Reproducible evolutionary search

Evolutionary parameter search whose every draw is addressable, so a fitted result replays bit-exactly on any machine and any core count.

evolutionary_search is a genetic optimiser whose randomness is counter-based and keyed, not sequential. Every draw is a pure function of (generation, individual, gene) and the master seed, so the whole search replays bit-exactly — on a different machine, at a different core count, in a different evaluation order.

from algenta import MojoRuntime

runtime = MojoRuntime(mode="local")

lo, hi = [0.0, 0.0], [10.0, 10.0]
pop = runtime.execute("evolutionary_search", "ga_init", [lo, hi, 24, 12345])["result"]

for gen in range(30):
    fitness = [-(pop[i * 2] - 3.0) ** 2 - (pop[i * 2 + 1] - 7.0) ** 2 for i in range(24)]
    pop = runtime.execute(
        "evolutionary_search", "ga_step",
        [pop, fitness, lo, hi, 3, 0.5, 0.3, 0.15, 2, 12345, gen],
    )["result"]

best = runtime.execute("evolutionary_search", "ga_best", [pop, fitness, 2])["result"]

The modules

Module	What it does
evolutionary_search	counter-based, addressable genetic optimiser — every draw is a pure function of its coordinates, so a run replays bit-exactly on any machine and any core count
genetic	the process-seeded genetic optimiser for one-off searches — same population/selection/mutation loop, but a run cannot be replayed

Functions

Function	What it does
ga_init(lo, hi, pop_size, seed)	Uniform initial population inside the bounds. Individual 0 is seeded at the box midpoint so an incumbent-like candidate is always present.
ga_step(pop, fitness, lo, hi, tournament, cx_rate, mut_rate, mut_scale, elite, seed, gen)	One generation: elitism, tournament selection, uniform crossover, Gaussian mutation. Fitness is maximised.
ga_best(pop, fitness, n_dim)	The fittest individual's genes, with its fitness appended.
ga_diversity(pop, n_dim)	Per-dimension population standard deviation — the identifiability diagnostic.
ga_saturation(pop, lo, hi)	Fraction of the population pinned within 1% of each bound: negative at lo, positive at hi.

Why reproducibility is the point

The runtime also ships genetic, which is fine for one-off optimisation. Its mutation draws from a process-seeded generator, so a run cannot be replayed. That is acceptable when you want an answer and unacceptable when the fitted parameters become evidence — a calibration nobody can reproduce is an assertion, not a result.

Because draws are addressed rather than consumed in sequence, nothing depends on evaluation order. Two runs on 8 and 64 cores produce identical populations.

The design is stateless, deliberately

ga_step maps one scored generation to the next and holds no problem state. Fitness is computed by the caller, so the search can drive anything — a simulation, a backtest, a physical experiment — without the optimiser needing to know what it is optimising.

Read the diagnostics, not just the best vector

A fitted value means nothing without knowing whether the objective could resolve it:

• ga_diversity — a dimension whose population spread never collapses is one the objective is indifferent to. Reporting a precise value for it overstates the evidence.
• ga_saturation — a value pinned to a bound is an artifact of the box, not an estimate. It is the difference between "the data prefers 0.6" and "0.6 is as far as we let it go".

Both are worth acting on. Widening a saturated bound is the obvious response and not always the right one: a bound can be doing useful regularisation, and a wider box that scores worse on held-out data is telling you so.

Verification

A one-dimensional search for the peak of −(x−3)² converges to 3.008 as a zero-argument selftest, so the same number is reproducible on any machine. The suite asserts that identical seeds give identical populations and different seeds do not, that genes never leave their bounds under mutation, that elitism never lets best fitness regress, and that ga_saturation reads exactly +1 at the upper bound, −1 at the lower and 0 in the interior.

Next: score what you fit with forecast scoring, or see model selection for held-out splits that keep a search honest.