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(action_size, state_size, (512, 256),
random_seed).to(device)
self.actor_target = Actor(action_size, state_size, (512, 256),
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(action_size, state_size, (512, 256),
random_seed).to(device)
self.critic_target = Critic(action_size, state_size, (512, 256),
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)
self.t_step = 0
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE,
random_seed, device)
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)
# Learn every UPDATE_EVERY time steps.
self.t_step = (self.t_step + 1) % UPDATE_EVERY
if self.t_step == 0:
# If enough samples are available in memory, get random subset and learn
if len(self.memory) > BATCH_SIZE:
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state, add_noise=True):
states = torch.from_numpy(state).float().to(device)
self.actor_local.eval()
with torch.no_grad():
actions = self.actor_local(states).cpu().data.numpy()
self.actor_local.train()
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
#print("states shape",states.shape)
# ---------------------------- 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
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 Agent():
def __init__(self, state_size, action_size, n_agents=1, seed=0):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
n_agents: number of agents it will control in the environment
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = np.random.seed(seed)
random.seed(seed)
self.n_agents = n_agents
# Actor Network (w/ Target Network)
self.actor_local = Actor(state_size,
action_size,
leak=LEAKINESS,
seed=seed).to(device)
self.actor_target = Actor(state_size,
action_size,
leak=LEAKINESS,
seed=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,
leak=LEAKINESS,
seed=seed).to(device)
self.critic_target = Critic(state_size,
action_size,
leak=LEAKINESS,
seed=seed).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=LR_CRITIC)
# self.critic_optimizer = optim.Adam(self.critic_local.parameters(), lr=LR_CRITIC, weight_decay=WEIGHT_DECAY)
# Noise process
self.noise = OUNoise(action_size, seed)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed)
self.timesteps = 0
def step(self, states, actions, rewards, next_states, dones):
""" Given a batch of S,A,R,S' experiences, it saves them into the
experience buffer, and occasionally samples from the experience
buffer to perform training steps.
"""
self.timesteps += 1
for i in range(self.n_agents):
self.memory.add(states[i], actions[i], rewards[i], next_states[i],
dones[i])
if (len(self.memory) > BATCH_SIZE) and (self.timesteps % 20 == 0):
for _ in range(10):
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, states, add_noise=True):
""" Given a list of states for each agent it returns the actions to be
taken by each agent based on the current policy.
Returns a numpy array of shape [n_agents, n_actions]
NOTE: clips actions to be between -1, 1
Args:
states: () one row of state for each agent [n_agents, n_actions]
add_noise: (bool) add noise to the actions?
"""
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() for _ in range(self.n_agents)]
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_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)
@property
def device(self):
return device
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, random_seed, learn_every):
"""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.seed = random.seed(random_seed)
# Actor Network (w/ Target Network)
self.state_size = state_size
self.action_size = action_size
self.learn_every = learn_every
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):
"""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)
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:
action += self.noise.sample()
return np.clip(action, -1, 1)
def reset(self):
self.noise.reset()
def learn_from_file(self):
try:
file = np.random.choice(glob('data/*.*'))
exps = pickle.load(open(file, 'rb'))
exps = [experience(**exp) for exp in exps]
exps = batch_from_sample(exps)
self.learn(exps, GAMMA)
except EOFError:
pass
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 Agent():
def __init__(self, state_size, action_size, random_seed, memory=None):
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
# Actor 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
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
self.noise = OUNoise(action_size, random_seed)
# Replay memory
if memory is None:
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE,
random_seed)
else:
self.memory = memory
def step(self, states, actions, rewards, next_states, dones, step):
for state, action, reward, next_state, done in zip(
states, actions, rewards, next_states, dones):
self.memory.add(state, action, reward, next_state, done)
if len(self.memory) > BATCH_SIZE and step % LEARN_EVERY == 0:
for _ in range(LEARN_N_TIMES):
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, states, add_noise=True):
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):
states, actions, rewards, next_states, dones = experiences
actions_next = self.actor_target(next_states)
Q_targets_next = self.critic_target(next_states, actions_next)
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
Q_expected = self.critic_local(states, actions)
critic_loss = F.mse_loss(Q_expected, Q_targets)
self.critic_optimizer.zero_grad()
critic_loss.backward()
torch.nn.utils.clip_grad_norm_(self.critic_local.parameters(), 1)
self.critic_optimizer.step()
actions_pred = self.actor_local(states)
actor_loss = -self.critic_local(states, actions_pred).mean()
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
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):
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, args, state_size, action_size):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
"""
self.args = args
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(self.args.seed)
# Actor Network (w/ Target Network)
self.actor_local = Actor(state_size,
action_size,
self.args.seed,
fc1_units=self.args.hidden_1_size,
fc2_units=self.args.hidden_2_size,
fc3_units=self.args.hidden_3_size).to(device)
self.actor_target = Actor(state_size,
action_size,
self.args.seed,
fc1_units=self.args.hidden_1_size,
fc2_units=self.args.hidden_2_size,
fc3_units=self.args.hidden_3_size).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=self.args.lr_Actor,
eps=self.args.adam_eps)
# Critic Network (w/ Target Network)
self.critic_local = Critic(
state_size,
action_size,
self.args.seed,
fc1_units=self.args.hidden_1_size,
fc2_units=self.args.hidden_2_size,
fc3_units=self.args.hidden_3_size).to(device)
self.critic_target = Critic(
state_size,
action_size,
self.args.seed,
fc1_units=self.args.hidden_1_size,
fc2_units=self.args.hidden_2_size,
fc3_units=self.args.hidden_3_size).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=self.args.lr_Critic,
eps=self.args.adam_eps,
weight_decay=self.args.weight_decay)
# Noise process
self.noise = OUNoise(action_size,
self.args.seed,
sigma=self.args.noise_std)
self.learning_starts = int(args.memory_capacity *
args.learning_starts_ratio)
self.learning_frequency = args.learning_frequency
#self.memory = ReplayBuffer(action_size, BUFFER_SIZE, LEARNING_BATCH_SIZE, seed)
def memorize(self, states, actions, rewards, next_states, dones, memory):
"""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):
#memory.add(state, action, reward, done)
memory.add(state, action, reward, next_state, done)
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:
action += self.noise.sample()
return np.clip(action, -1, 1)
def reset(self):
self.noise.reset()
def learn(self, memory, timestep):
"""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
======
"""
# Learn, if enough samples are available in memory and after "learning_frequency" steps since we last learnt
# if memory.num_memories > self.learning_starts and timestep % self.learning_frequency == 0:
# idxs, states, actions, rewards, next_states, dones, _ = memory.sample(self.args.batch_size)
if len(
memory
) > self.learning_starts and timestep % self.learning_frequency == 0:
states, actions, rewards, next_states, dones = memory.sample()
# ---------------------------- 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.args.discount * 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(),
self.args.reward_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()
# memory.update_priorities(idxs, actor_loss.detach()) # Update priorities of sampled transitions
# ----------------------- update target networks ----------------------- #
# if timestep % self.args.target_update == 0:
# Every time there is a leartning process happening, let's update
self.soft_update(self.critic_local, self.critic_target,
self.args.tau)
self.soft_update(self.actor_local, self.actor_target,
self.args.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)