I updated the batch size, it shows a more robust training for me.

What's causing all the test failures?

The performance of test scenarios are not same in case of cpu and gpu. I saw the same behavior in my machine ... so currently I have no idea why CI fails exactly.

btw can we remove the performance check?

@Trinkle23897 could you please look into this?

Sure, please wait for me to finish my homework though...

Is it because of pistonball_v4 vs pistonball_v6? @mahi97
I'm thinking about the dependency issue of both gym and pettingzoo. Maybe we should open another PR to fix that issue (because it breaks many tests such as Pendulum-v0 vs Pendulum-v1)?

@Trinkle23897 Can this be merged now since all tests are passing?

@mahi97 would you be willing to create a follow up PR to this that bumps the versions of Gym and PettingZoo?

https://readthedocs.org/projects/tianshou/builds/16093735/

That's all just pettingzoo/gym versioning errors if I'm reading correctly. Hopefully @mahi97 would be willing to make a follow up PR addressing this?

Sure things, I will continue this on another PR.

That's all just pettingzoo/gym versioning errors if I'm reading correctly. Hopefully @mahi97 would be willing to make a follow up PR addressing this?

Hopefully #534 will resolve this issue.

Fixed. However, the related documentation of multi-agent is outdated. We should switch the example to Pettingzoo. https://tianshou.readthedocs.io/en/master/tutorials/tictactoe.html

@Trinkle23897 @mahi97 we still aren't on the most recent release of PettingZoo though, right?

No, it's the newest.

However, pettingzoo doesn't have conda channel to publish package, which results in failure of publish tianshou's conda pkg: conda-forge/tianshou-feedstock#13
I'll fix it today by silently import pettingzoo instead of force-install. Or maybe you can add the conda channel for pettingzoo?

PettingZoo and Gym both aren't hosted on Conda. Doing this would require a very large amount of work, and for Gym won't actually be possible for awhile until certain dependencies are changed. The Gym version on Conda is not official and I would generally recommend against using it.

@mahi97 would you mind replacing the examples and built in environments in tianshou with PettingZoo ones like Jiayi mentioned?

Sure things, I'll do it on Monday 👍

Sorry for the huge delay. I'm caught up in some tasks with tight deadlines.
Unfortunately, I can't start working on this until May.

How can the pettingzoo wrapper support .state() api for some tasks?

How can the pettingzoo wrapper support .state() api for some tasks?

Firstly I try to implement a new buffer class that stores the global information such as the states separately, but found it incompatible within the collector.collect() function.

Should I just put the global information in the return info dict at the cost of multiple duplications?