class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, random_seed):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
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
self.noise = OUNoise(action_size, random_seed)
# Replay memory
self.memory = ReplayBuffer(
action_size, BUFFER_SIZE, BATCH_SIZE, random_seed)
def step(self, state, action, reward, next_state, done, timestep):
"""Save experience in replay memory, and use random sample from buffer to learn."""
# Save experience / reward
for i in range(20):
self.memory.add(state[i], action[i],
reward[i], next_state[i], done[i])
# Learn, if enough samples are available in memory
if timestep % 20 == 0:
if len(self.memory) > BATCH_SIZE:
for i in range(10):
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state, 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:
for i in range(20):
action[i] += 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()
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)
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)
class DDPGAgent:
"""Interacts with and learns from the environment."""
def __init__(self,
state_size,
action_size,
n_agents,
seed,
buffer_size=int(1e6),
batch_size=200,
lr_actor=1e-4,
lr_critic=1e-3,
gamma=0.99,
weight_decay=0,
tau=1e-3,
update_frequency=20,
n_learns=10):
"""Initialize a DDPG agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
n_agents (int): number of agents
random_seed (int): random seed
batch_size (int): minibatch size
lr_actor (float): learning rate of the actor
lr_critic (float): learning rate of the critic
gamma (float): discount factor
weight_decay (float): critic L2 weight decay
tau (float): value for soft update of target parameters
update_frequency (int): how much steps must be executed before starting learning
n_learns (int): how many learning for update
"""
self.state_size = state_size
self.action_size = action_size
self.n_agents = n_agents
self.gamma = gamma
self.batch_size = batch_size
self.tau = tau
self.seed = random.seed(seed)
self.update_frequency = update_frequency
self.n_learns = n_learns
# Actor Network (w/ Target Network)
self.actor_local = Actor(state_size, action_size, seed).to(device)
self.actor_target = Actor(state_size, action_size, 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, seed).to(device)
self.critic_target = Critic(state_size, action_size, seed).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=lr_critic,
weight_decay=weight_decay)
# Noise process
self.noise = Ornstein((n_agents, action_size), seed)
# Replay memory
self.memory = ReplayBuffer(action_size, buffer_size, seed, device)
# Initialize the time step (for every update_frequency steps)
self.t_step = 0
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 state, action, reward, next_state, done in zip(
states, actions, rewards, next_states, dones):
self.memory.add(state, action, reward, next_state, done)
self.t_step = (self.t_step + 1) % self.update_frequency
if self.t_step == 0:
# Learn, if enough samples are available in memory
for _ in range(self.n_learns):
if len(self.memory) > self.batch_size:
experiences = self.memory.sample(self.batch_size)
self.learn(experiences, self.gamma)
def act(self, states, add_noise=True):
"""Returns actions for given state as per current policy."""
states = torch.from_numpy(states).float().to(device)
self.actor_local.eval()
with torch.no_grad():
action = self.actor_local(states).cpu().data.numpy()
self.actor_local.train()
if add_noise:
action += 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()
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, self.tau)
self.soft_update(self.actor_local, self.actor_target, self.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)
class DDPG:
def __init__(self,
dim_obs,
dim_act,
actor_lr=0.001,
critic_lr=0.01,
gamma=0.9,
capacity=1000,
batch_size=64,
tau=0.01,
hidden_size=64):
self.gamma = gamma
self.memory = ReplayMemory(capacity)
self.batch_size = batch_size
self.tau = tau
self.device = 'gpu' if GPU_CONFIG.use_cuda else 'cpu'
self.learn_cnt = 0
self.FloatTensor = th.cuda.FloatTensor if GPU_CONFIG.use_cuda else th.FloatTensor
self.critic = Critic(dim_obs, dim_act, hidden_size).to(self.device)
self.actor = Actor(dim_obs, dim_act, hidden_size).to(self.device)
self.target_critic = Critic(dim_obs, dim_act,
hidden_size).to(self.device)
self.target_actor = Actor(dim_obs, dim_act,
hidden_size).to(self.device)
self.target_critic.load_state_dict(self.critic.state_dict())
self.target_actor.load_state_dict(self.actor.state_dict())
# for target_param, param in zip(self.target_actor.parameters(), self.actor.parameters()):
# target_param.data.copy_(param.data) # 方式二: todo 试试
self.critic_optimizer = th.optim.Adam(self.critic.parameters(),
lr=critic_lr)
self.actor_optimizer = th.optim.Adam(self.actor.parameters(),
lr=actor_lr)
def learn(self):
# sample batch from all memory
transitions = self.memory.sample(self.batch_size)
batch = Experience(*zip(*transitions)) # class(list)
obs_batch = self.FloatTensor(np.array(
batch.obs)) # torch.Size([32, 4])
# obs_batch = th.tensor(obs_batch, device=self.device, dtype=th.float) # 方法二
logger.debug("obs_batch: {}".format(obs_batch.shape))
action_batch = self.FloatTensor(np.array(batch.action)) # (batch, 1)
logger.debug('action batch: {}'.format(action_batch.shape))
reward_batch = self.FloatTensor(np.array(batch.reward)).view(
self.batch_size, 1) # (batch, 1)
logger.debug('reward_batch: {}'.format(reward_batch.shape))
next_obs_batch = self.FloatTensor(np.array(batch.next_obs))
done_batch = self.FloatTensor(np.array(batch.done)).view(
self.batch_size, 1) # (batch, 1)
logger.debug('done_batch: {}'.format(done_batch))
# c loss
self.critic_optimizer.zero_grad()
q_eval = self.critic(obs_batch, action_batch)
next_action = self.target_actor(next_obs_batch).detach()
q_next = self.target_critic(next_obs_batch, next_action).detach()
q_target = reward_batch + self.gamma * q_next * (1 - done_batch)
c_loss = nn.MSELoss()(q_eval, q_target)
c_loss.backward()
self.critic_optimizer.step()
# a loss
self.actor_optimizer.zero_grad()
current_action = self.actor(obs_batch)
policy_loss = self.critic(obs_batch, current_action)
a_loss = -policy_loss.mean()
a_loss.backward()
self.actor_optimizer.step()
# soft update actor_target, critic_target
soft_update(self.target_critic, self.critic, self.tau)
soft_update(self.target_actor, self.actor, self.tau)
a_loss = a_loss.detach().cpu().numpy(
) if GPU_CONFIG.use_cuda else a_loss.detach().numpy()
c_loss = c_loss.detach().cpu().numpy(
) if GPU_CONFIG.use_cuda else c_loss.detach().numpy()
return a_loss, c_loss
@th.no_grad()
def select_action(self, obs):
obs = self.FloatTensor(obs).unsqueeze(0)
action = self.actor(obs).detach()
action = action.cpu().numpy() if GPU_CONFIG.use_cuda else action.numpy(
)
return action
class Agent:
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, random_seed):
"""
Initialize the agent
:param state_size: state space size
:param action_size: action space size
:param random_seed: seed
"""
self.state_size = state_size
self.action_size = action_size
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)
self.hard_copy_weights(self.actor_target, self.actor_local)
self.hard_copy_weights(self.critic_target, self.critic_local)
# Noise process
self.noise = OUNoise(action_size, random_seed)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE,
random_seed)
# Iteration counter
self.step_counter = 0
@staticmethod
def hard_copy_weights(target, source):
""" copy weights from source to target network"""
for target_param, param in zip(target.parameters(),
source.parameters()):
target_param.data.copy_(param.data)
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 state, action, reward, next_state, done in zip(
states, actions, rewards, next_states, dones):
self.memory.add(state, action, reward, next_state, done)
self.step_counter += 1
# Learn, if enough samples are available in memory
if self.step_counter % UPDATE_EVERY == 0:
if len(self.memory) > BATCH_SIZE:
for i in range(0, UPDATE_TIMES):
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
self.step_counter = 0
def act(self, states, add_noise=False):
"""Returns actions for given state as per current policy."""
states = torch.from_numpy(states).float().to(device)
self.actor_local.eval()
with torch.no_grad():
actions = self.actor_local(states).cpu().data.numpy()
self.actor_local.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
:param experiences: tensor of (s,a,r,s') tuples
:param gamma: discount factor
:return:
"""
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()
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.critic_local, self.critic_target, TAU)
self.soft_update(self.actor_local, 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
:param local_model: local model where weights are copied from
:param target_model: target model where weights are copied to
:param tau: soft update rate
:return:
"""
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)
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, device, memory, config):
"""Initialize an Agent object.
Params
======
device (object): hardware device to run on CPU or GPU
memory (object): memory for replay buffer
config (dict)
- "state_size": dimension of each state
- "action_size": dimension of each action
- "buffer_size": replay buffer size
- "batch_size": minibatch size
- "random_seed": random seed
- "gamma": discount factor
- "tau": for soft update of target parameters
- "weight_decay": L2 weight decay
- "learn_every": learn from replay buffer every time step
- "learn_batch_size": number of batches to learn from replay buffer every learn_every time step
- "grad_clip": gradient value to clip at for critic
- "eps_start": starting value of epsilon, for epsilon-greedy action selection
- "eps_end": minimum value of epsilon
- "eps_decay": multiplicative factor (per episode) for decreasing epsilon
- "print_every": Print average every x episode,
- "episode_steps": Maximum number of steps to run for each episode
- "mu": mu for noise
- "theta": theta for noise
- "sigma": sigma for noise
- "actor": actor specific config object
- "fc": array of input sizes for hidden layers
- "learning_rate": learning rate
- "critic": actor specific config object
- "fc": array of input sizes for hidden layers
- "learning_rate": learning rate
"""
self.num_agents = config['num_agents']
self.state_size = config['state_size']
self.action_size = config['action_size']
if config['random_seed'] is not None:
self.seed = random.seed(config['random_seed'])
else:
self.seed = random.seed()
self.eps = config['eps_start']
self.eps_decay = config['eps_decay']
self.eps_end = config['eps_end']
self.device = device
# Replay memory
self.memory = memory
self.batch_size = config['batch_size']
self.gamma = config['gamma']
self.tau = config['tau']
self.lr_actor = config['actor']['learning_rate']
self.lr_critic = config['critic']['learning_rate']
self.weight_decay = config['weight_decay']
self.learn_every = config['learn_every']
self.learn_batch_size = config['learn_batch_size']
self.grad_clip = config['grad_clip']
# Actor Network (w/ Target Network)
self.actor_local = Actor(config).to(self.device)
self.actor_target = Actor(config).to(self.device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=self.lr_actor)
# Critic Network (w/ Target Network)
self.critic_local = Critic(config).to(self.device)
self.critic_target = Critic(config).to(self.device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=self.lr_critic,
weight_decay=self.weight_decay)
# Noise process
self.noise = OUNoise(config)
def step(self, states, actions, rewards, next_states, dones, timestep):
"""Save experience in replay memory, and use random sample from buffer to learn."""
# Save experience / reward
for i in range(self.num_agents):
self.memory.add(states[i], actions[i], rewards[i], next_states[i],
dones[i])
# Learn, if enough samples are available in memory and every update_every time steps
if len(self.memory
) > self.batch_size and timestep % self.learn_every == 0:
for i in range(self.learn_batch_size):
experiences = self.memory.sample()
self.learn(experiences)
def act(self, state, add_noise=True):
"""Returns actions for given state as per current policy."""
state = torch.from_numpy(state).float().to(self.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 += self.eps * self.noise.sample()
return np.clip(action, -1, 1)
def reset(self):
self.noise.reset()
def reset_episode(self):
self.reset()
self.memory.reset_episode()
def learn_best_episode(self):
self.learn(self.memory.sample(True))
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 + (self.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()
# gradient clipping for critic
if self.grad_clip > 0:
torch.nn.utils.clip_grad_norm_(self.critic_local.parameters(),
self.grad_clip)
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)
self.soft_update(self.actor_local, self.actor_target)
if self.eps_decay > 0:
self.eps = max(self.eps_end,
self.eps - self.eps_decay) # decrease epsilon
self.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)
"""
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)