OpenAI Gym | Hacker News Reader

OpenAI Gym

openai.com

76 pointsvarunagrawal13 comments

Open articleView on HN

Previously on HackerNews: https://news.ycombinator.com/item?id=11582345

Two repos to get started building reinforcement learning agents with Gym and Keras:

https://github.com/sherjilozair/dqn

https://github.com/osh/kerlym

This is awesome. All that's missing is a self driving car environment ;)

The focus on improving reinforcement learning is really exciting. It's the area of machine learning that seems to have the most untapped potential.

For those more familiar in this scape - would something like this be possible for NLP? Specifically sentence classification (ie, to produce an intent with metadata)?

OpenAI Gym is for Reinforcement Learning - a different kind of learning, where you don't have ground truth, but the agent gets a positive reward when it makes good guesses. Is this what you are looking for with NLP?

What i'm looking for from NLP is the classic holy grail i guess _(Ie, i know i'm asking a lot)_. To be able to translate english to and from actionable data.

A lot of companies like wit.ai are harnessing the power of machine learning, and having customers/users contribute heavily to this.

I'd love to see a service like OpenAI provide open source offerings to help the community grow language understanding together. Think of wit.ai, but fully open source and with the machine learned result downloadable - usable offline, licensed freely, etc.

Machine Learning for NLP has so many powerful private offerings, but i'm not aware of a single open source one. Especially one that you could run offline on your own servers.

kudos for citing Searn, it's one of the greatest joint prediction and learning algorithms that started a whole range of useful learning to search approaches.

LOLS is even better. all approaches inspired by the reinforcement learning but use the information from already present data to imitate the constructed policy.

training neural networks with the approach of LOLS could make so many tasks reachable.

one could probably even imitate a komodo 9 algorithm with it.

Would be great to have Unity3D or Unreal Engine integration for games/simulations

Yes, that'd be great. Gym is open-source — would love if someone built an integration!

Are there any existing projects based on this, just looking for some examples

Does anyone else think the title/url is misleading ?

Looking at 'Open AI', I expected a generic AI related open source project.

This project seems to be specifically related to "developing and comparing reinforcement learning (RL) algorithms" .... which looks like a more specific subset.

It is not misleading at all, that is the project.

Reply on news.ycombinator.com