adaa8aefa8
* fixes to enjoy_cartpole, enjoy_mountaincar.py
* fixed {train,enjoy}_pong, removed enjoy_retro
* set number of timesteps to 1e7 in train_pong
* flake8 complaints
* use synchronous version fo acktr in test_env_after_learn
* flake8
37 lines
823 B
Python
37 lines
823 B
Python
from baselines import deepq
|
|
from baselines import bench
|
|
from baselines import logger
|
|
from baselines.common.atari_wrappers import make_atari
|
|
|
|
|
|
def main():
|
|
logger.configure()
|
|
env = make_atari('PongNoFrameskip-v4')
|
|
env = bench.Monitor(env, logger.get_dir())
|
|
env = deepq.wrap_atari_dqn(env)
|
|
|
|
model = deepq.learn(
|
|
env,
|
|
"conv_only",
|
|
convs=[(32, 8, 4), (64, 4, 2), (64, 3, 1)],
|
|
hiddens=[256],
|
|
dueling=True,
|
|
lr=1e-4,
|
|
total_timesteps=int(1e7),
|
|
buffer_size=10000,
|
|
exploration_fraction=0.1,
|
|
exploration_final_eps=0.01,
|
|
train_freq=4,
|
|
learning_starts=10000,
|
|
target_network_update_freq=1000,
|
|
gamma=0.99,
|
|
)
|
|
|
|
model.save('pong_model.pkl')
|
|
env.close()
|
|
|
|
|
|
|
|
if __name__ == '__main__':
|
|
main()
|