asystemoffields/disco-torch

A PyTorch port of DeepMind's Disco103 — the meta-learned reinforcement learning update rule from Discovering State-of-the-art Reinforcement Learning Algorithms (Nature, 2025).

35
/ 100
Emerging

This project offers a pre-trained neural network, Disco103, that replaces traditional reinforcement learning loss functions. You feed it your agent's experiences, and it outputs specific loss targets, allowing your agent to learn by minimizing the difference. This tool is for AI researchers and practitioners who develop and train autonomous agents and want to simplify or improve their reinforcement learning pipelines.

Available on PyPI.

Use this if you are training a reinforcement learning agent and want to use a state-of-the-art, meta-learned update rule instead of manually designing complex loss functions like PPO or GRPO.

Not ideal if you need full transparency into the exact mathematical formulation of your loss function or if your project specifically requires fine-grained control over every aspect of the learning update rule.

reinforcement-learning AI-agent-training meta-learning machine-learning-research autonomous-systems
Maintenance 10 / 25
Adoption 5 / 25
Maturity 20 / 25
Community 0 / 25

How are scores calculated?

Stars

9

Forks

Language

Python

License

Apache-2.0

Last pushed

Mar 08, 2026

Commits (30d)

0

Dependencies

2

Get this data via API

curl "https://pt-edge.onrender.com/api/v1/quality/ml-frameworks/asystemoffields/disco-torch"

Open to everyone — 100 requests/day, no key needed. Get a free key for 1,000/day.