This repository contains the frozen PyTorch model weights for a custom architecture that shattered the theoretical physics ceiling of MountainCarContinuous-v0.
Methodology
Instead of Reward Shaping or augmented logic, this agent utilizes an unmodified PPO core wrapped in an FDF governor. The governor injects a massive curiosity penalty (Beta 5.0) based purely on neural topological stagnation during the Play Phase, seamlessly decaying to 0.0 to allow raw physics compilation during the Mastery Phase.
Governor Beta: 5.0
Schedule: 50% Play / 30% Decay
Hardware Bias: Apple Silicon (MPS)
Holdout Average (100-Episodes): 97.10
How to Evaluate
Download the fdf_seed_24_golden_mountaincar.pt file.
Download the fdf_inference.py script to instantiate the custom PyTorch ContinuousPPOMLP.
