ENOMAD

ENOMAD is a flexible evolutionary optimizer that couples global search (crossover + mutation) with local, derivative‑free refinement via the NOMAD solver, exposed through PyNomad.

Why ENOMAD?

• Derivative‑free continuous control – no action discretisation, no back‑prop through time; handles dense recurrence gracefully.

• Prior‑aware search – starts from a biological connectome (or any strong prior).

  • EA mode: many tiny edits → minimal drift.
  • rEA mode: < 50 edits → structural fidelity.

• Embarrassingly parallel – flip on Ray to scale linearly with CPU cores.

Installation

pip install ENOMAD

# Development version (clone + editable install)
git clone https://github.com/dsb-lab/ENOMAD
cd ENOMAD
pip install -e .[ray]  # optional: Ray for parallel NOMAD calls

Requirements: Python >= 3.9 · NumPy >= 1.23 · tqdm · PyNomad >= 0.9 · (optional: ray)

Quick start

Try ENOMAD instantly in your browser:

import numpy as np
from ENOMAD import ENOMAD

obj = lambda x: -np.sum(x**2)  # maximise => global optimum at x = 0

opt = ENOMAD(
    "EA",                          # or "hybrid"
    population_size=32,
    dimension=10,
    objective_fn=obj,
    subset_size=5,
    bounds=0.2,
    max_bb_eval=100,
    n_mutate_coords=2,
)

best_x, best_fit = opt.run(generations=100)
print(f"Best fitness: {best_fit:.4f}")

API

ENOMAD(
    optimizer_type: Literal["EA", "rEA"],
    population_size: int,
    dimension: int,
    objective_fn: Callable[[np.ndarray], float],
    subset_size: int = 20,
    bounds: float = 0.1,
    max_bb_eval: int = 200,
    n_elites: int | None = None,
    n_mutate_coords: int = 5,
    crossover_rate: float = 0.5,
    crossover_type: Literal["uniform", "fitness"] = "uniform",
    crossover_exponent: float = 1.0,
    init_pop: np.ndarray | None = None,
    init_vec: np.ndarray | None = None,
    low: float = -1.0,
    high: float = 1.0,
    use_ray: bool | None = None,
    seed: int | None = None,
)

Methods

ENOMAD offers two training strategies that differ only in when and how NOMAD is invoked within the evolutionary loop.

EA mode

Every generation, each individual in the population is passed to NOMAD for local refinement:

1. Slice selection – Pick subset_size coordinates at random (≤ 49, per NOMAD’s convergence guarantees).
2. Local search – Run PyNomad with a ±bounds hyper‑rectangle around that slice and a budget of max_bb_eval evaluations.
3. Replacement – If the refined individual improves its fitness, it replaces the original.
4. Reproduction – Select the top n_elites by fitness, then fill the rest of the population via fitness‑proportional crossover (probability crossover_rate) followed by random‑reset mutation (n_mutate_coords coordinates).

EA mode tends to make many small synaptic adjustments, keeping the overall L2 distance to the original connectome low while steadily improving reward.

rEA mode

An evolutionary mutation proposes a sparse change‑set first; NOMAD then fine‑tunes only those altered weights:

1. Mutation – Each offspring mutates a random subset of weights (usually < 50).
2. Targeted NOMAD – If the diff mask is novel and < 50 coords, run PyNomad only on that mask.
3. Evaluation & elitism – Update fitness, retain best individuals, proceed with crossover/mutation.

rEA mode yields comparable rewards to EA mode while changing far fewer synapses – ideal when biological plausibility demands minimal rewiring.

Key hyper‑parameters (shared):

Testing

pip install -e .[dev]  # includes pytest, ruff, black, etc.
pytest -q              # run smoke + reproducibility tests

Contributing

1. Fork + create a feature branch
2. Run pre-commit install
3. Add unit tests for new behavior
4. PR + short summary of the change

License

MIT License — see LICENSE file

Acknowledgements

• Hi my name is miles, I hope you enjoy these algorithms and optimize some cool stuff using them <3
• PyNomad
• NOMAD team at Polytechnique Montréal / GERADmiddlemouse