Use keras-rl for multi agent training.

[Cristián Serpell]

Hi all,

I have asked this in github's issue #53, but as it is closed, I repeat my question here:

Is it possible to use keras-rl for multi agent environments, where each one has its own model to train (each agent has a different model)?

I have been trying to do this using threads, but after trying many different ways to configure agents, environments and launching threads at different points, I always manage to get a TensorFlow exception of the kind:

ValueError: Tensor Tensor("activation_2/Softmax:0", shape=(?, 4), dtype=float32) is not an element of this graph.
 

Something interesting would be to start with an environment that allows different agents, such as go, and then generalize to others more complicated.

Another thing i tried is to create two agents, and then iteratively call fit with nb_steps=1 once per agent, for a number of iterations, expecting them to learn something. But I am sure this is not the easiest and fastest way to train them, and I am not sure if it will really work.

Thanks for your help.

[axel....@gmail.com]

Yes, I would also like to learn 2-player games, where two deep nets play against each other and learn via reinforcement learning. But looking at Keras-rl it does not seem to support this :-(

[Cristián Serpell]

I finally switched to tensorforce. They are keeping the library up to date and it was easy to use.

Cristián