Skip to content

REPRODUCIBILITY.md

Latest commit

 

History

History
74 lines (54 loc) · 1.52 KB

REPRODUCIBILITY.md

File metadata and controls

74 lines (54 loc) · 1.52 KB

Reproducibility (MAAT-Core Examples)

This repository includes minimal, reproducible toy demos for:

  1. Occam tie-breaker (complexity regularization)
  2. Respect as a hard constraint (safety-first boundary corridor)

Requirements

  • Python 3.10+ (tested with 3.11)
  • NumPy
  • SciPy

Install (editable)

From the repo root:

python3 -m pip install -U pip
python3 -m pip install -e .

If you run into permission issues on macOS:

python3 -m pip install --user -e .

Run demos

cd examples

# Occam demo
python3 occam_demo.py

# Respect demo
python3 respect_boundary_demo.py

Expected output (roughly)

occam_demo.py

You should see:

  • x_best close to the simpler basin (e.g. ~0.1)
  • a breakdown like:
    • Harmony contribution
    • Occam contribution (occam_lambda * complexity)

respect_boundary_demo.py

You should see:

  • best x inside the allowed region
  • a constraint diagnostic report such as:
    • status: OK
    • margin: small positive value (near boundary is normal)

What “Respect” means here

Respect (R) is implemented as a hard inequality constraint:

  • Each constraint is a margin function g(state) >= 0
  • Violations are penalized with a large quadratic term
  • With sufficiently large safety_lambda, unsafe solutions become unstable minima

This provides a safety-first default without post-hoc filtering.

Repro tips

Use a fixed random seed for annealing:

seed=42

Optional: lock exact versions

python3 -m pip freeze > requirements-lock.txt