def __init__(self,
max_velocity,
turn_speed,
max_health,
max_armor,
spawn_point=(200, 200),
starting_angle=0,
starter_weapon_pack=None,
starter_ammo_pack=None,
color='#303030',
radius=10):
BaseAgent.__init__(self,
max_velocity,
turn_speed,
max_health,
max_armor,
spawn_point,
starting_angle,
starter_weapon_pack,
starter_ammo_pack,
color,
radius)
input_layer = Input(shape=(17, 13))
flattened_input = Flatten()(input_layer)
inner_layer = Dense(20, activation='relu')(flattened_input)
output_layer = Dense(11, activation='tanh')(inner_layer)
self.model = Model(input_layer, output_layer)
self.model.compile(RMSprop(),
loss='hinge')
self.delta = 1-1e-5
self.epsilon = 1
def __init__(self,
max_velocity,
turn_speed,
max_health,
max_armor,
spawn_point=(200, 200),
starting_angle=0,
starter_weapon_pack=None,
starter_ammo_pack=None,
color='#303030',
radius=10):
BaseAgent.__init__(self,
max_velocity,
turn_speed,
max_health,
max_armor,
spawn_point,
starting_angle,
starter_weapon_pack,
starter_ammo_pack,
color,
radius)
#input_layer = Input(shape=(17, 13))
#inner_layer1 = Convolution1D(20, 5, activation='relu')(input_layer)
#pooling1 = MaxPooling1D(2)(inner_layer1)
#inner_layer2 = Convolution1D(20, 3, activation='relu')(pooling1)
#pooling2 = MaxPooling1D(2)(inner_layer2)
#flattened = Flatten()(pooling2)
#inner_layer3 = Dense(20, activation='relu')(flattened)
#bn = BatchNormalization()(inner_layer3)
#output_layer = Dense(11, activation='tanh')(bn)
#self.model = Model(input_layer, output_layer)
#self.model.compile(RMSprop(),
# loss='hinge')
self.delta = 1-1e-5 #decrease coefficient of epsilon-greedy
self.epsilon = 1 #probability of random action
self.max_memory_size = 50000
self.observation_memory = []
self.action_memory = []
self.max_buffer_size = 100
self.observation_buffer = []
self.action_buffer = []
self.reward_buffer = []
self.tau = 0.97
self.batch_size = 16
self.skip = 5
self.t = 0
self.episode_rewards = []
self.age = 0
self.to_learn = True
def __init__(self, config, session):
BaseAgent.__init__(self, config, session)
self.action_modes = {str(config.testing_epsilon)+"_greedy":self.e_greedy_action,
"plan_"+str(config.testing_epsilon)+"_greedy":self.plan_e_greedy_action}
self.default_action_mode = self.action_modes.items()[0]
self.action_mode = self.default_action_mode
# build the net
with tf.device(config.device):
# Create all variables and the FIFOQueue
self.state_ph = tf.placeholder(
tf.float32, [None, 84, 84, 4], name="state_ph")
self.action_ph = tf.placeholder(tf.int64, [None], name="action_ph")
self.reward_ph = tf.placeholder(tf.float32, [None], name="reward_ph")
self.terminal_ph = tf.placeholder(tf.float32, [None], name="terminal_ph")
self.stateT_ph = tf.placeholder(
tf.float32, [None, 84, 84, 4], name="stateT_ph")
# Define all the ops
with tf.variable_scope("Q"):
self.h_state = self.state_to_hidden(self.state_ph, config, "Normal")
self.Q = self.hidden_to_Q(self.h_state, config, "Normal")
self.predicted_reward = self.hidden_to_reward(self.h_state, config, "Normal")
self.predicted_h_state = self.hidden_to_hidden(self.h_state, self.action_ph, config, "Normal")
tf.get_variable_scope().reuse_variables()
self.predicted_next_Q = self.hidden_to_Q(self.predicted_h_state, config, "Normal")
with tf.variable_scope("QT"):
self.h_stateT = self.state_to_hidden(self.stateT_ph, config, "Target")
self.QT = self.hidden_to_Q(self.h_stateT, config, "Target")
self.train_op = self.train_op(self.Q, self.predicted_reward,
self.predicted_next_Q, self.QT, self.reward_ph,
self.action_ph, self.terminal_ph, config, "Normal")
self.sync_QT_op = []
for W_pair in zip(
tf.get_collection("Target_weights"),
tf.get_collection("Normal_weights")):
self.sync_QT_op.append(W_pair[0].assign(W_pair[1]))
# Define the summary ops
self.Q_summary_op = tf.merge_summary(
tf.get_collection("Normal_summaries"))
self.QT_summary_op = tf.merge_summary(
tf.get_collection("Target_summaries"))
if config.logging:
self.summary_writter = tf.train.SummaryWriter(
self.config.log_path, self.sess.graph, flush_secs=20)
def __init__(self,
max_velocity,
turn_speed,
max_health,
max_armor,
spawn_point=(200, 200),
starting_angle=0,
starter_weapon_pack=None,
starter_ammo_pack=None,
color='#303030',
radius=10):
BaseAgent.__init__(self,
max_velocity,
turn_speed,
max_health,
max_armor,
spawn_point,
starting_angle,
starter_weapon_pack,
starter_ammo_pack,
color,
radius)
def __init__(self, config, url):
BaseAgent.__init__(self, config, url)
def __init__(self, config, url):
self.url = url
self.config = config
BaseAgent.__init__(self, config, url)