Try ray/RLlib

[lefnire]

[Update 2018-07-27] Update: seems Coach has slowed down (w/o much community), and rllab has stopped. A more recently popular framework is rllib (one letter different than rllab).

---

I'd like to try replacing TensorForce with Coach and see if we get any better performance. Use the Clipped PPO for an apples-to-apples comparison; instructions on converting our btc_env.py here.

History: before landing on TensorForce I'd tried rll/rllab and openai/baselines. Baselines is backed by OpenAI, the company behind half these algorithms (they're behind PPO, the model we're using). But baselines isn't a plug-n-play framework intended for developer use; instead it's a dumping-ground for each paper's corresponding sample code. I couldn't get any of their stuff customized to our use-case; all runs resulted eventually in NaNs everywhere. Coach is new. No hunch as to whether it'll outperform, though it is backed by Intel which bodes well. I want to give it a whirl, but don't have time right now - so if anyone wants to take a stab at it, please indeed!

[datctbk]

hi Sir! can I help you that?

[lefnire]

Sorry late, I'm AFK from the project till July. Would be great if you could take a look! I've seen a lot of activity on both projects, just from Git commits, so I think it'd be good to have a comparison between the two.