-
Notifications
You must be signed in to change notification settings - Fork 157
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Double IQN #69
Double IQN #69
Conversation
/test |
Successfully created a job for commit 453a4f8: |
/test |
|
/test |
Successfully created a job for commit a98d6f5: |
@cross32768 Sorry for the delayed response! Do you have time/resource for running Atari experiments like chainer/chainerrl#503 (comment) to verify its relative performance to IQN? |
I have some computing resource and conducted some experiments to verify correctness of implementation. But to conduct experiment in all Atari environments is difficult for my computing power. Is it okay to verify performance in some parts of Atari environments? |
Yes, it is fine to run only the subset. It would be perfect if you run the same set of chainer/chainerrl#503 (comment) so that we can compare scores. |
I'll add results of experiments on same environments as in chainer/chainerrl#503 (comment) by editing this comment. All parameter for experiments is default parameter of training code, and reported scores are "mean" of evaluation after 5e7 training steps. Each score is average of 3 experiments with different seeds.
|
I will withdraw the pull request once and try to investigate the cause of problem because all scores observed is lower than PFRL original scores of IQN (https://github.com/pfnet/pfrl/tree/master/examples/atari/reproduction/iqn ). |
This PR adds Double IQN as in ChainerRL as suggested in #4.
This PR includes new agent
DoubleIQN
and it's training example as well as tests, based on ChainerRL implementation and IQN in PFRL .