Diversity

By default the GeneticEngine evolves one flat population where every individual competes against every other. That works in a majority of problems, but it can fail into what is known as premature convergence - where the population piles into the first decent solution it finds and stops exploring. A novel individual that needs a few generations to really evolve and pay off gets out-competed and lost before it ever matures.

Diversity is radiate's opt-in defense against premature convergence. Instead of one undifferentiated pool, the engine groups genetically-similar individuals into species and lets them compete as groups. A promising-but-immature lineage then competes mainly against its own kind, giving it room to refine before it has to stand against the rest of the population.

How it fits together

Diversity is really two halves of a single loop that runs each generation:

• A distance measure answers "how genetically similar are these two individuals?" given the structure of their genotypes. This effectively turns a pair of genotypes into a single number.
• Species are what the engine builds from those numbers, or distances: it clusters the population, shares fitness within each cluster, and allocates offspring between clusters.

Here we can see the whole cycle before diving into either piece:

flowchart LR
    A[Population] --> B[Measure distance<br/>between individuals]
    B --> C[Group into species<br/>within a threshold]
    C --> D[Compete as groups:<br/>share fitness, allocate offspring]
    D --> A

The distance measure is the input to everything while the species machinery is what consumes it.

Page	What it covers
Distance	The built-in measures (Hamming, Euclidean, Cosine, NEAT), what each one captures, and the value range it produces.
Species	How the engine forms species from those measurements, shares fitness, and allocates offspring + the knobs that tune it.
Example	Wiring it all into a working engine.

Opt-in, and not free

Speciation only runs when you attach a distance measure. It also adds computational cost because every generation the engine must measure distances and re-cluster the population — so reach for it when a problem is converging too early or not reaching its global optimum, not by default.

The whole feature, in one call

Turning diversity on comes down to two settings: the measure (which distance) and the threshold (how close counts as "the same species"). Everything in this sections builds on these two, so here is a brief look at how they work together:

Python Rust

import radiate as rd

engine = (
    rd.Engine(your_codec)
    .fitness(your_fitness_func)
    # A distance measure turns speciation on; the threshold sets how close
    # two individuals must be (per the measure) to share a species.
    .diversity(rd.Dist.euclidean(), species_threshold=0.5)
    # ... other parameters ...
)

let engine = GeneticEngine::builder()
    .codec(your_codec)
    .fitness_fn(your_fitness_fn)
    // A distance measure turns speciation on; the threshold sets how close
    // two individuals must be (per the measure) to share a species.
    .diversity(EuclideanDistance)
    .species_threshold(0.5)
    // ... other parameters ...
    .build();