def dqn_per_gridworld():
hp = DictConfig({})
hp.steps = 1000
hp.batch_size = 500
hp.replay_batch = 100
hp.replay_size = 1000
hp.delete_freq = 100 * (hp.batch_size + hp.replay_size) # every 100 steps
hp.env_record_freq = 100
hp.env_record_duration = 25
hp.max_steps = 50
hp.grid_size = 4
hp.lr = 1e-3
hp.epsilon_exploration = 0.1
hp.gamma_discount = 0.9
model = (GenericConvModel(height=4,
width=4,
in_channels=4,
channels=[50],
out_size=4).float().to(device))
train_dqn_per(
GridWorldEnvWrapper,
model,
hp,
project_name="SimpleGridWorld",
run_name="dqn_per",
)
def test_dqn_vanilla(self, *_):
from dqn.dqn import train_dqn
hp = DictConfig({})
hp.steps = 2
hp.batch_size = 2
hp.env_record_freq = 0
hp.env_record_duration = 0
hp.max_steps = 50
hp.grid_size = 4
hp.lr = 1e-3
hp.epsilon_exploration = 0.1
hp.gamma_discount = 0.9
for case in env_cases:
print(case["env"].__name__)
model = GenericLinearModel(
in_size=case["input"],
units=[10],
out_size=case["output"],
flatten=case.get("flatten", False),
)
train_dqn(case["env"], model, hp)
def breakout_double_dqn():
hp = DictConfig({})
hp.steps = 2000
hp.batch_size = 50
hp.replay_batch = 50
hp.replay_size = 1000
hp.delete_freq = 50 * (hp.batch_size + hp.replay_size) # every 100 steps
hp.delete_percentage = 0.2
hp.env_record_freq = 100
hp.env_record_duration = 50
hp.lr = 1e-3
hp.gamma_discount = 0.9
# hp.epsilon_exploration = 0.1
hp.epsilon_flatten_step = 1500
hp.epsilon_start = 1
hp.epsilon_end = 0.1
hp.epsilon_decay_function = decay_functions.LINEAR
hp.target_model_sync_freq = 50
model = GenericConvModel(42, 42, 3, [50, 50, 50], [100], 4)
train_dqn_double(
BreakoutEnvWrapper, model, hp, project_name="Breakout", run_name="double_dqn"
)
def breakout_dqn():
hp = DictConfig({})
hp.steps = 2000
hp.batch_size = 32
hp.env_record_freq = 500
hp.env_record_duration = 100
hp.max_steps = 1000
hp.lr = 1e-3
hp.epsilon_exploration = 0.1
hp.gamma_discount = 0.9
model = GenericLinearModel(42 * 42 * 3, [100, 100], 4, flatten=True)
train_dqn(
BreakoutEnvWrapper, model, hp, project_name="Breakout", run_name="vanilla_dqn"
)
def dqn_double():
hp = DictConfig({})
hp.steps = 1000
hp.batch_size = 500
hp.replay_batch = 100
hp.replay_size = 1000
hp.delete_freq = 100 * (hp.batch_size + hp.replay_size) # every 100 steps
hp.env_record_freq = 100
hp.env_record_duration = 25
hp.max_steps = 50
hp.grid_size = 4
hp.lr = 1e-3
hp.gamma_discount = 0.9
# hp.epsilon_exploration = 0.1
hp.epsilon_flatten_step = 700
hp.epsilon_start = 1
hp.epsilon_end = 0.001
hp.epsilon_decay_function = decay_functions.LINEAR
hp.target_model_sync_freq = 50
model = (GenericConvModel(height=4,
width=4,
in_channels=4,
channels=[50],
out_size=4).float().to(device))
train_dqn_double(
GridWorldEnvWrapper,
model,
hp,
project_name="SimpleGridWorld",
run_name="dqn_target",
)
max_steps = 500
reward_range = (-10, 10) # TODO: Fix this
def __init__(self):
super().__init__()
self.env = gym.make(
"GDY-Sokoban---2-v0",
global_observer_type=gd.ObserverType.VECTOR,
player_observer_type=gd.ObserverType.VECTOR,
level=0,
)
if __name__ == "__main__":
hp = DictConfig({})
hp.steps = 10000
hp.batch_size = 1000
hp.env_record_freq = 500
hp.env_record_duration = 50
hp.max_steps = 200
hp.lr = 1e-3
hp.epsilon_exploration = 0.1
hp.gamma_discount = 0.9
model = GenericLinearModel(5 * 7 * 8, [10], 5,
flatten=True).float().to(device)
train_dqn(SokobanV2L0EnvWrapper, model, hp, name="SokobanV2L0")
