README.md

metadata

viewer: false
pretty_name: One Solve, Every Step
tags:
  - reinforcement-learning
  - actor-critic
  - reproducibility
  - checkpoints

One Solve, Every Step

Checkpoints for Accelerating Actor–Critic Learning

Code  ·  Usage

Saved checkpoints, simulator inputs and raw outcomes for the four benchmarks in the report. Training and evaluation code lives in the GitHub repository.

Benchmark	Setting	Runs	Size
KeyCorridor	S3R2 · S3R3 · S4R3 · S5R3	640	6.56 GB
MiniGridLoCA	Reward change from AWAC, SAC and QAC starts	480	8.05 GB
DeepSea	N = 10 · 20 · 30 · 40	520	3.51 GB
SecureCourier	Gate skill and route choice	50 pairs	10.63 GB

Usage

pip install -U huggingface_hub zstandard
huggingface-cli download denghanning/COMP8800_Research_Project --repo-type dataset --local-dir data
cd data
python archive.py verify  --release DeepSea
python archive.py restore --release DeepSea --output DeepSea

Files are packed into 256 shards (shards/) indexed by manifest.sqlite; SHARDS.json lists their SHA-256 checksums. restore checks every file against its checksum.