pisinger-knapsack

A deterministic generator of 0/1 knapsack problem instances based on Pisinger's scheme — three value–weight correlation classes, an explicit seed, canonical JSON output and a dataset manifest with checksums. Built so that an experiment's input data is fully reproducible and verifiable.

Why

Benchmark results are only trustworthy if the inputs can be regenerated exactly. This generator is designed around that requirement:

• Deterministic. Generation is driven by an injected numpy.random.Generator and never touches global RNG state — the same seed always yields the same instance, regardless of call order.
• Canonical & portable. Instances serialize to byte-stable JSON (sorted keys, no platform newline translation), so a file's SHA-256 checksum is identical on every operating system.
• Verifiable. A manifest indexes a dataset and lets you confirm later that no instance file has changed.

Installation

pip install pisinger-knapsack

Development version (from a checkout):

pip install -e ".[dev]"

Requires Python 3.10+ and NumPy 2.x.

Quick start (Python)

import numpy as np
from pisinger_knapsack import (
    CorrelationType,
    generate_instance,
    save_instance,
    load_instance,
)

# Inject an explicit generator — this is what makes results reproducible.
rng = np.random.default_rng(20260101)

instance = generate_instance(rng, n=20, correlation_type=CorrelationType.WEAKLY_CORRELATED)
print(instance.n, instance.R, instance.capacity)   # 20 1000 <W>
print(instance.values, instance.weights)           # aligned int64 arrays

save_instance(instance, "instance.json", metadata={"seed": 20260101})
assert load_instance("instance.json") == instance

generate_instance(rng, n, correlation_type, R=1000, offset=None) returns a frozen KnapsackInstance(n, R, correlation_type, values, weights, capacity). The capacity follows the "50% knapsack" rule, W = floor(0.5 · Σ weights).

Correlation classes

Given a weight w ~ U(1, R) and an offset d (default R // 10):

Weights always lie in [1, R]; values are always ≥ 1 and may exceed R for the weakly and strongly correlated classes (by design). Correlation strength increases across the classes, which is what drives the difficulty spread.

Command-line interface

generate

pisinger-knapsack generate --n 20 --correlation weakly --seed 1 --out instance.json

manifest

pisinger-knapsack manifest build  --dir ./data --out ./data/manifest.json
pisinger-knapsack manifest verify --dir ./data --manifest ./data/manifest.json

manifest verify exits with a non-zero status if any instance file no longer matches the checksum recorded in the manifest.

Output format

Each instance is a JSON object with sorted keys (arrays shown compact here; on disk each element is on its own line):

{
    "R": 1000,
    "capacity": 812,
    "correlation_type": "weakly",
    "metadata": { "seed": 1 },
    "n": 3,
    "schema_version": 1,
    "values": [415, 720, 88],
    "weights": [402, 731, 95]
}

metadata is optional, free-form provenance — the library copies it verbatim and assigns no meaning to its contents. The manifest indexes a directory of instances; the instance_id is the file stem, so the caller controls naming:

{
    "schema_version": 1,
    "instances": [
        {
            "R": 1000,
            "capacity": 812,
            "checksum": "sha256:…",
            "correlation_type": "weakly",
            "instance_id": "n20_weakly",
            "metadata": { "seed": 1 },
            "n": 20
        }
    ]
}

Examples

The examples/data directory ships a small, deterministic dataset (two sizes × three correlations) and its manifest. Regenerate it byte-for-byte at any time:

bash examples/generate.sh

A test (tests/test_examples.py) regenerates these files and fails if they drift from what the current generator produces.

Public API

Re-exported from the top-level package:

• CorrelationType, KnapsackInstance
• generate_instance, capacity
• to_dict, from_dict, save_instance, load_instance, read_metadata
• build_manifest, write_manifest, read_manifest, verify_manifest, file_checksum, instance_checksum

Development

python -m venv .venv
source .venv/bin/activate          # Windows: .venv\Scripts\activate
pip install -e ".[dev]"

ruff check .          # lint
ruff format .         # formatting
mypy                  # type checking (strict)
pytest                # tests

Optional pre-commit hooks:

pip install pre-commit
pre-commit install

References

The correlation classes and the instance design follow David Pisinger's work on hard knapsack instances:

• D. Pisinger. Where are the hard knapsack problems? Computers & Operations Research, 32(9):2271–2284, 2005. doi:10.1016/j.cor.2004.03.002
• D. Pisinger. Optimization codes (knapsack instance generators). http://hjemmesider.diku.dk/~pisinger/codes.html

License

MIT © Damian Wiśniewski