cmenpy

Compute Metaheuristic Engine for Python (cmenpy or /siːmɛnpaɪ/) is an extensible, flexible, and vectorized Python framework for designing and prototyping metaheuristic and population-based optimization algorithms.

It provides the clean abstractions necessary to build custom evolutionary algorithms, swarm intelligence methods, and hybrid heuristics from scratch — with absolute architectural control.

Installation

pip install cmenpy

Requires Python >=3.11. Numba is optional: kernels compile when it is installed (pip install "cmenpy[numba]"), and run as pure NumPy/Python when it is not — or whenever a model asks with numba=False.

To develop or build the docs:

git clone https://github.com/ltsim/cmenpy.git
cd cmenpy
uv sync --extra dev        # tests, type checks, linters
uv sync --extra docs       # ... plus the MkDocs toolchain
uv sync --extra benchmarks # ... plus opfunu-core / enoppy-core test functions

Quick example

import numpy as np
import cmenpy as cm


@cm.declare(x=cm.Argument[int, (0, 3), 2])
def my_algorithm(args, epoch, ctx):
    pop = cm.PopulationSwarm(ctx.buff)
    pop.compute << ctx.rng.uniform(
        ctx.bounds.low, ctx.bounds.up, (pop.size, ctx.bounds.ndim)
    )
    for _ in epoch:  # one epoch of your operator, per iteration
        ...


best_fitness, best_x = my_algorithm(x=1).solve(
    lambda x: np.sum(x**2), cm.Bounds([(-1, 1)] * 3), epochs=150, pop_size=15
)

See the Quickstart for both the function (@cm.declare) and class (@cm.template) styles.

Core pillars

• Vectorized data structures. Population arrays, agent states, and search spaces use a flattened, vectorized layout with no per-individual objects. Buffer views (pop.X, pop.F) interoperate with NumPy through the explicit cm.cast(...) helper and can be passed directly to cmenpy kernels, which unwrap them internally.
• Extensible architecture. Custom operators, mutation rules, or entirely new paradigms require overriding only a few decoupled, well-defined methods. Reusable Operator Modules (@cm.module) share kernels, helpers, and cm.Constant fields across models.
• Non-monolithic flexibility. No visualization tools, file exporters, or rigid loggers — just the mathematical engine and its structural flow.
• Research ready. Licensed under the GNU GPLv3. Design, modify, and distribute your algorithms freely; derivative works stay open, keeping the software a public good for research and commercial development.

Documentation

Full guides and the API reference are published at ltsim.github.io/cmenpy.

To preview locally: uv sync --extra docs && .venv/bin/mkdocs serve.

Contributing

Contributions are welcome — see CONTRIBUTING.md and the list of contributors.

Developed by LTSIM @ 2026. Licensed under GNU GPLv3.