LTU-AI/hdppo-FrozenLake-v1

HD-PPO Agent playing FrozenLake-v1

This is a trained HD-PPO (Hyperdimensional Proximal Policy Optimization) agent playing FrozenLake-v1 using gradient-adaptive Fractional Power Encoding (FPE) with a prune-and-fine-tune pipeline.

Published by LTU-AI.

Pipeline

1. Train a teacher at D=64 with gradient-adaptive single-beta FPE.
2. Prune by actor-weight importance through D=64 → 16 → 4.
3. Fine-tune each pruned checkpoint with PPO.

Published checkpoint: seed 123, compact D=4 model (held-out eval mean reward 0.67 ± 0.47).

Usage

Install dependencies:

pip install -r requirements.txt

Evaluate the local checkpoint:

python enjoy.py --weights hdppo-FrozenLake-v1/weights.npz --episodes 10

Render episodes:

python enjoy.py --weights hdppo-FrozenLake-v1/weights.npz --render --episodes 3

Record a replay video:

python record_video.py --weights hdppo-FrozenLake-v1/weights.npz --output replay.mp4

Load from Hugging Face Hub:

python enjoy.py --weights LTU-AI/hdppo-FrozenLake-v1 --episodes 10

Training pipeline

Reproduce the teacher → prune → fine-tune workflow:

python run_prune_finetune_5seed.py

Hyperparameters

{
    "env": "FrozenLake-v1",
    "algo": "HD-PPO (gradient-adaptive FPE, discrete)",
    "teacher_D": 64,
    "pruned_D": 4,
    "timesteps_per_stage": 1000000,
    "seed": 123,
    "is_slippery": true
}

Environment Arguments

{
    "render_mode": "rgb_array",
    "is_slippery": true
}

Model files

File	Description
hdppo-FrozenLake-v1/weights.npz	Published actor (+ critic if HD) and FPE encoder (D=4)
hdppo-FrozenLake-v1/weights_D64_teacher.npz	Teacher checkpoint (D=64)
replay.mp4	Sample rollout video from the published min-D checkpoint
results.json	Evaluation summary for the published checkpoint
results_D64_teacher.json	Evaluation summary for the teacher
config.yml	Training hyperparameters
train_hdppo.py / training modules	Self-contained training code

Citation

If you use this model, please cite the HD-PPO / Hybrid-HD-PPO work.