FluxML
Prioritized Experience Replay DQN
Implemented DQN with PER. Kindly review the code. Looking forward towards the feedback. Ref:
- https://arxiv.org/pdf/1511.05952.pdf
- https://github.com/MorvanZhou/Reinforcement-learning-with-tensorflow/blob/master/contents/5.2_Prioritized_Replay_DQN/RL_brain.py
This looks awesome. Unfortunately on my setup it throws an error:
MethodError: Cannot `convert` an object of type Flux.Tracker.TrackedReal{Float64} to an object of type Float64
Let me know if you can't reproduce this and we'll narrow it down.
Once you get it going this is a pretty cool demo, because you can see the thing learning in real time.
Is it possible to get it working on the GPU?
Is there any way to skip ahead a few iterations without the visualisation? Just to see how the model performs after the full training.
Yes, I'll get it on GPU. Skipping ahead is possible by not rendering the frams, but I am not sure if that'll improve the training speed. I'll try that.