After hearing Jeremy talk about RL in part 14, just wanted to share that the folks at Uber have been able to train DNNs that can play Atari simulations on a single regular computer recently.
They have achieved this by training on populations of different NNs, and modifying core tensorflow operations to reduce the overall training time (estimated training time is 4 hrs. according to the post.)
Here is the code https://github.com/uber-common/deep-neuroevolution/tree/master/gpu_implementation and blog article which talks about the same: https://eng.uber.com/accelerated-neuroevolution/
Most of the stuff is still in prototype phase, but let’s hope more research is done in RL in this direction. In their words:
We hope others will use our code to accelerate their own research activities. We also invite the community to build off our code to improve it. For example, further speedups are possible with distributed GPU training and with adding other TensorFlow operations customized for this type of computation.