This repository provides a TensorFlow implementation for the paper "Large-Scale Study of Curiosity-Driven Learning." It enables researchers and practitioners to explore curiosity-driven reinforcement learning agents without extrinsic rewards across 54 benchmark environments, investigating the impact of feature spaces on prediction error for intrinsic reward calculation.

How It Works

The project implements curiosity as an intrinsic reward signal derived from prediction error. It explores the efficacy of using random features versus learned features for computing this prediction error, demonstrating that while random features suffice for many benchmarks, learned features offer better generalization, particularly to novel game levels.

Quick Start & Requirements

• Primary install / run command: bash python run.py
• Prerequisites: TensorFlow, MPI (for multi-GPU/multi-machine training).
• Links: Project Website, Demo Video

Highlighted Details

• Large-scale study across 54 environments.
• Investigates random vs. learned features for prediction error.
• Demonstrates generalization benefits of learned features.

Maintenance & Community

Status: Archived (code provided as-is, no updates expected).

Licensing & Compatibility

The repository does not explicitly state a license.

Limitations & Caveats

The project is archived and will not receive further updates. The lack of an explicit license may pose compatibility issues for commercial or closed-source use.