Last active
July 18, 2023 19:18
-
-
Save awjuliani/35d2ab3409fc818011b6519f0f1629df to your computer and use it in GitHub Desktop.
An implementation of a Deep Recurrent Q-Network in Tensorflow.
Thank you very much for the code.
I got a question. It seems that the "deep" part is on the CNN but not on the recurrent part. And I found that the multiple-layer RNN doesn't quite popular such that I hardly found examples on the web. Is it true?
Thanks for these amazing codes!
Why did you define LSTMCell outside the class Qnetwork, rather than inside the Qnetwork?
and why did you split the outputs of recurrent layer? can I use the outputs of recurrent layer directly for the inputs of advantage layer and value layer?
Because I didn't see the connections between targetQN and targetOps in the codes. So I really would like to know how exactly to update the parameters of targetQN by updateTarget(targetOps,sess).
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Thank you, excellent work!
I would like to discuss many points … :)
It was very cool to use DRQN, Double-DQN and Dueling-DQN in one setup! (impressive)
I was thinking it could be very nice if the network can learn how much gradient to use, using some sort of gating mechanism, what do you think? is it doable? how? would it help?
Could you please send the exact reference where the idea of using only half of gradients
Finally, I am a PhD candidate and was thinking of making a contribution, However it looks very hard given all the scientific work described above!
For example, here are some ideas:
1- Do you think using Skip-Connections like ResNet could enable deeper networks and then improve the results? https://arxiv.org/pdf/1512.03385v1.pdf
2- Instead of random initialization, using simple idea such as auto-encoders for game images as pre-training the network, could speed up the training? what about transfer learning?
3- A3C: Can we use the same methods (DRQN) with A3C,and have a super A3C-DRQN-DD-DQN algorithm?
https://arxiv.org/pdf/1602.01783.pdf
4- Planning: Do you think it will be good if we train a model for the environment, using neural network as function approximator, and then use algorithms like Dyna2 to plan for the next move?Generally speaking, can we think of planned-DQN?
What direction, in your opinion, could make a good contribution?
Thank you