Skip to content
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

Potential descrepancies between paper and code #2

Closed
ikostrikov opened this issue Feb 15, 2021 · 2 comments
Closed

Potential descrepancies between paper and code #2

ikostrikov opened this issue Feb 15, 2021 · 2 comments

Comments

@ikostrikov
Copy link

Hi Philip,

Thanks for sharing this amazing repository and the paper!

The performance reported in the paper seems to outperform the author's original implementation and I was trying to incorporate the changes into my personal repository.

While looking at the code I've noticed that there is a difference in one of the hyper parameters defined in Table 1 and this code. In particular, in the table it says that the MLP has 2 hidden layers while in the code is seems to be set to 3: https://github.com/fiorenza2/OffCon3/blob/master/networks.py#L14

Please let me know if I missed something.

@philipjball
Copy link
Owner

Hi Ilya,

Great catch! Just looked at author's code too just to make sure and they have only 2 hidden layers; I will make it clear in the README and update the paper accordingly :)

I'll also rerun some experiments to see if the diff is actually down to these "superior" hyperparams.

@philipjball
Copy link
Owner

Hi Ilya,

Apologies for the delay in getting back; I've rerun some quick experiments with the 2 hidden layer MLPs and will include the comparison in the appendix for the updated report. Long story short is that in HalfCheetah the results are strikingly worse with 2 layers, but in Hopper SAC improves significantly to match TD3. Otherwise it's basically the same in the limit, with some environments having maybe faster convergence with the shallower networks.

I will make it clear that this implementation has the diff you outlined compared to the authors' code :)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants