PropFlow Documentation

PropFlow is a Python toolkit for building and experimenting with belief propagation and other distributed constraint optimization (DCOP) algorithms on factor graphs.

PyPI version Python versions License

Quick Start

Install PropFlow from PyPI:

pip install propflow

Basic Example:

from propflow import BPEngine, FGBuilder
from propflow.configs import create_random_int_table

# Create a simple factor graph
fg = FGBuilder.build_cycle_graph(
    num_vars=5,
    domain_size=2,
    ct_factory=create_random_int_table,
    ct_params={'low': 1, 'high': 10}
)

# Run belief propagation
engine = BPEngine(fg)
engine.run(max_iter=20)

print(f"Solution: {engine.assignments}")

Learning Path

PropFlow’s documentation follows the same top-down pipeline implemented in the runtime:

  1. Quick Start Guide gives you runnable snippets using the recommended flow (FGBuilder → engine → simulator).

  2. PropFlow User Guide dives into each layer in detail—from agents up through snapshot tooling—highlighting when to rely on helpers versus manual wiring.

  3. Snapshot How-To (Updated) provides a comprehensive guide to capturing, analyzing, and visualizing algorithm state at each iteration—essential for debugging and validating convergence.

  4. PropFlow Handbook covers operational practices, deployments, and deeper troubleshooting once you’re running larger studies.

Contents:

Key Features

  • Belief propagation variants: Min-Sum, Max-Sum, Sum-Product, and Max-Product computators

  • Policy engines: damping, Q/R damping, splitting, mid-run splitting, cost reduction, diffusion, pruning, and TRW

  • Dynamic graph construction: cycle, random, lemniscate, unary-augmented, and manual factor graphs

  • Simulation and analysis: multiprocessing-backed comparisons through Simulator

  • Snapshot tooling: automatic EngineSnapshot capture plus analyzer and visualizer helpers

Indices and tables