OpenAI Gym: a toolkit for comparing reinforcement learning algorithms


#1

OpenAI Gym is a toolkit for developing and comparing reinforcement learning algorithms. It supports teaching agents everything from walking to playing games like Pong or Go.

The goal is to have a framework for running and comparing different RL algorithms, as well as sharing their sources for later reproduction.

Website: https://gym.openai.com/
Lauch post: https://openai.com/blog/openai-gym-beta/


#2

This type of initiative will greatly benefit scientific practice in computer science, we can expect to see more and more of them emerging in the coming years. Hopefully they will converge to a unified online platform allowing to:

  • easily create and share standardise problems/datasets that can be used and tested online
  • easily develop, share and compare algorithms online (visualizing the resulting policies, curves for many standardized test, interactive visualization)
  • associate scientific publication to each algorithms, most certainly in the form of a blog post
  • associate a forum to each project for discussion and the ability to dynamically link to different projects
  • fork and extend algorithms of others
  • organize, link, and search projects with their ‘history tree’
  • associate media to each project that could be reused by other for presentation and outreach

I am sure many are working on this, especially initiatives like: http://jupyter.org/, https://cloud.sagemath.com/ and many others.It might be good to make a list of similar initiative on this forum.