NEAT Cube Foragers — Neuroevolution Sandbox

Setup (locked once training starts)

Cubes per set

Population

Push every N ticks

Episode length (ticks)

Food per set

Obstacles

Capture radius

Protect output activations

The action decoder assumes the 4 outputs are logistic, i.e. in [0,1]: force is clamped to that range and the face selector bins it. But MOD_ACTIVATION mutates output nodes too by default, and a force output on SINUSOID or BIPOLAR spends much of its range negative — which clamps to exactly zero force, so the cube can never push. Leaving this on keeps output activations logistic while hidden nodes still mutate freely.

Network inputs (locked once training starts)

Direction to nearest foods

how many foods (X)

+ distance to each

Direction to nearest rivals

how many rivals (X)

+ distance to each

Own position

Own velocity

Own orientation (up axis)

Obstacle whiskers (lidar)

rays around cube

Input neurons4

If fewer live foods/rivals remain than X, the nearest one is repeated to fill the extra slots (zeros only when none exist). Whiskers are evenly-spread world-frame rays (range 6) sensing obstacles and walls: 0 = clear, →1 at contact. The input layout is stored in saved genome files, so replays wire up correctly.

Run

Headless (no render)

Seed population from loaded genomes

Training runs the physics as fast as your CPU allows (fixed 60 Hz substeps, so it stays stable). Headless skips all drawing. Training keeps running when this tab is in the background — a worker clock sidesteps the browser's background throttling.

Progress

Modeidle

Generation0

Set–

Best fitness (all-time)–

Best fitness (last gen)–

Mean fitness (last gen)–

Sim speed–

Green: generation best. Amber: generation mean.

NEAT / Species

Species–

Compat. threshold–

Best genome size–

	id	size	best	mean	age

Species are clustered by structural compatibility distance (disjoint connection genes + weight difference); the threshold self-tunes toward a healthy species count. During training, cubes are tinted by their genome's species.

Genomes

Show whisker rays (watch/sandbox)

Loaded genomesnone

One file format for everything: saving from training, from a single inspected cube, or from the ecosystem all produce the same envelope, carrying the sensor spec and settings the brains were evolved under. Loading any file restores those settings into the controls, arms the genomes for both training and the ecosystem, and either replays them at normal speed or — if the file carries a world — puts that world back. Sandbox runs the brains open-ended with no timer: click the floor to drop food, shift+click for an obstacle.

Ecosystem — persistent biological sandbox

No fitness function. Cubes burn energy to live and to push, eat to refill it, starve at zero, die of old age at the lifespan, and bud mutated offspring when they are mature and rich. Food rains back at a fixed rate, which sets the carrying capacity. Natural selection does the rest.

World

World half-size

Obstacles

Population

Founders

Population ceiling

Seed from

Food supply

Food spawns / sec

Max food on field

Energy per food

Metabolism

Energy burn / sec

Energy per full push

Starting energy

Life cycle

Reproduce at energy

Energy given to child

Birth overhead

Maturity age (s)

Lifespan (s)

Heredity

Mutation rate

Mutations per birth

Sexual reproduction

Sensing radius (0 = all)

Carrying capacity (est.)–

Food density–

Run

Run as fast as possible

Follow selected cube

Every file this program writes has the same shape and carries the settings that produced it — sensor spec, run config and ecosystem parameters. Any of them can be loaded to watch, to seed further training, or to seed a new ecosystem. A snapshot is simply a genome file that also carries the world, so loading one puts that world back as it was.

Census

Population–

Elapsed–

Food on field–

Births–

Deaths–

Mean age at death–

Lineage depth–

Mean brain size–

Brain resets (diverged)–

Green: population. Blue: food on the field. Amber: mean lineage depth (generations since the founders). The species table above switches to the living population while the ecosystem runs — its columns become mean energy and mean age.