class DDPGAgent():
def __init__(self, state_size, action_size, par):
self.par = par
# Actor Network (w/ Target Network)
self.actor_local = Actor(state_size, action_size, par).to(device)
self.actor_target = Actor(state_size, action_size, par).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=par.lr_actor)
print('actor')
print(self.actor_local)
# Critic Network (w/ Target Network)
self.critic_local = Critic(state_size, action_size, par).to(device)
self.critic_target = Critic(state_size, action_size, par).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=par.lr_critic,
weight_decay=par.weight_decay)
# Noise process
self.noise = OUNoise(action_size, par.random_seed, par.ou_mu,
par.ou_theta, par.ou_sigma)
def save_model(self, experiment_name, i_episode):
path = self.par.save_path
torch.save(
self.actor_local.state_dict(),
experiment_name + '_checkpoint_actor_' + str(i_episode) + '.pth')
torch.save(
self.critic_local.state_dict(),
experiment_name + '_checkpoint_critic_' + str(i_episode) + '.pth')
class Agent(AgentABC):
def __init__(self, state_size, action_size, num_agents, random_seed):
"""
Initialize an DDPG Agent object.
:param state_size (int): dimension of each state
:param action_size (int): dimension of each action
:param num_agents (int): number of agents in environment ot use ddpg
:param random_seed (int): random seed
"""
super().__init__(state_size, action_size, num_agents, random_seed)
self.state_size = state_size
self.action_size = action_size
self.num_agents = num_agents
self.seed = random.seed(random_seed)
# Actor Network (w/ Target Network)
self.actor_local = Actor(state_size, action_size,
random_seed).to(device)
self.actor_target = Actor(state_size, action_size,
random_seed).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=LR_ACTOR)
# Critic Network (w/ Target Network)
self.critic_local = Critic(state_size, action_size,
random_seed).to(device)
self.critic_target = Critic(state_size, action_size,
random_seed).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
# Noise process for each agent
self.noise = OUNoise((num_agents, action_size), random_seed)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE,
random_seed)
# debug of the MSE critic loss
self.mse_error_list = []
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
for agent in range(self.num_agents):
self.memory.add(states[agent, :], actions[agent, :],
rewards[agent], next_states[agent, :],
dones[agent])
# Learn, if enough samples are available in memory
if len(self.memory) > BATCH_SIZE:
experiences = self.memory.sample()
self.learn(experiences)
self.debug_loss = np.mean(self.mse_error_list)
def act(self, state, add_noise=True):
"""Returns actions for given state as per current policy."""
state = torch.from_numpy(state).float().to(device)
acts = np.zeros((self.num_agents, self.action_size))
self.actor_local.eval()
with torch.no_grad():
for agent in range(self.num_agents):
acts[agent, :] = self.actor_local(
state[agent, :]).cpu().data.numpy()
self.actor_local.train()
if add_noise:
noise = self.noise.sample()
acts += noise
return np.clip(acts, -1, 1)
def reset(self):
""" see abstract class """
super().reset()
self.noise.reset()
self.mse_error_list = []
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.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.view(BATCH_SIZE,
-1) + (GAMMA * Q_targets_next *
(1 - dones).view(BATCH_SIZE, -1))
# Compute critic loss
Q_expected = self.critic_local(states, actions)
critic_loss = F.mse_loss(Q_expected, Q_targets)
self.mse_error_list.append(critic_loss.detach().cpu().numpy())
# Minimize the loss
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
# ---------------------------- update actor ---------------------------- #
# Compute actor loss
actions_pred = self.actor_local(states)
actor_loss = -self.critic_local(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_local, self.critic_target, TAU)
self.soft_update(self.actor_local, self.actor_target, TAU)
@staticmethod
def soft_update(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 load_weights(self, directory_path):
""" see abstract class """
super().load_weights(directory_path)
self.actor_target.load_state_dict(
torch.load(os.path.join(directory_path, an_filename),
map_location=device))
self.critic_target.load_state_dict(
torch.load(os.path.join(directory_path, cn_filename),
map_location=device))
self.actor_local.load_state_dict(
torch.load(os.path.join(directory_path, an_filename),
map_location=device))
self.critic_local.load_state_dict(
torch.load(os.path.join(directory_path, cn_filename),
map_location=device))
def save_weights(self, directory_path):
""" see abstract class """
super().save_weights(directory_path)
torch.save(self.actor_local.state_dict(),
os.path.join(directory_path, an_filename))
torch.save(self.critic_local.state_dict(),
os.path.join(directory_path, cn_filename))
def save_mem(self, directory_path):
""" see abstract class """
super().save_mem(directory_path)
self.memory.save(os.path.join(directory_path, "ddpg_memory"))
def load_mem(self, directory_path):
""" see abstract class """
super().load_mem(directory_path)
self.memory.load(os.path.join(directory_path, "ddpg_memory"))
class MADDPGAgent:
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, num_agents, random_seed):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
num_agents (int): number of agents
random_seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.num_agents = num_agents
self.seed = random.seed(random_seed)
# Actor Networks (Local and Target)
self.actor_local = Actor(state_size, action_size,
random_seed).to(device)
self.actor_target = Actor(state_size, action_size,
random_seed).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=LR_ACTOR)
# Critic Networks (Local and Target)
self.critic_local = Critic(state_size, action_size,
random_seed).to(device)
self.critic_target = Critic(state_size, action_size,
random_seed).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
# Noise Process
self.noise = OUNoise(action_size, random_seed)
# Replay memory
self.memory = ReplayBuffer(action_size, REPLAY_BUFFER_SIZE,
MINIBATCH_SIZE, random_seed)
# Count t_steps
self.time_step = 0
def step(self, state, action, reward, next_state, done, time_step):
"""Save experience in replay memory, and use random sample from buffer to learn."""
# Save experience / reward
self.memory.add(state, action, reward, next_state, done)
# Learn, if enough samples are available in memory
if len(self.memory) > MINIBATCH_SIZE and time_step % UPDATE_EVERY == 0:
for _ in range(LEARN_NUM):
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state, eps=0., add_noise=True):
"""Returns actions for given state as per current policy."""
state = torch.from_numpy(state).float().to(device)
self.actor_local.eval()
with torch.no_grad():
action = self.actor_local(state).cpu().data.numpy()
self.actor_local.train()
if add_noise:
action += eps * self.noise.sample()
return np.clip(action, -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_local(states, actions)
critic_loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
self.critic_optimizer.zero_grad()
critic_loss.backward()
# clipping gradient to 1 for stable learning
torch.nn.utils.clip_grad_norm_(self.critic_local.parameters(), 1)
self.critic_optimizer.step()
# ---------------------------- update actor ---------------------------- #
# Compute actor loss
actions_pred = self.actor_local(states)
actor_loss = -self.critic_local(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.actor_local, self.actor_target, TAU)
self.soft_update(self.critic_local, self.critic_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 create_agent(state_size,
action_size,
seed=0,
actor_fc1_units=400,
actor_fc2_units=300,
actor_lr=1e-4,
critic_fc1_units=400,
critic_fc2_units=300,
critic_lr=1e-4,
weight_decay=0,
buffer_size=int(1e5),
batch_size=128,
gamma=0.99,
tau=0.1,
noise_dev=0.3):
"""
This function creates an agent with specified parameters for training.
Arguments:
state_size: An integer count of dimensions for each state.
action_size: An integer count of dimensions for each action.
seed: Random seed specified to diversify results.
actor_fc1_units: An integer number of units used in the first FC
layer for the Actor object.
actor_fc2_units: An integer number of units used in the second FC
layer for the Actor object.
actor_lr: A float designating the learning rate of the Actor's
optimizer.
critic_fc1_units: An integer number of units used in the first FC
layer for the Critic object.
critic_fc2_units: An integer number of units used in the second FC
layer for the Critic object.
critic_lr: A float designating the learning rate of the Critic's
optimizer.
weight_decay: Float multiplicative factor to stabilize complexity
penalization.
buffer_size: An integer for replay buffer size.
batch_size: An integer for minibatch size.
gamma: A float designating the discount factor.
tau: A float designating multiplication factor for soft update of
target parameters.
noise_dev: Float designating the noise to be added to action decisions.
Returns:
agent: An Agent object used for training.
"""
# Initialize the replay buffer from which experiences are gathered for
# training the agent.
replay_buffer = ReplayBuffer(seed=seed,
buffer_size=buffer_size,
batch_size=batch_size)
# Initialize local and target Actor Networks and optimizer.
actor_local = Actor(state_size, action_size, seed, actor_fc1_units,
actor_fc2_units).to(device)
actor_target = Actor(state_size, action_size, seed, actor_fc1_units,
actor_fc2_units).to(device)
actor_optimizer = optim.Adam(actor_local.parameters(), lr=actor_lr)
# Initialize local and target Critic Networks and optimizer.
critic_local = Critic(state_size, action_size, seed, critic_fc1_units,
critic_fc2_units).to(device)
critic_target = Critic(state_size, action_size, seed, critic_fc1_units,
critic_fc2_units).to(device)
critic_optimizer = optim.Adam(critic_local.parameters(),
lr=critic_lr,
weight_decay=weight_decay)
# Initialize Gaussian noise to reduce generalization error.
noise = GaussianNoise(action_size, seed, mu=0.0, sigma=noise_dev)
# Create agent object used for training.
agent = Agent(seed=seed,
memory=replay_buffer,
batch_size=batch_size,
actor_local=actor_local,
actor_target=actor_target,
actor_optimizer=actor_optimizer,
critic_local=critic_local,
critic_target=critic_target,
critic_optimizer=critic_optimizer,
noise=noise,
gamma=gamma,
tau=tau)
return agent