can the train process with CPU backend utilize multi-cores?

[loofahcus]

The neon example code like examples/mnist_mlp.py with CPU backend can take full advantage of multi-cores, but simple_dqn seems to use one core while running. Is it possible for simple_dqn to use all the cpu cores?

[tambetm]

I didn't find the documentation for multi-core, but I doubt it would be trivial to take advantage of this in simple_dqn. simple_dqn doesn't use the high level fit() method, but calls lower level fprop() and bprop() instead (to implement error clipping).

I you are going for CPU training, then it would be interesting to apply algorithms from Asynchronous Methods for Deep Reinforcement Learning.

[loofahcus]

Thanks for your advice~