loguso 0.1.0

Logarithmically partitioned mixed-variable surrogate optimization.

LOGUSO

loguso is a Python package for mixed-variable surrogate optimization on discretized real, integer, and categorical search spaces. It also includes a categorical extension of the COCO bbob-mixint benchmark family, where categorical assignments are routed to COCO instance numbers through a stable hash.

Installation

Install the package in editable mode from the repository root:

pip install -e . --no-build-isolation

If you also want to build categorical COCO benchmarks, install COCO's Python bindings so that import cocoex works in your environment:

python -c "import cocoex"

What Is Included

• loguso.loguso(...): the main optimizer
• loguso.make_coco_categorical_benchmark(...): build one categorical COCO benchmark task
• loguso.build_coco_categorical_suite(...): build a benchmark suite over selected functions and dimensions
• loguso.CocoCategoricalSuite: lower-level categorical COCO wrapper

Minimal Example

The example below optimizes one categorical COCO function, namely the separable sphere function F1 in total dimension 24.

from loguso import loguso, make_coco_categorical_benchmark

spec = make_coco_categorical_benchmark(
    function_index=1,
    total_dimension=24,
    instances=100,
    real_levels=101,
)

result = loguso(
    exp_func=spec.exp_func,
    arg_dict=spec.arg_dict,
    opt=spec.opt,
    model="xgboost",
    uncertainty_metric="mipt",
    initial_sample_size=24,
    iterative_sample_size=8,
    parallel_tasks=8,
    max_evaluations=160,
    local_search=True,
    local_search_max_iters=20,
    local_search_ordered_radius_fraction=0.075,
    local_search_categorical_radius=1,
    log_scale_increase_rate=1.2,
    seed=0,
)

print("Best value:", result["value"])
print("Best point:", result["point"])

result["point"] is a dictionary in the original argument structure expected by the expensive function, while result["value"] stores the best objective value found.

Notes

• Categorical COCO tasks require COCO's Python bindings at runtime.
• The categorical benchmark family supports the public total dimensions 6, 12, 24, 48, 96, 192.
• Supported surrogate backends are xgboost, catboost, and lightgbm.