Demos

Vivarium — feature gallery

A visual tour of what the simulation does. See the README for build and run instructions.

The living world: green/teal dots are two plant types, blue circles are herbivores, red triangles are carnivores. Gray discs are terrain. Coloured rings are agents broadcasting a signal.

---

The ecosystem

A toroidal (wrap-around) plane with three trophic tiers — regenerating plants, herbivores that eat them, and carnivores that hunt herbivores. Every agent spends energy each tick (basal cost + movement) and dies at zero; eating restores it. Predator–prey dynamics are kept from collapsing by predator metabolism, a reproduction cooldown/energy cap, and an optional rescue effect (rare immigrants when a tier nears extinction).

The brain

Each agent is driven by a tiny hand-rolled recurrent neural net (19 inputs → 12 hidden → 4 outputs, tanh, no ML libraries). The hidden layer feeds its previous activations back in (Elman-style memory). Inputs are directional vision: a ring of 6 sectors, each carrying the proximity of the nearest target (food/prey), nearest threat (predator), and the voice of the nearest same-kind neighbour, plus the agent's energy. Outputs are turn, speed, eat, and a broadcast signal.

Evolution, learning & curiosity

• Evolution — reproduction passes on the genome (brain weights) and traits (size, speed, sense radius, learning rate, curiosity, diet) with small mutations. Reproduction is sexual by default: a nearby mate yields a crossover of both parents, otherwise an asexual clone.
• In-lifetime learning — a reward-modulated Hebbian rule adapts a working copy of the weights from energy reward during life; the evolved Plasticity trait sets the rate. Learning is Baldwinian (offspring inherit the genome, not what a parent learned).
• Curiosity — each agent trains a small forward model (online gradient descent) predicting its next senses; the prediction error becomes an intrinsic reward, scaled by the evolved Curiosity trait — exploration for its own sake.

Click any agent to open the inspector: energy, age, generation, traits, live vision channels, brain outputs, and recurrent memory.

Communication

Agents broadcast a scalar signal that same-kind neighbours hear through a "voice" vision channel — a substrate for emergent alarm calls / flocking. Active broadcasts show as faint halos (warm = positive, cool = negative).

Environment

Four environmental systems add structure and time-varying pressure (all tunable, 0 disables):

• Seasons — a slow cycle scales plant regrowth (boom/bust).
• Day/night — a faster cycle shrinks effective vision at night (the world dims).
• Terrain — impassable rocks block both movement and line of sight.
• Food types — two plant types; herbivores evolve a Diet trait and so can specialise into niches.

Close-up (zoomed in): terrain rocks, both food colours, signalling halos, and the dimmed night-time tint.

Analytics views

Built-in, hand-rolled machine-learning views of the population (no ML libraries):

Species — g

k-means clusters the population in brain-genome space (so it discovers lineages that share a genome), recolours agents, and shows a 2D PCA map.

Lineage — l

A stacked abundance-over-time chart of the most prominent lineages, with agents coloured by lineage — watch lineages rise, sweep, and go extinct.

Phylogeny — p

The coalescent genealogy tree of the living population: leaves traced back through shared ancestors, time on the x-axis, branches coloured by lineage.