def defaults(cls):
# Preprocessing.
subsample = 2
frame_skip = 4
history = 4
delta = False
frame_max = 2
noop_max = 30
# Architecture.
network = 'network_dqn_2015'
replay_capacity = 1e5 # 1e6
start_learning = 5e4
# Exploration.
epsilon = dict(from_=1.0,
to=0.1,
test=0.05,
over=1e6,
offset=start_learning)
# Learning.
batch_size = 32
sync_target = 2500
# Optimizer.
initial_learning_rate = 2.5e-4
optimizer = tf.train.RMSPropOptimizer
rms_decay = 0.95
rms_epsilon = 0.1
return merge_dicts(super().defaults(), locals())
def defaults(cls):
# Preprocessing.
subsample = 2
frame_skip = 4
history = 4
delta = False
frame_max = 2
noop_max = 30
# Architecture.
learners = 16
apply_gradient = 5
network = 'network_a3c_lstm'
scale_critic_loss = 0.5
regularize = 0.01
# Optimizer.
initial_learning_rate = 7e-4
optimizer = tf.train.RMSPropOptimizer
rms_decay = 0.99
rms_epsilon = 0.1
return merge_dicts(super().defaults(), locals())
def defaults(cls):
frameskip = 3
fps = 30 / frameskip
sensitivity = 0.3
viewer = Viewer
return merge_dicts(super().defaults(), locals())