class DQNAgent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, seed):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
# Q-Network
self.qnetwork_local = QNetwork(state_size, action_size,
seed).to(device)
self.qnetwork_target = QNetwork(state_size, action_size,
seed).to(device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE, seed)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
def step(self, state, action, reward, next_state, done):
# Save experience in replay memory
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, eps=0.):
"""Returns actions for given state as per current policy.
Params
======
state (array_like): current state
eps (float): epsilon, for epsilon-greedy action selection
"""
state = torch.from_numpy(state).float().unsqueeze(0).to(device)
self.qnetwork_local.eval()
with torch.no_grad():
action_values = self.qnetwork_local(state)
self.qnetwork_local.train()
# Epsilon-greedy action selection
if random.random() > eps:
return np.argmax(action_values.cpu().data.numpy())
else:
return random.choice(np.arange(self.action_size))
def learn(self, experiences, gamma):
"""Update value parameters using given batch of experience tuples.
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# Get max predicted Q values (for next states) from target model
Q_targets_next = self.qnetwork_target(next_states).detach().max(
1)[0].unsqueeze(1)
# Compute Q targets for current states
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Get expected Q values from local model
Q_expected = self.qnetwork_local(states).gather(1, actions)
# Compute loss
loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# ------------------- update target network ------------------- #
self.soft_update(self.qnetwork_local, self.qnetwork_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():
"""Interacts with and learns from the environment."""
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
def __init__(self,
state_size,
action_size,
seed,
sample_method='uniform',
method='doubledqn',
device=None,
**kwargs):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(seed)
self.device = Agent.device if device is None else device
# Network
self.qnetwork_local = QNetwork(state_size, action_size,
seed).to(self.device)
self.qnetwork_target = QNetwork(state_size, action_size,
seed).to(self.device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
self.dqnmethod = method
self.sample_method = sample_method
# Replay memory
if sample_method == 'minority_resampled':
self.memory = MinorityResampledReplayBuffer(action_size,
BUFFER_SIZE,
BATCH_SIZE,
seed,
device=self.device)
elif sample_method == 'prioritized':
if 'alpha' in kwargs:
alpha = kwargs['alpha']
else:
alpha = None
if 'beta0' in kwargs:
beta = kwargs['beta0']
else:
beta = None
self.memory = PrioritizedReplayBuffer(action_size,
BUFFER_SIZE,
BATCH_SIZE,
seed,
device=self.device,
alpha=alpha,
beta0=beta)
elif sample_method == 'uniform':
self.memory = ReplayBuffer(action_size,
BUFFER_SIZE,
BATCH_SIZE,
seed,
device=self.device)
else:
raise Exception('Unrecognized sampling method')
# Initialize time step (for updating every UPDATE_EVERY steps)
self.update_step = 0
self.replay_step = 0
self.episode_count = 0
def step(self, state, action, reward, next_state, done):
# Save experience in replay memory
self.memory.add_new_experience(state, action, reward, next_state, done)
self.update_step = (self.update_step +
1) % UPDATE_EVERY # This is checked in learn
self.replay_step = (self.replay_step + 1) % REPLAY_EVERY
self.episode_count = self.episode_count + 1 if done else self.episode_count
if len(self.memory) > BATCH_SIZE:
if self.sample_method == 'prioritized':
if self.replay_step == 0:
self.learn(GAMMA, method=self.dqnmethod)
else:
self.learn(GAMMA, method=self.dqnmethod)
def act(self, state, eps=0.):
"""Returns actions for given state as per current policy.
Params
======
state (array_like): current state
eps (float): epsilon, for epsilon-greedy action selection
"""
state = torch.from_numpy(state).float().unsqueeze(0).to(self.device)
self.qnetwork_local.eval()
with torch.no_grad():
action_values = self.qnetwork_local(state)
self.qnetwork_local.train()
# Epsilon-greedy action selection
if random.random() > eps:
return np.argmax(action_values.cpu().data.numpy())
else:
return random.choice(np.arange(self.action_size))
def dqn(self, rewards, next_states, dones, gamma):
y = torch.zeros_like(rewards)
# For end of episode, the return is just the final reward
y[(dones == 1).squeeze(), ...] = rewards[(dones == 1).squeeze(), ...]
# Compute the aproximation of the optimal target reward values (Q*)
with torch.no_grad():
logits = self.qnetwork_target(next_states[(dones == 0).squeeze(),
...])
next_values, _ = torch.max(logits, 1) # Values of next max actions
y[(dones == 0).squeeze(),
...] = rewards[(dones == 0).squeeze(),
...] + gamma * next_values.unsqueeze(-1)
return y
def doubledqn(self, rewards, next_states, dones, gamma):
y = torch.zeros_like(rewards)
# For end of episode, the return is just the final reward
y[(dones == 1).squeeze(), ...] = rewards[(dones == 1).squeeze(), ...]
# Compute the aproximation of the optimal target reward values (Q*)
with torch.no_grad():
# 1 - Get the local net next action
next_local_logits = self.qnetwork_local(
next_states[(dones == 0).squeeze(), ...])
_, max_next_local_act = torch.max(next_local_logits,
1) # Values of next max actions
# 2 - Get target network's value of the local's max next action
next_target_logits = self.qnetwork_target(
next_states[(dones == 0).squeeze(), ...])
values = next_target_logits.gather(
1, max_next_local_act.unsqueeze(-1))
# 3 - Obtain the approximation of
y[(dones == 0).squeeze(),
...] = rewards[(dones == 0).squeeze(), ...] + gamma * values
return y
def learn(self, gamma, method='doubledqn'):
"""Update value parameters using given batch of experience tuples.
Params
======
experiences (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones, idc, weights = self.memory.sample(
self.episode_count)
dones = torch.round(dones).int()
if method == 'dqn':
y = self.dqn(rewards=rewards,
next_states=next_states,
dones=dones,
gamma=gamma)
elif method == 'doubledqn':
y = self.doubledqn(rewards=rewards,
next_states=next_states,
dones=dones,
gamma=gamma)
else:
raise Exception('Unrecognized method')
## TODO: compute and minimize the loss
# GT: We train the local network,
# and update the target network parameters
# zero the parameter gradients
self.qnetwork_local.train()
# 1 - Clear out gradients from the local network: perform detach() and zero_() on network parameters
self.optimizer.zero_grad()
# 2 - Local estimation of action values
local_q = self.qnetwork_local(states)
# 3 - Loss between the approximation of optimal target reward values (Q*) and local estimates
local_q = local_q.gather(1, actions) # Prior expected returns
# Temporal Difference (TD) error
td_error = y - local_q
# Update the priorities
self.memory.update_priorities(
np.abs(td_error.data.clone().cpu().numpy()) + 1.0e-5, idc)
if self.sample_method == 'prioritized':
local_q.backward(-weights * td_error)
else:
loss = torch.nn.MSELoss(reduce=False)(local_q, y)
# 4 - Gradient descend on local network
loss.backward(weights)
# 5 - Gradient update
self.optimizer.step()
# ------------------- update target network ------------------- #
if self.update_step == 0:
# 6 - Set local network to eval
# self.qnetwork_local.eval() ... this messes things up !!
self.soft_update(TAU)
def soft_update(self, 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(self.qnetwork_target.parameters(),
self.qnetwork_local.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, 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):
"""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) > BATCH_SIZE:
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:
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()
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)