How I created an evolving neural network ecosystem

Summary of technical concepts, features, and analysis

• Project premise (“bibbots”): An evolving artificial ecosystem of “creatures” that:
  • Eat food pellets to gain energy
  • Use energy to survive longer
  • Lay eggs and pass down genes
  • Are not initially programmed for specific behaviors (e.g., “move forward” isn’t hard-coded). Instead, behaviors emerge through evolution.

Evolution via senses/actions mapping

• Each creature receives inputs (senses) and produces outputs (actions).
• At the start, there are no brain connections, so behavior is essentially random/absent.
• Over time:
  • Random mutations can create useful connections (e.g., evolving a “move toward pellet” behavior by chance).
  • Natural selection amplifies improvements.

Genotype vs. phenotype separation

• Genes include traits like:
  • Size, speed, color
  • Mutation factors
• Genes also influence metabolism cost (energy cost per second), derived from size/speed.
• “Memory/brain” behavior is treated separately from genes:
  • The brain begins with empty connections
  • Then evolves its network structure and parameters

Neural network implementation and mutation mechanics

• The brain initially appears art-coded, but then the creator shifts toward using neural networks for behavioral evolution.
• Evolution modifies network topology and weights:
  • Mutations can add synapses
  • Mutations can change synapse strength
  • Mutations can disable links
  • Create hidden neurons and evolve their activation functions
  • Remove synapses/neurons to prevent “floating” structures, keeping behavior consistent
• Forward propagation occurs every time step to produce actions.

Neuroevolution algorithm (NEAT-inspired, continual selection)

• The creator uses NEAT-style evolution, referencing NEAT / RTNEAT ideas.
• Key departure from generation-based genetic algorithms:
  • There is no discrete generation evaluation
  • Evolution resembles continuous natural selection, where fitness is based on:
    • Survival
    • Successful reproduction
• The fitness core is effectively: “how well they survive and pass their genes.”

Emergence and performance constraints

• Even simple evolving network components can produce complex emergent behaviors.
• Large ecosystem sizes can cause frame-rate collapse if too many agents exist.
• The creator regulates population/complexity by controlling food pellet spawn rate to maintain an equilibrium around ~200 bibbots:
  • If too many exist, they reduce pellet spawning (food supply) proportionally
  • They adjust gradually to steer toward equilibrium

Unclarity / interpretability issues (user-facing analysis)

• The creator notes many events are hard to understand visually.
• Viewer suggestions include encoding stats into visuals (e.g., blue color = speed).
• The creator prefers using color to indicate:
  • Genetic drift
  • Species identification
  • Independent of other genes

Pheromones / communication (feature under development)

• Agents can produce and sense pheromones to enable coordination and structures larger than individuals.
• Current state:
  • Implemented algorithmically (e.g., pheromone detected within a distance like a “broadcast tower”)
  • No visible on-screen representation of pheromone effects
• Proposed improvements:
  • Visualize pheromones as clouds
  • Or replace them with leaving trails (or both)

Planned ecosystem complexity: predation

• The creator wants to add carnivores/predators to introduce competition and greater ecosystem interest.
• Anticipated benefit:
  • Helping solve population control via herbivore–carnivore balance
• Proposed implementation:
  • Each creature has a continuous “spectrum” float for herbivorous ↔ carnivorous tendency
  • Omnivory is allowed (eat both pellets and meat)
  • Digestion becomes less efficient for mixed diets
• Also suggested:
  • Reworking the energy system for more stable realism

Energy conservation and food dependency problem (technical limitation)

• The system currently depends on continuously spawning food because energy transfer is described as not conserved.
• If food spawning stops, all agents die.
• The creator plans to fix this using an idea aligned with:
  • “nothing is lost, nothing is created, everything is transformed”
• A follow-up video is planned to explain a diagram related to this change.

Delivery / updates

• The video is presented as a project demo with future updates.
• The creator requests community feedback in the comments.
• Mentions subscribing for updates (frequency not guaranteed).

---

Main speakers / sources

• Main speaker: The video creator (author of the “bibbots” project).
• Referenced source/model: NEAT / RTNEAT (real-time evolution of augmenting topologies).
• Referenced principle: Antoine Lavoisier’s conservation-of-mass/energy quote (“nothing is lost…”), used to motivate a correction to the energy system.