This repository contains the code for a tv maze environment that is used to show the shortcomings of standard curiosity-driven learning.

Download the built environments here:

https://drive.google.com/drive/folders/1fFzDrs78teoBXtuUHJT6bvUN6MSoL_Dc?usp=sharing

To test after cloning and downloading:

python3 tv_test.py

To edit the environment, download Unity (https://unity.com/) and open this repo as a new project.

This environment has appeared in (these papers will be updated to cite this work):

https://pathak22.github.io/large-scale-curiosity/

https://pathak22.github.io/exploration-by-disagreement/

As well as in blog posts here:

https://ai.googleblog.com/2018/10/curiosity-and-procrastination-in.html

https://blog.openai.com/reinforcement-learning-with-prediction-based-rewards

The unity interface is from a very early version of this:

https://github.com/Unity-Technologies/ml-agents