def __init__(self, num_states, num_actions, action_high, action_low, gamma=GAMMA, rho=RHO,
std_dev=STD_DEV):
# initialize everything
self.actor_network = ActorNetwork(num_states, num_actions, action_high)
self.critic_network = CriticNetwork(num_states, num_actions, action_high)
self.actor_target = ActorNetwork(num_states, num_actions, action_high)
self.critic_target = CriticNetwork(num_states, num_actions, action_high)
# Making the weights equal initially
self.actor_target.set_weights(self.actor_network.get_weights())
self.critic_target.set_weights(self.critic_network.get_weights())
self.buffer = ReplayBuffer(BUFFER_SIZE, BATCH_SIZE)
self.gamma = tf.constant(gamma)
self.rho = rho
self.action_high = action_high
self.action_low = action_low
self.num_states = num_states
self.num_actions = num_actions
self.noise = OUActionNoise(mean=np.zeros(1), std_deviation=float(std_dev) * np.ones(1))
# optimizers
self.critic_optimizer = tf.keras.optimizers.Adam(CRITIC_LR, amsgrad=True)
self.actor_optimizer = tf.keras.optimizers.Adam(ACTOR_LR, amsgrad=True)
# temporary variable for side effects
self.cur_action = None
# define update weights with tf.function for improved performance
@tf.function(
input_signature=[
tf.TensorSpec(shape=(None, num_states), dtype=tf.float32),
tf.TensorSpec(shape=(None, num_actions), dtype=tf.float32),
tf.TensorSpec(shape=(None, 1), dtype=tf.float32),
tf.TensorSpec(shape=(None, num_states), dtype=tf.float32),
tf.TensorSpec(shape=(None, 1), dtype=tf.float32),
])
def update_weights(s, a, r, sn, d):
"""
Function to update weights with optimizer
"""
with tf.GradientTape() as tape:
# define target
y = r + self.gamma * (1 - d) * self.critic_target([sn, self.actor_target(sn)])
# define the delta Q
critic_loss = tf.math.reduce_mean(tf.math.abs(y - self.critic_network([s, a])))
critic_grad = tape.gradient(critic_loss, self.critic_network.trainable_variables)
self.critic_optimizer.apply_gradients(
zip(critic_grad, self.critic_network.trainable_variables))
with tf.GradientTape() as tape:
# define the delta mu
actor_loss = -tf.math.reduce_mean(self.critic_network([s, self.actor_network(s)]))
actor_grad = tape.gradient(actor_loss, self.actor_network.trainable_variables)
self.actor_optimizer.apply_gradients(
zip(actor_grad, self.actor_network.trainable_variables))
return critic_loss, actor_loss
self.update_weights = update_weights
def __init__(self, input_size, output_size, hidden = 256, lr_actor=1.0e-3, lr_critic=1.0e-3, agent_number=0, tau=1.0e-2,
gamma=0.99, epsilon=1.0, epsilon_decay=0.99, weight_decay=0, clipgrad=.1, seed = 42):
super(Agent, self).__init__()
self.seed = seed
self.actor = ActorNetwork(input_size, output_size, name=f"Actor_Agent{agent_number}").to(device)
self.critic = CriticNetwork(input_size, output_size, name=f"Critic_Agent{agent_number}").to(device)
self.target_actor = ActorNetwork(input_size, output_size, name=f"Actor_Target_Agent{agent_number}").to(device)
self.target_critic = CriticNetwork(input_size, output_size, name=f"Critic_Target_Agent{agent_number}").to(device)
self.noise = OUActionNoise(mu=np.zeros(output_size))
self.tau = tau
self.epsilon = epsilon
self.epsilon_decay=epsilon_decay
self.gamma = gamma
self.clipgrad = clipgrad
self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=lr_actor)
self.critic_optimizer = optim.Adam(self.critic.parameters(), lr=lr_critic, weight_decay=weight_decay)
def __init__(self,
env_id,
alpha,
beta,
input_dims,
tau,
n_actions,
gamma=0.99,
max_size=1000000,
fc1_dims=256,
fc2_dims=256,
batch_size=256):
print("agent fc1 dims {}, fc2 dims {}".format(fc1_dims, fc2_dims))
self.gamma = gamma
self.tau = tau
self.batch_size = batch_size
self.alpha = alpha
self.beta = beta
self.memory = ReplayBuffer(max_size, input_dims, n_actions)
self.noise = OUActionNoise(mu=np.zeros(n_actions))
self.actor = ActorNetwork(alpha,
input_dims,
fc1_dims,
fc2_dims,
n_actions=n_actions,
name=env_id + '_actor')
self.critic = CriticNetwork(beta,
input_dims,
fc1_dims,
fc2_dims,
n_actions=n_actions,
name=env_id + '_critic')
self.target_actor = ActorNetwork(alpha,
input_dims,
fc1_dims,
fc2_dims,
n_actions=n_actions,
name=env_id + '_target_actor')
self.target_critic = CriticNetwork(beta,
input_dims,
fc1_dims,
fc2_dims,
n_actions=n_actions,
name=env_id + '_target_critic')
self.update_network_parameters(tau=1)
def test(env_info, total_episodes=3, noise_std=0.2):
ou_noise = OUActionNoise(mean=np.zeros(1),
std_deviation=float(noise_std) * np.ones(1))
for _ in range(total_episodes):
prev_state = env.reset()
while True:
env.render()
tf_prev_state = tf.expand_dims(tf.convert_to_tensor(prev_state), 0)
action = policy(actor_model, tf_prev_state, ou_noise, **env_info)
state, reward, done, info = env.step(action)
if done:
break
prev_state = state
def __init__(self,
alpha,
beta,
input_dims,
tau,
n_actions,
gamma=0.99,
max_size=1000000,
fc1_dims=400,
fc2_dims=300,
batch_size=64):
self.gamma = gamma
self.tau = tau
self.batch_size = batch_size
self.alpha = alpha
self.beta = beta
self.memory = ReplayBuffer(max_size, input_dims, n_actions)
self.noise = OUActionNoise(mu=np.zeros(n_actions))
self.actor = ActorNetwork(alpha,
input_dims,
fc1_dims,
fc2_dims,
n_actions=n_actions,
name='actor')
self.critic = CriticNetwork(beta,
input_dims,
fc1_dims,
fc2_dims,
n_actions=n_actions,
name='critic')
self.target_actor = ActorNetwork(alpha,
input_dims,
fc1_dims,
fc2_dims,
n_actions=n_actions,
name='target_actor')
self.target_critic = CriticNetwork(beta,
input_dims,
fc1_dims,
fc2_dims,
n_actions=n_actions,
name='target_critic')
self.update_network_parameters(tau=1)
def train(env_info, buffer, total_episodes=100, noise_std=0.2, gamma=0.99, tau=0.005):
ou_noise = OUActionNoise(mean=np.zeros(1), std_deviation=float(noise_std) * np.ones(1))
# Making the weights equal initially
target_actor.set_weights(actor_model.get_weights())
target_critic.set_weights(critic_model.get_weights())
# To store reward history of each episode
ep_reward_list = []
for ep in tqdm(range(total_episodes)):
prev_state = env.reset()
episodic_reward = 0
while True:
tf_prev_state = tf.expand_dims(tf.convert_to_tensor(prev_state), 0)
action = policy(actor_model, tf_prev_state, ou_noise, **env_info)
# Recieve state and reward from environment.
state, reward, done, info = env.step(action)
buffer.record((prev_state, action, reward, state))
episodic_reward += reward
buffer.learn(gamma)
update_target(target_actor.variables, actor_model.variables, tau)
update_target(target_critic.variables, critic_model.variables, tau)
if reward >= 40:
print(f"-- Episode: {ep} "+"-"*20)
print(f"Action: {action}")
print(f"Reward: {reward}\tDone: {done}\nState: {state}")
# End this episode when `done` is True
if done:
break
prev_state = state
ep_reward_list.append(episodic_reward)
def __init__(self,
state_size,
action_size,
lr_actor=LR_ACTOR,
lr_critic=LR_CRITIC,
random_seed=42,
num_agents=1):
"""Initialize Agent object.
Params
====
state_size (int): Dimension of each state
action_size (int): Dimension of each action
lr_actor (float): Learning rate for actor model
lr_critic (float): Learning Rate for critic model
random_seed (int): Random seed
num_agents (int): Number of agents
return
====
None
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.num_agents = num_agents
# Initialize time step (for updating every hyperparameters["update_every"] steps)
self.t_step = 0
# Actor network
self.actor = ActorNetwork(lr_actor,
state_size,
action_size,
random_seed,
name="actor")
self.actor_target = ActorNetwork(lr_actor,
state_size,
action_size,
random_seed,
name="actor_target")
self.soft_update(self.actor, self.actor_target, tau=1)
# Critic network
self.critic = CriticNetwork(lr_critic,
state_size,
action_size,
random_seed,
name="critic")
self.critic_target = CriticNetwork(lr_critic,
state_size,
action_size,
random_seed,
name="critic_target")
self.soft_update(self.critic, self.critic_target, tau=1)
# Noise process
self.noise = OUActionNoise(mu=np.zeros(action_size))
# Replay buffer memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE,
random_seed)
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self,
state_size,
action_size,
lr_actor=LR_ACTOR,
lr_critic=LR_CRITIC,
random_seed=42,
num_agents=1):
"""Initialize Agent object.
Params
====
state_size (int): Dimension of each state
action_size (int): Dimension of each action
lr_actor (float): Learning rate for actor model
lr_critic (float): Learning Rate for critic model
random_seed (int): Random seed
num_agents (int): Number of agents
return
====
None
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.num_agents = num_agents
# Initialize time step (for updating every hyperparameters["update_every"] steps)
self.t_step = 0
# Actor network
self.actor = ActorNetwork(lr_actor,
state_size,
action_size,
random_seed,
name="actor")
self.actor_target = ActorNetwork(lr_actor,
state_size,
action_size,
random_seed,
name="actor_target")
self.soft_update(self.actor, self.actor_target, tau=1)
# Critic network
self.critic = CriticNetwork(lr_critic,
state_size,
action_size,
random_seed,
name="critic")
self.critic_target = CriticNetwork(lr_critic,
state_size,
action_size,
random_seed,
name="critic_target")
self.soft_update(self.critic, self.critic_target, tau=1)
# Noise process
self.noise = OUActionNoise(mu=np.zeros(action_size))
# Replay buffer memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE,
random_seed)
def step(self, states, actions, rewards, next_states, dones):
"""Save experience in replay memory, and use random sample from buffer to learn."""
# Save experience / reward
# Support for multi agents learners
for state, action, reward, next_state, done in zip(
states, actions, rewards, next_states, dones):
self.memory.add(state, action, reward, next_state, done)
# Update timestep to learn
self.t_step = (self.t_step + 1) % UPDATE_EVERY
# Learn, if enough samples are available in memory
if len(self.memory) > BATCH_SIZE and self.t_step == 0:
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state, add_noise=True):
"""Returns actions for given state as per current policy."""
states = T.from_numpy(state).float().to(device)
self.actor.eval()
with T.no_grad():
actions = self.actor(states).cpu().data.numpy()
self.actor.train()
if add_noise:
actions += self.noise.sample()
return np.clip(actions, -1, 1)
def reset(self):
self.noise.reset()
def learn(self, experiences, gamma):
"""Update policy and value parameters using given batch of experience tuples.
Q_targets = r + γ * critic_target(next_state, actor_target(next_state))
where:
actor_target(state) -> action
critic_target(state, action) -> Q-value
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# ---------------------------- update critic ---------------------------- #
# Get predicted next-state actions and Q values from target models
actions_next = self.actor_target(next_states)
Q_targets_next = self.critic_target(next_states, actions_next)
# Compute Q targets for current states (y_i)
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Compute critic loss
Q_expected = self.critic(states, actions)
critic_loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
self.critic.optimizer.zero_grad()
critic_loss.backward()
T.nn.utils.clip_grad_norm_(self.critic.parameters(), 1.0)
self.critic.optimizer.step()
# ---------------------------- update actor ---------------------------- #
# Compute actor loss
actions_pred = self.actor(states)
actor_loss = -self.critic(states, actions_pred).mean()
# Minimize the loss
self.actor.optimizer.zero_grad()
actor_loss.backward()
self.actor.optimizer.step()
# ----------------------- update target networks ----------------------- #
self.soft_update(self.critic, self.critic_target, TAU)
self.soft_update(self.actor, self.actor_target, TAU)
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model: PyTorch model (weights will be copied from)
target_model: PyTorch model (weights will be copied to)
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(tau * local_param.data +
(1.0 - tau) * target_param.data)
def save_models(self):
""" Save models weights """
self.actor.save_checkpoint()
self.critic.save_checkpoint()
self.actor_target.save_checkpoint()
self.critic_target.save_checkpoint()
def load_models(self):
""" Load models weights """
self.actor.load_checkpoint()
self.critic.load_checkpoint()
self.actor_target.load_checkpoint()
self.critic_target.load_checkpoint()
class Agent:
def __init__(self, input_size, output_size, hidden = 256, lr_actor=1.0e-3, lr_critic=1.0e-3, agent_number=0, tau=1.0e-2,
gamma=0.99, epsilon=1.0, epsilon_decay=0.99, weight_decay=0, clipgrad=.1, seed = 42):
super(Agent, self).__init__()
self.seed = seed
self.actor = ActorNetwork(input_size, output_size, name=f"Actor_Agent{agent_number}").to(device)
self.critic = CriticNetwork(input_size, output_size, name=f"Critic_Agent{agent_number}").to(device)
self.target_actor = ActorNetwork(input_size, output_size, name=f"Actor_Target_Agent{agent_number}").to(device)
self.target_critic = CriticNetwork(input_size, output_size, name=f"Critic_Target_Agent{agent_number}").to(device)
self.noise = OUActionNoise(mu=np.zeros(output_size))
self.tau = tau
self.epsilon = epsilon
self.epsilon_decay=epsilon_decay
self.gamma = gamma
self.clipgrad = clipgrad
self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=lr_actor)
self.critic_optimizer = optim.Adam(self.critic.parameters(), lr=lr_critic, weight_decay=weight_decay)
def act(self, state, add_noise=True):
"""Returns actions for given state as per current policy."""
state = torch.from_numpy(state).float().unsqueeze(0).to(device) #.unsqueeze(0)
self.actor.eval()
with torch.no_grad():
action = self.actor(state).cpu().squeeze(0).data.numpy()
self.actor.train()
if add_noise:
action += self.noise.sample() * self.epsilon
return np.clip(action, -1, 1)
def learn(self, experiences):
"""Update policy and value parameters using given batch of experience tuples.
Q_targets = r + γ * critic_target(next_state, actor_target(next_state))
where:
actor_target(state) -> action
critic_target(state, action) -> Q-value
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# ---------------------------- update critic ---------------------------- #
# Get predicted next-state actions and Q values from target models
actions_next = self.target_actor(next_states.to(device))
#set_trace()
Q_targets_next = self.target_critic(next_states.to(device), actions_next.to(device))
# Compute Q targets for current states (y_i)
Q_targets = rewards + (self.gamma * Q_targets_next * (1 - dones))
# Compute critic loss
Q_expected = self.critic(states, actions)
critic_loss = f.mse_loss(Q_expected, Q_targets)
# Minimize the loss
self.critic_optimizer.zero_grad()
critic_loss.backward()
clip_grad_norm_(self.critic.parameters(), self.clipgrad)
self.critic_optimizer.step()
# update actor
# Compute actor loss
actions_pred = self.actor(states)
actor_loss = -self.critic(states, actions_pred).mean()
# Minimize the loss
self.actor_optimizer.zero_grad()
actor_loss.backward()
#clip_grad_norm_(self.actor.parameters(), self.clipgrad)
self.actor_optimizer.step()
# update target networks
self.soft_update(self.critic, self.target_critic )
self.soft_update(self.actor, self.target_actor)
# update epsilon and noise
self.epsilon *= self.epsilon_decay
self.noise.reset()
def reset(self):
self.noise.reset()
def soft_update(self, local_model, target_model):
"""Soft update model parameters.
θ_target = τ*θ_local + (1 - τ)*θ_target
Params
======
local_model: PyTorch model (weights will be copied from)
target_model: PyTorch model (weights will be copied to)
tau (float): interpolation parameter
"""
for target_param, local_param in zip(target_model.parameters(), local_model.parameters()):
target_param.data.copy_(self.tau*local_param.data + (1.0-self.tau)*target_param.data)
def save_models(self):
""" Save models weights """
self.actor.save_checkpoint()
self.critic.save_checkpoint()
self.target_actor.save_checkpoint()
self.target_critic.save_checkpoint()
def load_models(self):
""" Load models weights """
self.actor.load_checkpoint()
self.critic.load_checkpoint()
self.target_actor.load_checkpoint()
self.target_critic.load_checkpoint()