a distribution {"dist": ..., "parameters": [...]}: "gauss" (mean, std) or "uniform" (low, high) for continuous parameters, in natural-log space for parameters with log=True; "cat" (one weight per choice) or "int_uniform" (low and exclusive high choice index) for categorical parameters and integer parameters with fewer than 10 values.
  • DisCoMBOSampler.get_disco_scores() -> list[dict[str, Any]]: The distributional conformance (DisCo) score of each applied intervention: the trial number, the beliefs (in DisCoMBO’s representation) and the mean DisCo score disco_score of the beliefs w.r.t. the circuit conditioned on the best observed value. Low scores indicate that the beliefs contradict the observations.

DisCoMBOSampler supports single-objective optimization. Failed and pruned trials are ignored.

Installation

DisCoMBO is available on PyPI. Install the dependencies of this sampler with

pip install -r https://hub.optuna.org/samplers/discombo/requirements.txt

DisCoMBO runs on the CPU. On Linux, pip installs the CUDA build of PyTorch by default, which takes several GB; to get the much smaller CPU build, run pip install torch --index-url https://download.pytorch.org/whl/cpu first.

Example

import optuna
import optunahub


def objective(trial: optuna.Trial) -> float:
    x = trial.suggest_float("x", -5, 5)
    n = trial.suggest_int("n", 1, 8)
    kind = trial.suggest_categorical("kind", ["a", "b", "c"])
    return x**2 + abs(n - 3) + (kind != "b")


sampler = optunahub.load_module(package="samplers/discombo").DisCoMBOSampler(seed=0)
sampler.intervene({"kind": "b"}, trial_number=30)  # an expert believes that "b" is a good choice
study = optuna.create_study(sampler=sampler)
study.optimize(objective, n_trials=100)
print(study.best_params, sampler.get_disco_scores()[0]["disco_score"])

See example.py for an example with a distribution as belief.

Others

DisCoMBO itself is developed at github.com/bewit/DisCoMBO and licensed under the Apache License 2.0; this sampler wraps it under the MIT License.

Reference

Jonas Seng, Bennet Wittelsbach, and Kristian Kersting. DisCoMBO: Steering Expert-in-the-Loop Black-Box Optimization via Distributional Conformance. In Advances in Neural Information Processing Systems (NeurIPS), 2026.

Bibtex

@inproceedings{seng2026discombo,
  title         = {{DisCoMBO}: Steering Expert-in-the-Loop Black-Box Optimization via Distributional Conformance},
  author        = {Seng, Jonas and Wittelsbach, Bennet and Kersting, Kristian},
  booktitle     = {Advances in Neural Information Processing Systems},
  volume        = {40},
  year          = {2026},
  eprint        = {2609.36472},
  archivePrefix = {arXiv},
  primaryClass  = {cs.LG},
  url           = {https://arxiv.org/abs/2609.36472}
}
Package
samplers/discombo
Author
Jonas Seng, Bennet Wittelsbach, Kristian Kersting
License
MIT License
Verified Optuna version
  • 4.9.0
  • 5.0.0
Dependencies (.txt)
  • optuna
  • optunahub
  • discombo>=1.2,<2
Last update
2026-11-13
Discussions & Issues
Create a discussion
Create a bug report