Werld: Open-Ended Artificial Life Simulation with Evolving Neural Networks

What Werld Is

Werld is a pure Python artificial life simulation built with standard library only that drops agents into a computational ecosystem with blank neural networks and zero knowledge of human existence. Agents evolve through survival and reproduction rather than backpropagation, with no hardcoded behaviors or reward functions.

Technical Implementation

The simulation runs on 800 nodes arranged as a Watts-Strogatz small-world graph. It starts with 30 agents equipped with NEAT (NeuroEvolution of Augmenting Topologies) neural networks that can evolve their own topology. Each agent has:

• 64 sensory channels covering energy gradients, pheromone trails, nearby agents, seasonal rhythms, internal state, and 19 latent channels that start as unknown
• 7 continuous motor effectors for action
• Up to 16 broadcast channels for communication
• 29 heritable genome traits including communication bandwidth, memory decay, aggression vs cooperation

Brains can grow new neurons, prune connections, and evolve any of 7 activation functions per node. Every cognitive component has a metabolic cost - more neurons, connections, or communication requires more energy.

Evolution Mechanics

Agents have two primary goals: harvest enough energy to stay alive and live long enough to reproduce. When they reproduce, offspring inherit mutated copies of neural traits from both parents through full sexual crossover with NEAT gene alignment. This includes sensory processing, behavioral drives, and all 29 genome traits.

The system allows evolution in any direction. Agents can discover previously unknown sensory channels, develop communication patterns, create motor patterns that become heritable compound actions, and form distinct species based on evolving genome traits.

Observations from Initial Runs

In one 12-hour run, the initial 30 agents grew to over 7,000. The simulation experienced 20+ population crises and famines that wiped out most populations, followed by recovery from handfuls of survivors. Over 18,000 agents died during this period.

Surviving agents evolved more efficient energy consumption, pruned unnecessary neural complexity, developed basic communication, and reproduced frequently. Some lineages evolved out of cortex structures entirely while improving brain capacity. In other cases, populations collapsed to single survivors who repopulated the world with mutated descendants.

Dashboard and Monitoring

Werld includes a Next.js dashboard called "Werld Observatory" that provides live views of:

• Population dynamics
• Brain complexity metrics
• Species trajectories
• Narrative story generator
• Live world map visualization

Project Status

The project is open source under MIT license with 42 stars on GitHub. The creator notes it can consume significant storage when running locally. The codebase includes agents, dashboard, engine, persistence, reasoning, systems, and utils directories.