How does this differ from novelty-seeking (RND)?

Novelty-seeking often gets stuck on random noise (the 'noisy TV' problem), whereas curiosity-driven learning specifically seeks learnable patterns that offer compression progress.

What is 'Compression Progress'?

It is the mathematical difference between the description length of a data history before and after a model update, signaling that the agent has learned something meaningful.

What technologies does this skill support?

This skill provides implementation examples and logic patterns for Python, JAX, and PyTorch frameworks.

What is curiosity-driven learning in AI?

It is an approach where agents are rewarded for finding data that improves their ability to compress and understand their environment, rather than just seeking external rewards.

Curiosity-Driven Learning

Name: Curiosity-Driven Learning
Author: plurigrid

byplurigrid

•

Ciencia de Datos y ML

Implements intrinsic motivation for AI agents using Schmidhuber's formal theory of creativity and compression progress.

This skill provides a framework for implementing curiosity-driven learning within AI agents, enabling them to seek data that improves their internal world models. Based on Jürgen Schmidhuber's 1991-2010 theories, it shifts focus from extrinsic rewards to 'Compression Progress'—the measure of how much better an agent can compress information after observing it. This approach allows agents to autonomously discover learnable patterns while ignoring both redundant data and unlearnable noise, making it ideal for creating more creative, exploratory, and efficient machine learning models.

Características Principales

017 GitHub stars

02World model update patterns for Python, JAX, and PyTorch

03Intrinsic reward computation based on compression progress

04Logic for distinguishing learnable patterns from random noise

05Integration with GF(3) triads for cognitive superposition

06Meta-learning implementations for predicting state learnability

Casos de Uso

01Implementing advanced exploratory behaviors in robotic or simulated agents

02Developing autonomous RL agents for environments with sparse or delayed rewards

03Building 'creative' AI systems that prioritize the generation of novel, structured data

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add plurigrid/asi curiosity-driven

For use in Claude.ai and ChatGPT

Download Skill