search

tensorflow / agents

TF-Agents: A reliable, scalable and easy to use TensorFlow library for Contextual Bandits and Reinforcement Learning.
Apache License 2.0
2.8k stars 722 forks source link

SimPLe efficient Agent? #56

Open jarlva opened 5 years ago

jarlva commented 5 years ago

Hi, any chance to port the SimPLe Agent from https://github.com/tensorflow/tensor2tensor/tree/master/tensor2tensor/rl to work with basic GYM (like cartpole)?

Based on the Google blog, SimPLe "is significantly more efficient than current state-of-the-art techniques".

sguada commented 5 years ago

That would require some work, and probably not very relevant for a simple problem like cartpole

jarlva commented 5 years ago

Yes I know that cartpole is a simple one.. I picked cartpols as a simple example that we can learn to apply to more complicated ones. Again, based on Google's report this agent is unlike others. and will reduce training time more than any other agent to-date. Thus I think it's worth looking at. The red line is the figure is the Simple agent....

Untitled

sibyjackgrove commented 5 years ago

Yes, adding SimPLe agent would be really beneficial. I am willing to help out if this is feasible with tf-agents.

jarlva commented 5 years ago

That would be awesome! (if it can be applied to GYM (non-Atari/ALE) environments. Thanks.

sguada commented 5 years ago

It is feasible, but it would take a bit of work so before start working on it, you should create a proposal with detailed steps and milestones. Also we should see how easy would be connecting TF-Agents with Tensor2Tensor, since you need that to train a good model.

adityauser commented 5 years ago

@sguada can you explain what do you mean by "connecting TF-Agents with Tensor2Tensor"? Do you want Tensor2Tensor compatible in TF-Agents or implement SimPLe in TF-Agents from scratch or something else.

jarlva commented 5 years ago

Hello, checking to see if there is any progress?