def __init__(self, *args):
paramLen = len(DRQNPrioritized.newParameters)
super().__init__(*args[:-paramLen])
self.alpha = float(args[-paramLen])
empty_state = self.get_empty_state()
self.memory = ExperienceReplay.PrioritizedReplayBuffer(self, self.memory_size, TransitionFrame(empty_state, -1, 0, empty_state, False),
history_length = self.historylength, alpha = self.alpha)
def __init__(self, *args):
# Initializing model parameters
paramLen = len(DDPG.newParameters)
super().__init__(*args[:-paramLen])
self.batch_size, self.memory_size, self.target_update_interval, self.tau = [
int(arg) for arg in args[-paramLen:]
]
empty_state = self.get_empty_state()
self.memory = ExperienceReplay.ReplayBuffer(
self, self.memory_size,
TransitionFrame(empty_state, -1, 0, empty_state, False))
# Learning rate for actor-critic models
critic_lr = 0.002
actor_lr = 0.001
self.critic_optimizer = tf.keras.optimizers.Adam(critic_lr)
self.actor_optimizer = tf.keras.optimizers.Adam(actor_lr)
# self.ou_noise = OUNoise(self.action_size)
self.actor_model = self.get_actor()
self.critic_model = self.get_critic()
self.target_actor = self.get_actor()
self.target_critic = self.get_critic()
# Making the weights equal initially
self.target_actor.set_weights(self.actor_model.get_weights())
self.target_critic.set_weights(self.critic_model.get_weights())
self.total_steps = 0
self.allMask = np.full((1, self.action_size), 1)
self.allBatchMask = np.full((self.batch_size, self.action_size), 1)
def __init__(self, *args):
paramLen = len(DeepQHindsight.newParameters)
super().__init__(*args)
empty_state = self.get_empty_state()
self.memory = ExperienceReplay.HindsightReplayBuffer(
self, self.memory_size,
TransitionFrame(empty_state, -1, 0, empty_state, False))
def __init__(self, *args):
paramLen = len(Rainbow.newParameters)
super().__init__(*args[:-paramLen])
Qparams = []
for i in range(3):
Qparams.append(DeepQ.newParameters[i].default)
'''self.batch_size, self.memory_size, self.target_update_interval = [int(param) for param in Qparams]
#self.batch_size, self.memory_size, self.target_update_interval, _ = [int(arg) for arg in args[-paramLen:]]
_, _, _, self.learning_rate = [arg for arg in args[-paramLen:]]
self.memory = ExperienceReplay.ReplayBuffer(self, self.memory_size, TransitionFrame(empty_state, -1, 0, empty_state, False))
self.total_steps = 0
self.allMask = np.full((1, self.action_size), 1)
self.allBatchMask = np.full((self.batch_size, self.action_size), 1)'''
empty_state = self.get_empty_state()
self.total_steps = 0
self.model = self.buildQNetwork()
self.target = self.buildQNetwork()
self.lr = 0.001
self.memory = ExperienceReplay.ReplayBuffer(
self, self.memory_size,
TransitionFrame(empty_state, -1, 0, empty_state, False))
# Parameters used for Bellman Distribution
self.num_atoms = 51
self.v_min = -10
self.v_max = 10
self.delta_z = (self.v_max - self.v_min) / float(self.num_atoms - 1)
self.z = [self.v_min + i * self.delta_z for i in range(self.num_atoms)]
self.sample_size = min(self.batch_size, self.memory_size)
# Initialize prioritization exponent
self.p = 0.5
self.allBatchMask = np.full((self.sample_size, self.num_atoms), 1)
def __init__(self, *args):
super().__init__(*args)
empty_state = self.get_empty_state()
self.memory = ExperienceReplay.ReplayBuffer(
self,
self.memory_size,
ActionTransitionFrame(-1, empty_state, -1, 0, empty_state, False),
history_length=self.historylength)
def __init__(self, *args):
paramLen = len(ADRQNHindsight.newParameters)
super().__init__(*args)
empty_state = self.get_empty_state()
self.memory = ExperienceReplay.HindsightReplayBuffer(
self,
self.memory_size,
ActionTransitionFrame(-1, empty_state, -1, 0, empty_state, False),
history_length=self.historylength)
def __init__(self, *args):
# Initializing model parameters
paramLen = len(SAC.newParameters)
super().__init__(*args[:-paramLen])
self.batch_size, self.memory_size, self.target_update_interval, self.tau, self.temperature = [
int(arg) for arg in args[-paramLen:]
]
self.polyak = 0.01
self.total_steps = 0
empty_state = self.get_empty_state()
self.memory = ExperienceReplay.ReplayBuffer(
self, self.memory_size,
TransitionFrame(empty_state, -1, 0, empty_state, False))
# Learning rate for actor-critic models
critic_lr = 0.002
actor_lr = 0.001
self.critic_optimizer = tf.keras.optimizers.Adam(critic_lr)
self.actor_optimizer = tf.keras.optimizers.Adam(actor_lr)
self.actor_network = actorNetwork(self.action_size)
self.soft_Q_network = self.q_network()
self.soft_Q_targetnetwork = self.q_network()
self.soft_Q_network1 = self.q_network()
self.soft_Q_targetnetwork1 = self.q_network()
# Building up 2 soft q-function with their relative targets
in1 = tf.keras.Input(shape=self.state_size, dtype=tf.float64)
in2 = tf.keras.Input(shape=self.action_size, dtype=tf.float64)
self.soft_Q_network([in1, in2])
self.soft_Q_targetnetwork([in1, in2])
force_update(self.soft_Q_network.variables,
self.soft_Q_targetnetwork.variables)
self.soft_Q_network1([in1, in2])
self.soft_Q_targetnetwork1([in1, in2])
force_update(self.soft_Q_network1.variables,
self.soft_Q_targetnetwork1.variables)
# Optimizers for the networks
self.softq_optimizer = tf.keras.optimizers.Adam(
learning_rate=critic_lr)
self.softq_optimizer2 = tf.keras.optimizers.Adam(
learning_rate=critic_lr)
self.actor_optimizer = tf.keras.optimizers.Adam(learning_rate=actor_lr)
def __init__(self, *args):
paramLen = len(DeepQ.newParameters)
super().__init__(*args[:-paramLen])
self.batch_size, self.memory_size, self.target_update_interval = [
int(arg) for arg in args[-paramLen:]
]
self.model = self.buildQNetwork()
self.target = self.buildQNetwork()
empty_state = self.get_empty_state()
self.memory = ExperienceReplay.ReplayBuffer(
self, self.memory_size,
TransitionFrame(empty_state, -1, 0, empty_state, False))
self.total_steps = 0
self.allMask = np.full((1, self.action_size), 1)
self.allBatchMask = np.full((self.batch_size, self.action_size), 1)
def __init__(self, *args):
paramLen = len(PPO.newParameters)
super().__init__(*args[:-paramLen])
empty_state = self.get_empty_state()
# Initialize parameters
self.memory = ExperienceReplay.ReplayBuffer(
self, self.memory_size,
TransitionFrame(empty_state, -1, 0, empty_state, False))
self.total_steps = 0
self.actorIts = 2
self.allMask = np.full((1, self.action_size), 1)
self.allBatchMask = np.full((self.batch_size, self.action_size), 1)
self.policy_lr = 0.001
self.value_lr = 0.001
self.policy_model = Actor(self.state_size, self.action_size,
self.policy_lr).policy_network()
self.value_model = Critic(self.state_size, self.action_size,
self.value_lr).value_network()
def __init__(self, *args):
print("Stuff PPO:")
print(str(args))
paramLen = len(PPO.newParameters)
super().__init__(*args[:-paramLen])
empty_state = self.get_empty_state()
# Initialize parameters
self.memory = ExperienceReplay.ReplayBuffer(
self, self.memory_size,
TransitionFrame(empty_state, -1, 0, empty_state, False))
self.total_steps = 0
self.allMask = np.full((1, self.action_size), 1)
self.allBatchMask = np.full((self.batch_size, self.action_size), 1)
#self.batch_size, _, _, self.horizon, self.epochSize, _, _ = [int(arg) for arg in args[-paramLen:]]
#_, self.policy_lr, self.value_lr, _, _, self.epsilon, self.lam = [arg for arg in args[-paramLen:]]
self.policy_lr = 0.001
self.value_lr = 0.001
def resetBuffer(self):
self.memory = ExperienceReplay.ReplayBuffer(self, self.memory_size, self.historylength)