def __init__(self):
super(GridSearchPong, self).__init__(
DQNPong, {
"learning_rate": [1e-3, 1e-5],
"greedy_epsilon": [
utils.CappedLinear(2e5, 0.1, 0.01),
utils.CappedLinear(5e5, 0.3, 0.01)
]
})
def __init__(self):
env = gym.make("BreakoutNoFrameskip-v4")
env = full_wrap_dqn(env)
env = RewardLongerEnv(env)
f = f_dqn_atari(env.action_space.n)
super(DQNBreakout, self).__init__(
env,
f,
episode_n=10000,
greedy_epsilon=utils.CappedLinear(2e5, 0.1, 0.01),
batch_size=64,
network_optimizer_ctor=lambda: hrl.network.LocalOptimizer(
tf.train.AdamOptimizer(1e-4), grad_clip=10.0))
def __init__(self,
greedy_epsilon=utils.CappedLinear(5e5, 0.1, 0.01),
learning_rate=1e-3):
env = gym.make("PongNoFrameskip-v4")
env = full_wrap_dqn(env)
f = f_dqn_atari(env.action_space.n)
network_optimizer_ctor = lambda: hrl.network.LocalOptimizer(
tf.train.AdamOptimizer(learning_rate), grad_clip=10.0)
super(DQNPong,
self).__init__(env,
f,
greedy_epsilon=greedy_epsilon,
network_optimizer_ctor=network_optimizer_ctor)
def __init__(self,
env,
f_create_q,
episode_n=1000,
discount_factor=0.99,
ddqn=False,
target_sync_interval=100,
target_sync_rate=1.0,
update_interval=4,
replay_size=5000,
batch_size=32,
greedy_epsilon=utils.CappedLinear(5e5, 0.1, 0.01),
network_optimizer_ctor=lambda: hrl.network.LocalOptimizer(
tf.train.AdamOptimizer(1e-3), grad_clip=10.0)):
super(DQNAtari,
self).__init__(env, f_create_q, episode_n, discount_factor, ddqn,
target_sync_interval, target_sync_rate,
update_interval, replay_size, batch_size,
greedy_epsilon, network_optimizer_ctor)
def __init__(self,
env=None,
f_create_q=None,
episode_n=10000,
discount_factor=0.99,
ddqn=False,
target_sync_interval=100,
target_sync_rate=1.0,
update_interval=4,
replay_size=1000,
batch_size=8,
lower_weight=1.0,
upper_weight=1.0,
neighbour_size=8,
greedy_epsilon=utils.CappedLinear(2e5, 0.1, 0.01),
learning_rate=1e-4,
sampler_creator=None,
**kwargs):
if env is None:
env = gym.make("BreakoutNoFrameskip-v4")
env = full_wrap_dqn(env)
env = RewardLongerEnv(env)
f_create_q = f_dqn_atari(env.action_space.n)
def f(args):
sampler = sampling.TruncateTrajectorySampler(
replay_memory=playback.NearPrioritizedPlayback(5000),
batch_size=batch_size,
trajectory_length=8,
interval=update_interval)
return sampler
sampler_creator = f
super(APEROTDQNBreakout,
self).__init__(env, f_create_q, episode_n, discount_factor, ddqn,
target_sync_interval, target_sync_rate,
update_interval, replay_size, batch_size,
lower_weight, upper_weight, neighbour_size,
greedy_epsilon, learning_rate, sampler_creator,
**kwargs)