Skip to content

Latest commit

 

History

History

atari

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"