Evolutionary Particle Life

An evolutionary particle life simulation using JAX for high-performance computation. This system implements evolving particles with local interaction rules, species diversity through mutation and selection, and comprehensive experimental tracking.

Features

Core Simulation

• JAX-accelerated physics: Fast particle dynamics using JAX and jax-md
• Periodic boundary conditions: Toroidal space for continuous interaction
• Neighbor lists: Efficient spatial queries using jax-md neighbor lists
• Species evolution: Particles copy successful neighbors and mutate over time

Circle-based Mutation System

• Spatial mutations: Random circular regions selected for mutation (not point mutations)
• Three transformation types:
  • Noise addition: Random gaussian noise to species vectors
  • Rotation: Species vectors rotated in 2D space
  • Scaling: Species vectors scaled by random factors
• Group-level evolution: Mutations affect spatially coherent groups

Optimized JAX While Loop Stepping

• Adaptive stepping: Runs steps until neighbor list overflow
• Maximized efficiency: Reduces overhead by maximizing steps between rebuilds
• Configurable limits: Balance efficiency and memory usage

Comprehensive Metrics System

Physics Metrics

• Momentum, mean velocity, velocity variance
• Spatial extent

Species Diversity Metrics

• Number of unique species
• Shannon entropy
• Species variance and range
• Pairwise diversity

Spatial Structure Metrics

• Activity level
• Clustering coefficient
• Nearest neighbor distances

Visual Complexity (Flow-Lenia inspired)

• Compression complexity: PNG compression ratio as proxy for Kolmogorov complexity
• Spatial frequency: FFT-based high-frequency content measure

Based on: Flow-Lenia.png: Evolving Multi-Scale Complexity by Means of Compression

VLM Evaluation

• Life-likeness scoring: Uses Qwen3-VL-8B to evaluate "life-like" qualities
• Local inference: vLLM for fast, free evaluation
• Temporal analysis: Tracks growth, movement, interaction patterns

Inspired by: ASAL: Agent-Supervised Artificial Life

Experiment Tracking

• Wandb integration: Track experiments, metrics, videos
• Parameter sweeps: Automated grid search over evolutionary parameters
• Hydra configuration: YAML-based config management

Installation

# Create virtual environment with uv
uv venv
source .venv/bin/activate

# Install dependencies
uv pip install -r requirements.txt

# For CUDA support (GPU acceleration)
uv pip install --upgrade "jax[cuda12]"

Quick Start

Test Your Setup

python test_setup.py

Run Single Experiment

# Default config
python run_experiment.py

# Small simulation for testing
python run_experiment.py simulation=small

# Disable wandb
python run_experiment.py wandb.mode=disabled

# Override parameters
python run_experiment.py \
    simulation.num_particles=2000 \
    mutation.mutation_prob=0.02

Run Parameter Sweep

# Evolution-focused sweep
python run_sweep.py experiment=evolution_sweep

# Smaller focused sweep
python run_sweep.py experiment=evo_focused

Analyze Sweep Results

python analyze_sweep.py

Configuration

All experiments use Hydra YAML configs in conf/:

conf/
├── config.yaml              # Main config
├── simulation/
│   ├── default.yaml        # Default parameters
│   ├── small.yaml          # Small test
│   └── large.yaml          # Large simulation
├── mutation/
│   └── default.yaml        # Mutation parameters
├── metrics/
│   └── default.yaml        # Metrics config
└── experiment/
    ├── single.yaml         # Single run
    ├── evolution_sweep.yaml # Full sweep
    └── evo_focused.yaml    # Smaller sweep

Override from command line:

# Override parameters
python run_experiment.py simulation.mass=0.05 mutation.copy_prob=0.01

# Use different configs
python run_experiment.py simulation=large experiment=evolution_sweep

# See full config
python run_experiment.py --cfg job

Key Parameters

Evolutionary Parameters (main focus)

• copy_dist: Maximum distance for copying successful neighbors
• copy_prob: Probability of copying per frame
• mutation_prob: Probability of mutation per frame
• species_dim: Dimensionality of species vectors

Physics Parameters

• mass: Particle mass
• rmax: Maximum interaction radius
• repulsion_dist: Repulsion distance
• repulsion: Repulsion strength

Mutation Parameters

• min_radius: Minimum mutation circle radius
• max_radius: Maximum mutation circle radius

Python API

import jax.numpy as jnp
from evo_particle_life import ParticleLife
from metrics import MetricsTracker

# Create simulation
sim = ParticleLife(
    num_particles=4000,
    species_dims=2,
    size=jnp.array([3.0, 3.0]),
)

# Run with adaptive stepping
positions, step_count = sim.step_while(max_steps=200)

# Track metrics
tracker = MetricsTracker(sim.displacement_fn)
metrics = tracker.compute_all_metrics(sim, step=0)

Research Questions

The metrics system enables investigation of:

1. Species diversification dynamics: How does diversity evolve over time?
2. Complexity emergence: When do complex visual patterns appear?
3. VLM correlation: Do compression-complex systems score higher on life-likeness?
4. Parameter effects: How do copy_prob, mutation_prob affect diversity/complexity?
5. Phase transitions: Identify sudden changes in system behavior

Architecture

• evo_particle_life.py: Core simulation with JAX physics and mutation
• metrics.py: Comprehensive metrics (species, complexity, spatial)
• vlm_evaluator.py: VLM-based life-likeness evaluation
• run_experiment.py: Single experiment runner
• run_sweep.py: Parameter sweep runner
• analyze_sweep.py: Results analysis
• render.py: Fast JAX-based visualization

Performance

• JIT compilation: All core functions JIT compiled
• Neighbor lists: O(N) force computation
• While loop optimization: Minimizes Python overhead
• GPU compatible: Runs on GPU with CUDA JAX

References

Inspired by:

• Flow-Lenia: Massively Parallel Continuous Cellular Automata
• Flow-Lenia.png: Evolving Multi-Scale Complexity by Means of Compression
• ASAL: Agent-Supervised Artificial Life

License

MIT