Can not gain same values as in code.

[VCasecnikovs]

Good evening,
I have tried to train SAC on PACMAN 2 times with the default config file and also with the slightly changed config file.
I change the batch size to 2048 and use_per to True. Unfortunately, I didn't get a good policy.
Could You tell me where could be an error and how to solve it?
Maybe I chose bad params...

Best regards,
Vadims.

[toshikwa]

Maybe I chose bad params...

If you think so, why don't you try the default parameters to see what is wrong?

First of all, batch_size=2048 is too big.
Also, if you modified the code, I can't help you with debugging.

Here are the results of SAC-Discrete in MsPacmanNoFramsskep-v4 I ran today, and it learned well.

Default Config

Default Config + use_per=True

[VCasecnikovs]

Good morning,
The red one is the default parameters.
I'll try to retest with default config again without any changes in code.
Thank you for saying that I should control BS, what is more interesting bigger BS gave me 1200 on the test.
Best regards,
Vadims.

[VCasecnikovs]

Good evening,
I have tested with the default settings, all works!
Also, I have tested with my pull request. FloatTensor doesn't lower the performance. Results are the same, but these changes make code more universal.

Hope You will merge it,
Best regards,
Vadims.