Blackbox Optimization

The blackboxopt Python package contains blackbox optimization algorithms with a common interface, along with useful helpers like parallel optimization loops, analysis and visualization tools.

Key Features

Common Interface

The blackboxopt base classes along with the EvaluationSpecification and Evaluation data classes specify a unified interface for different blackbox optimization method implementations. In addition to these interfaces, a standard pytest compatible testsuite is available to ensure functional compatibility of an optimizer implementation with the blackboxopt framework.

Optimizers

Aside from random search and a Sobol sequence based space filling method, the main ones in this package are Hyperband, BOHB and a BoTorch based Bayesian optimization base implementation. BOHB is provided as a cleaner replacement of the former implementation in HpBandSter.

Optimization Loops

As part of the blackboxopt.optimization_loops module compatible implementations for optimization loops are avilable bot for local, serial execution as well as for distributed optimization via dask.distributed.

Visualizations

Interactive visualizations like objective value over time or duration for single objective optimization, as well as an objectives pair plot with a highlighted pareto front for multi objective optimization is available as part of the blackboxopt.visualizations module.

Getting Started

The following example outlines how a quadratic function can be optimized with random search in a distributed manner.

# Copyright (c) 2020 - for information on the respective copyright owner
# see the NOTICE file and/or the repository https://github.com/boschresearch/blackboxopt
#
# SPDX-License-Identifier: Apache-2.0

import parameterspace as ps

try:
    import dask.distributed as dd
except ImportError:
    raise ImportError(
        "Unable to import Dask Distributed specific dependencies. "
        + "Make sure to install blackboxopt[dask]"
    )

from blackboxopt import Evaluation, EvaluationSpecification, Objective
from blackboxopt.optimization_loops.dask_distributed import (
    run_optimization_loop,
)
from blackboxopt.optimizers.random_search import RandomSearch


def evaluation_function(eval_spec: EvaluationSpecification) -> Evaluation:
    return eval_spec.create_evaluation(
        objectives={"loss": eval_spec.configuration["p1"] ** 2},
        user_info={"weather": "sunny"},
    )


if __name__ == "__main__":
    space = ps.ParameterSpace()
    space.add(ps.ContinuousParameter("p1", (-1.0, 1.0)))
    optimizer = RandomSearch(
        space,
        [Objective("loss", greater_is_better=False)],
        max_steps=1000,
    )

    evaluations = run_optimization_loop(
        optimizer, evaluation_function, dd.Client(), max_evaluations=100
    )

    n_successes = len([e for e in evaluations if not e.all_objectives_none])
    print(f"Successfully evaluated {n_successes}/{len(evaluations)}")

License

blackboxopt is open-sourced under the Apache-2.0 license. See the LICENSE file for details.

For a list of other open source components included in blackboxopt, see the file 3rd-party-licenses.txt.