662af52820
- [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!
14 KiB
Atari Environment
EnvPool
We highly recommend using envpool to run the following experiments. To install, in a linux machine, type:
pip install envpool
After that, atari_wrapper will automatically switch to envpool's Atari env. EnvPool's implementation is much faster (about 2~3x faster for pure execution speed, 1.5x for overall RL training pipeline) than python vectorized env implementation, and it's behavior is consistent to that approach (OpenAI wrapper), which will describe below.
For more information, please refer to EnvPool's GitHub, Docs, and 3rd-party report.
ALE-py
The sample speed is ~3000 env step per second (~12000 Atari frame per second in fact since we use frame_stack=4) under the normal mode (use a CNN policy and a collector, also storing data into the buffer).
The env wrapper is a crucial thing. Without wrappers, the agent cannot perform well enough on Atari games. Many existing RL codebases use OpenAI wrapper, but it is not the original DeepMind version (related issue). Dopamine has a different wrapper but unfortunately it cannot work very well in our codebase.
DQN (single run)
One epoch here is equal to 100,000 env step, 100 epochs stand for 10M.
| task | best reward | reward curve | parameters | time cost |
|---|---|---|---|---|
| PongNoFrameskip-v4 | 20 |  |
python3 atari_dqn.py --task "PongNoFrameskip-v4" --batch-size 64 |
~30 min (~15 epoch) |
| BreakoutNoFrameskip-v4 | 316 |  |
python3 atari_dqn.py --task "BreakoutNoFrameskip-v4" --test-num 100 |
3~4h (100 epoch) |
| EnduroNoFrameskip-v4 | 670 |  |
python3 atari_dqn.py --task "EnduroNoFrameskip-v4 " --test-num 100 |
3~4h (100 epoch) |
| QbertNoFrameskip-v4 | 7307 |  |
python3 atari_dqn.py --task "QbertNoFrameskip-v4" --test-num 100 |
3~4h (100 epoch) |
| MsPacmanNoFrameskip-v4 | 2107 |  |
python3 atari_dqn.py --task "MsPacmanNoFrameskip-v4" --test-num 100 |
3~4h (100 epoch) |
| SeaquestNoFrameskip-v4 | 2088 |  |
python3 atari_dqn.py --task "SeaquestNoFrameskip-v4" --test-num 100 |
3~4h (100 epoch) |
| SpaceInvadersNoFrameskip-v4 | 812.2 |  |
python3 atari_dqn.py --task "SpaceInvadersNoFrameskip-v4" --test-num 100 |
3~4h (100 epoch) |
Note: The eps_train_final and eps_test in the original DQN paper is 0.1 and 0.01, but some works found that smaller eps helps improve the performance. Also, a large batchsize (say 64 instead of 32) will help faster convergence but will slow down the training speed.
We haven't tuned this result to the best, so have fun with playing these hyperparameters!
C51 (single run)
One epoch here is equal to 100,000 env step, 100 epochs stand for 10M.
| task | best reward | reward curve | parameters |
|---|---|---|---|
| PongNoFrameskip-v4 | 20 |  |
python3 atari_c51.py --task "PongNoFrameskip-v4" --batch-size 64 |
| BreakoutNoFrameskip-v4 | 536.6 |  |
python3 atari_c51.py --task "BreakoutNoFrameskip-v4" --n-step 1 |
| EnduroNoFrameskip-v4 | 1032 |  |
python3 atari_c51.py --task "EnduroNoFrameskip-v4 " |
| QbertNoFrameskip-v4 | 16245 |  |
python3 atari_c51.py --task "QbertNoFrameskip-v4" |
| MsPacmanNoFrameskip-v4 | 3133 |  |
python3 atari_c51.py --task "MsPacmanNoFrameskip-v4" |
| SeaquestNoFrameskip-v4 | 6226 |  |
python3 atari_c51.py --task "SeaquestNoFrameskip-v4" |
| SpaceInvadersNoFrameskip-v4 | 988.5 |  |
python3 atari_c51.py --task "SpaceInvadersNoFrameskip-v4" |
Note: The selection of n_step is based on Figure 6 in the Rainbow paper.
QRDQN (single run)
One epoch here is equal to 100,000 env step, 100 epochs stand for 10M.
| task | best reward | reward curve | parameters |
|---|---|---|---|
| PongNoFrameskip-v4 | 20 |  |
python3 atari_qrdqn.py --task "PongNoFrameskip-v4" --batch-size 64 |
| BreakoutNoFrameskip-v4 | 409.2 |  |
python3 atari_qrdqn.py --task "BreakoutNoFrameskip-v4" --n-step 1 |
| EnduroNoFrameskip-v4 | 1055.9 |  |
python3 atari_qrdqn.py --task "EnduroNoFrameskip-v4" |
| QbertNoFrameskip-v4 | 14990 |  |
python3 atari_qrdqn.py --task "QbertNoFrameskip-v4" |
| MsPacmanNoFrameskip-v4 | 2886 |  |
python3 atari_qrdqn.py --task "MsPacmanNoFrameskip-v4" |
| SeaquestNoFrameskip-v4 | 5676 |  |
python3 atari_qrdqn.py --task "SeaquestNoFrameskip-v4" |
| SpaceInvadersNoFrameskip-v4 | 938 |  |
python3 atari_qrdqn.py --task "SpaceInvadersNoFrameskip-v4" |
IQN (single run)
One epoch here is equal to 100,000 env step, 100 epochs stand for 10M.
| task | best reward | reward curve | parameters |
|---|---|---|---|
| PongNoFrameskip-v4 | 20.3 |  |
python3 atari_iqn.py --task "PongNoFrameskip-v4" --batch-size 64 |
| BreakoutNoFrameskip-v4 | 496.7 |  |
python3 atari_iqn.py --task "BreakoutNoFrameskip-v4" --n-step 1 |
| EnduroNoFrameskip-v4 | 1545 |  |
python3 atari_iqn.py --task "EnduroNoFrameskip-v4" |
| QbertNoFrameskip-v4 | 15342.5 |  |
python3 atari_iqn.py --task "QbertNoFrameskip-v4" |
| MsPacmanNoFrameskip-v4 | 2915 |  |
python3 atari_iqn.py --task "MsPacmanNoFrameskip-v4" |
| SeaquestNoFrameskip-v4 | 4874 |  |
python3 atari_iqn.py --task "SeaquestNoFrameskip-v4" |
| SpaceInvadersNoFrameskip-v4 | 1498.5 |  |
python3 atari_iqn.py --task "SpaceInvadersNoFrameskip-v4" |
FQF (single run)
One epoch here is equal to 100,000 env step, 100 epochs stand for 10M.
| task | best reward | reward curve | parameters |
|---|---|---|---|
| PongNoFrameskip-v4 | 20.7 |  |
python3 atari_fqf.py --task "PongNoFrameskip-v4" --batch-size 64 |
| BreakoutNoFrameskip-v4 | 517.3 |  |
python3 atari_fqf.py --task "BreakoutNoFrameskip-v4" --n-step 1 |
| EnduroNoFrameskip-v4 | 2240.5 |  |
python3 atari_fqf.py --task "EnduroNoFrameskip-v4" |
| QbertNoFrameskip-v4 | 16172.5 |  |
python3 atari_fqf.py --task "QbertNoFrameskip-v4" |
| MsPacmanNoFrameskip-v4 | 2429 |  |
python3 atari_fqf.py --task "MsPacmanNoFrameskip-v4" |
| SeaquestNoFrameskip-v4 | 10775 |  |
python3 atari_fqf.py --task "SeaquestNoFrameskip-v4" |
| SpaceInvadersNoFrameskip-v4 | 2482 |  |
python3 atari_fqf.py --task "SpaceInvadersNoFrameskip-v4" |
Rainbow (single run)
One epoch here is equal to 100,000 env step, 100 epochs stand for 10M.
| task | best reward | reward curve | parameters |
|---|---|---|---|
| PongNoFrameskip-v4 | 21 |  |
python3 atari_rainbow.py --task "PongNoFrameskip-v4" --batch-size 64 |
| BreakoutNoFrameskip-v4 | 684.6 |  |
python3 atari_rainbow.py --task "BreakoutNoFrameskip-v4" --n-step 1 |
| EnduroNoFrameskip-v4 | 1625.9 |  |
python3 atari_rainbow.py --task "EnduroNoFrameskip-v4" |
| QbertNoFrameskip-v4 | 16192.5 |  |
python3 atari_rainbow.py --task "QbertNoFrameskip-v4" |
| MsPacmanNoFrameskip-v4 | 3101 |  |
python3 atari_rainbow.py --task "MsPacmanNoFrameskip-v4" |
| SeaquestNoFrameskip-v4 | 2126 |  |
python3 atari_rainbow.py --task "SeaquestNoFrameskip-v4" |
| SpaceInvadersNoFrameskip-v4 | 1794.5 |  |
python3 atari_rainbow.py --task "SpaceInvadersNoFrameskip-v4" |
PPO (single run)
One epoch here is equal to 100,000 env step, 100 epochs stand for 10M.
| task | best reward | reward curve | parameters |
|---|---|---|---|
| PongNoFrameskip-v4 | 20.2 |  |
python3 atari_ppo.py --task "PongNoFrameskip-v4" |
| BreakoutNoFrameskip-v4 | 441.8 |  |
python3 atari_ppo.py --task "BreakoutNoFrameskip-v4" |
| EnduroNoFrameskip-v4 | 1245.4 |  |
python3 atari_ppo.py --task "EnduroNoFrameskip-v4" |
| QbertNoFrameskip-v4 | 17395 |  |
python3 atari_ppo.py --task "QbertNoFrameskip-v4" |
| MsPacmanNoFrameskip-v4 | 2098 |  |
python3 atari_ppo.py --task "MsPacmanNoFrameskip-v4" |
| SeaquestNoFrameskip-v4 | 882 |  |
python3 atari_ppo.py --task "SeaquestNoFrameskip-v4" --lr 1e-4 |
| SpaceInvadersNoFrameskip-v4 | 1340.5 |  |
python3 atari_ppo.py --task "SpaceInvadersNoFrameskip-v4" |
SAC (single run)
One epoch here is equal to 100,000 env step, 100 epochs stand for 10M.
| task | best reward | reward curve | parameters |
|---|---|---|---|
| PongNoFrameskip-v4 | 20.1 |  |
python3 atari_sac.py --task "PongNoFrameskip-v4" |
| BreakoutNoFrameskip-v4 | 211.2 |  |
python3 atari_sac.py --task "BreakoutNoFrameskip-v4" --n-step 1 --actor-lr 1e-4 --critic-lr 1e-4 |
| EnduroNoFrameskip-v4 | 1290.7 |  |
python3 atari_sac.py --task "EnduroNoFrameskip-v4" |
| QbertNoFrameskip-v4 | 13157.5 |  |
python3 atari_sac.py --task "QbertNoFrameskip-v4" |
| MsPacmanNoFrameskip-v4 | 3836 |  |
python3 atari_sac.py --task "MsPacmanNoFrameskip-v4" |
| SeaquestNoFrameskip-v4 | 1772 |  |
python3 atari_sac.py --task "SeaquestNoFrameskip-v4" |
| SpaceInvadersNoFrameskip-v4 | 649 |  |
python3 atari_sac.py --task "SpaceInvadersNoFrameskip-v4" |