Tensorforce


#1

Related to: [Name of Request for Research]

Has anyone here been working with Tensorforce ? I am a PyTorch user, which framework I enjoy more than the Tensorflow setting. However, due to successive failures in training an agent in a robotic task , I wanted to verify whether the problem lied in my implementation by using another person’s code. Usually, I found that trying to use other’s code is impractical, maybe because I’m a poor coder or because it is specific. Anyway, I heard of TensorForce and their implementations of PPO, TRPO… But, I have troubles understanding the setting. They do have a documentation page (http://tensorforce.readthedocs.io/en/latest/agents_models.html) but is it not a clear as I would have liked. Also, trying to find a bug in Tensorflow is a real pain (hundreds of lines of error, especially here, given that there’s a whole architecture behind myAgent = PPOAgent() )

Anyway, if anyone has some examples, tuto or whatever, I’d love to hear about ! Thanks !


#2

I’ve found Ikostrikov’s pytorch implementations more than sufficient and working well for most problems. https://github.com/ikostrikov/pytorch-a2c-ppo-acktr

Perhaps you can give this a try? Or check his implementation against yours?


#3

Hey, thanks for answering.
Actually, I find his implementation a bit hard to understand. I wanted something more plug-and-play. Where you specify your observation and action spaces, the type of learning algorithm you wish to use, and off you go. :wink:

I’m still fighting with my own implementation.


#4

I understand. I’m working on a general rl package with friendlier usage as you described in pytorch. I’ll post back in this forum once it’s ready to go.


#5

Awesome. Let’s race :wink: