JG1310/repro-chebyshev-policies-mountain-bundle

Reproduction bundle — Chebyshev Policies and the Mountain Car Problem (aNWIVNjocB)

From-scratch reproduction (no Gymnasium / Stable-Baselines3 / RL-Zoo). All environments, policies (Chebyshev + MLP) and trainers (ARS / REINFORCE / PPO) reimplemented in numpy/torch.

Logbook: https://huggingface.co/spaces/JG1310/repro-chebyshev-policies-and-the-mountain-car-problem-reinforcement-learning-for-low

Layout

• core/ — envs (MountainCarContinuous-v0, Pendulum-v1), Chebyshev & MLP policies, ARS/REINFORCE/PPO, eval harness
• scripts/exp00..exp03 — build smoke-test + the three claim experiments
• specs/ — experiment specifications
• DERIVATIONS.md — Claim-1 theorem re-derivations (D1–D7) with executable checks
• results/*.json — full-scale run outputs (numbers quoted in the logbook)
• diag2.py — exploration-scale diagnostic (CH-3-ARS inversion + recovery on Mountain Car)
• gates.py — fidelity gates

Rerun (CPU only)

PYTHONPATH=. python3 scripts/exp01_analytical.py --full   # Claim 1
PYTHONPATH=. python3 scripts/exp02_mc.py --full           # Claim 2 + 3 (MC)
PYTHONPATH=. python3 scripts/exp03_pendulum.py --full     # Claim 3 (Pendulum)
PYTHONPATH=. python3 diag2.py                             # diagnostic

Reproduction outcome

• Claim 1 (analytical solution): VERIFIED — π_ana R̄=99.37, 100/100 goal (paper 99.39)
• Claim 2 (regret 4.18× / 268× params): MIXED — 272× param identity verified; regret factor not reproducible; assigned numbers absent from paper (paper says 6.18×/277×)
• Claim 3 (consistent improvement, PPO/ARS/REINFORCE, multiple envs): SUPPORTED WITH CAVEAT — hyperparameter-sensitive; reproduces with policy-matched exploration