creature.md

• Cell-specific Genome Neural Cellular Automata

  • Bit Packing
    • Assume rgba32ui (can go smaller)
    • Bits 0-a (Booleans/1 bit) (maybe for program use / optimization)
      • Self (Properties):
        • isAlive
        • markedForBirth
        • markedForDeath
        • hasNeighbor(Up/Down/Left/Right)
        • isReproducing
        • isMobile
        • isHungry
        • isCommunicating

      • Self (Actions):

      • Neighbors (* 8):
        • isAlive
        • hasNeighbor(Up/Down/Left/Right)
        • isMutating
        • isInReproductionProcess
        • isUnderAttack
        • isMobile
        • isHungry
        • isCommunicating
  • Bits a-b (Integers/2-8 bits)
    • Self:
      • energy (0 to 255)
      • timeAlive
      • reproductionCount
      • mutationRate
      • attackPower
      • defensePower
      • reproductionSpeed
      • mobilitySpeed*
    • Neighbors:
      • energy
      • timeAlive
      • reproductionCount
      • mutationRate
      • attackPower
      • defensePower
      • communicationSkill
      • mobilitySpeed
  • Arbitrary Memory?
    • Self memory: (stores historical actions, status, etc.)
    • Incoming memory: (communication from other cells)
    • Outgoing memory: (communication to other cells)

  Possible Actions:

  • Move up, left, right, down
  • Attack up, left, right, down
  • Reproduce up, left, right, down (with mutations)
  • Communicate up, left, right, down
  • Consume energy
  • Place pheremone
  • Die

  Need Some Update Rule Scheme

  • New layer, rgba32f (2x 16f = 8 FP16 total)
    • Each of the 8 FP16 is a weight
  • Mini Neural Network:
    • Input layer: (cell status, neighbors' status, etc.)
    • 2 Hidden layers: (with activation functions)
    • Output layer: (action, softmax)

  Update Pipeline:

  • Set states
    • energy
    • timeAlive
    • ...
  • (Sync)
  • PlanMove
  • (Sync)
  • Perform NN forward pass
  • Execute action (move, attack, reproduce, etc.)
  • Update environment (energy consumption, cell deaths, etc.)