- [x] I have marked all applicable categories:
+ [ ] exception-raising fix
+ [x] algorithm implementation fix
+ [ ] documentation modification
+ [ ] new feature
- [x] I have reformatted the code using `make format` (**required**)
- [x] I have checked the code using `make commit-checks` (**required**)
- [x] If applicable, I have mentioned the relevant/related issue(s)
- [x] If applicable, I have listed every items in this Pull Request
below
While trying to debug Atari PPO+LSTM, I found significant gap between
our Atari PPO example vs [CleanRL's Atari PPO w/
EnvPool](https://docs.cleanrl.dev/rl-algorithms/ppo/#ppo_atari_envpoolpy).
I tried to align our implementation with CleaRL's version, mostly in
hyper parameter choices, and got significant gain in Breakout, Qbert,
SpaceInvaders while on par in other games. After this fix, I would
suggest updating our [Atari
Benchmark](https://tianshou.readthedocs.io/en/master/tutorials/benchmark.html)
PPO experiments.
A few interesting findings:
- Layer initialization helps stabilize the training and enable the use
of larger learning rates; without it, larger learning rates will trigger
NaN gradient very quickly;
- ppo.py#L97-L101: this change helps training stability for reasons I do
not understand; also it makes the GPU usage higher.
Shoutout to [CleanRL](https://github.com/vwxyzjn/cleanrl) for a
well-tuned Atari PPO reference implementation!
662af52820