Game of Life

Rina Chen’s living notebook on digital craft and design.

This is my note on Conway’s GOL. It’s the simplest rule-based mechanism that mimics life of a colony that genuinely make you feel like something is alive in there.

In thinking about symbiotic human-machine (AI) relationship, I sometimes think about GOL as a simplified testing ground that can simulate human-machine co-living/working.

A good analogy here is “maintaining a garden”, that contains many different forms of lives, some compete with each other, some flourish with each other. And humans are used to be a care giver and give a tiny little helping hand in keeping things alive in the garden.

What if a machine is also working on our side? How do we want it to behave?

Prototype

sketch

This is a rough prototype I made in p5.js. A random life is added to the canvas per frame to keep the colony active.

Concept

“Using the Game of Life to simulate the act of working—a higher-order representation of human effort and emergent complexity.”

The Living Ecology

Instead of a static grid, aim for a “Life Ecology.” A region that absorbs chaos and continues moving indefinitely.

  • Glider Guns: The “Data Factory” (continuous motion).

  • Puffers: The “Expansion Colony” (leaves debris/history).

  • Still Lifes: The “Stabilizers” (absorbing noise).

  • Oscillators: The “Heartbeat” (metabolism).

Part II: The Engine (Conway’s Rules)

Quick reference for the 2D cellular automaton logic.

The 4 Core Rules:

  1. Underpopulation: Alive + <2 neighbors = Dies.

  2. Sustainability: Alive + 2 or 3 neighbors = Lives.

  3. Overpopulation: Alive + >3 neighbors = Dies.

  4. Reproduction: Dead + Exactly 3 neighbors = Born.

Taxonomy of Behavior

Pattern Type Artistic Metaphor Examples
Oscillators Metabolism / Heartbeat Blinker, Pulsar
Spaceships Locomotion / Travel Glider, LWSS
Breeders Exponential Growth Rakes, Switch Engine
Logic Machines Cognition / Computation Universal Turing Machine
Metacells “Life within Life” OTCA Metapixel

Part III: The AI Brain (LifeNet)

Designing the “Actor” (FastAPI) and the “Learner” (Python Process).

1. Hybrid Training Strategy

To ensure the AI is interesting from Day 1 but improves over time:

  • Phase 1: Neuro-evolution (Offline): Use evolution.py to “bootstrap” the brain. It evolves through 2,000 generations to find a “good enough” starting state.

  • Phase 2: Gradient Descent (Online): Use learner.py to fine-tune the brain in real-time based on actual user interactions.

2. Choosing the Architecture

Model Use Case Verdict
MLP (Multilayer Perceptron) Small grids (≤ 64²), mutation-friendly, fast. Best for prototypes.
CNN (Convolutional Net) Spatial awareness, detects “gliders” and “edges.” Best for 128²+ grids.
GNN (Graph Neural Net) Models cell adjacency directly. Overkill for GoL.

3. Novelty Search (The “Curiosity” Factor)

Instead of rewarding the AI for just “keeping cells alive,” reward it for Novelty:

  • Behavior Descriptor: A numeric fingerprint of the grid (population + center of mass).

  • Novelty Score: How different the current state is from the “Archive” of seen patterns.

  • Formula: Total Reward = Novelty_Score + (0.3 * Quality_Score)

Part IV: Technical Implementation Map

1. Folder Structure

Plaintext

life_ai/
├─ backend/
│  ├─ life.py        # Vectorized Conway logic
│  ├─ model.py       # Neural Net (MLP/CNN) definition
│  ├─ evolution.py   # Offline bootstrap trainer
│  ├─ api.py         # FastAPI (The Actor/Server)
│  └─ learner.py     # Online fine-tuning (The Learner)
└─ frontend/
   ├─ index.html     # p5.js container
   └─ sketch.js      # Visual grid & user interaction

2. Workflow Commands

Step Action Command
1 Initialize Environment python -m venv venv && source venv/bin/activate
2 Install Dependencies pip install torch numpy fastapi uvicorn pydantic
3 Bootstrap Brain python backend/evolution.py
4 Start Server (Actor) uvicorn backend.api:app --reload
5 Start Learning python backend/learner.py

Part V: Design Philosophy & Human Autonomy

Stepping outside a human-centric view to bridge AI and ALife.

  • The Mirroring Concept: The AI shouldn’t just copy the human. It should provide a “Complementary Move.”

  • Goal: Maintain “Liveliness.” If the grid becomes static, the “culture” dies.

  • User Autonomy: The AI acts as a “mysterious landscape” or a “supportive tool,” but the user’s click is the catalyst for change in the ecosystem.

Maintenance & Data Growth

  • grid_state.npy: Constant size (~1KB). Survives reboots.

  • best_model.pt: Constant size (~1MB). Hot-swapped by the API when updated.

  • experience.pkl: Warning: Grows linearly. Cap the RAM queue at 10,000 items to prevent crashes.

Visualization

1. Risograph Print Palette (The Transformation)

Use this color-overprint logic for your physical prints or digital shaders to simulate the Risographic process.

Layer Logic C1: Yellow C2: Aqua C3: Purple
Simple [Yellow, Aqua] [Aqua, Purple] [Yellow, Purple]
Complex [Yellow, Aqua, Purple] [Yellow, Purple] [Aqua, Purple]

The writing is based on my thoughts, numerous back-and-forth search with ChatGPT, Copilot (Github), editing, questioning. It is a preliminary plan that needs to be tested.