class AgentDQ(AgentAbstract):
"""
Implement Deep Q-Net with Fixed TD-Target computation and Experience Replay
Fixed TD-Target: TD-Error computed on a target (offline) and local (online) network,
where local network weights are copied to target network every `update_every` batches
"""
def __init__(self, state_size, action_size, gamma, hidden_layers, drop_p,
batch_size, learning_rate, soft_upd_param, update_every,
buffer_size, seed):
super(AgentDQ,
self).__init__(state_size, action_size, gamma, hidden_layers,
drop_p, batch_size, learning_rate, soft_upd_param,
update_every, buffer_size, seed)
# Q-Network Architecture
self.qnetwork_local = QNetwork(self.state_size, self.action_size,
self.seed, self.hidden_layers,
self.drop_p).to(device)
self.qnetwork_target = QNetwork(self.state_size, self.action_size,
self.seed, self.hidden_layers,
self.drop_p).to(device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(),
lr=self.learning_rate)
# Experience Replay
self.memory = ReplayBuffer(action_size, buffer_size, batch_size, seed)
def _forward_local(self, states, actions):
"""
Returns
======
ps_local (torch.tensor)
"""
ps_local = self.qnetwork_local.forward(states).gather(1, actions)
return ps_local
def _forward_targets(self, rewards, next_states, dones):
"""
Use Fixed TD-Target Algorithm
Returns
======
ps_target (torch.tensor)
"""
# Fixed Q-Targets
# use target network compute r + g*max(q_est[s',a, w-]), this tensor should be detached in backprop
ps_target = rewards + self.gamma * (1 - dones) * self.qnetwork_target.forward(next_states).detach().\
max(dim=1)[0].view(-1, 1)
return ps_target
class AgentDoubleDQ(AgentAbstract):
"""
Implement Dueling Q-Net with Double QNet (fixed) TD-Target computation and Experience Replay
Double Q-Net: Split action selection and Q evaluation in two steps
"""
def __init__(self, state_size, action_size, gamma, hidden_layers, drop_p,
batch_size, learning_rate, soft_upd_param, update_every,
buffer_size, seed):
super(AgentDoubleDQ,
self).__init__(state_size, action_size, gamma, hidden_layers,
drop_p, batch_size, learning_rate, soft_upd_param,
update_every, buffer_size, seed)
# Q-Network Architecture: Dueling Q-Nets
self.qnetwork_local = QNetwork(self.state_size, self.action_size,
self.seed, self.hidden_layers,
self.drop_p).to(device)
self.qnetwork_target = QNetwork(self.state_size, self.action_size,
self.seed, self.hidden_layers,
self.drop_p).to(device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(),
lr=self.learning_rate)
# Experience Replay
self.memory = ReplayBuffer(action_size, buffer_size, batch_size, seed)
def _forward_local(self, states, actions):
"""
Returns
======
ps_local (torch.tensor)
"""
ps_local = self.qnetwork_local.forward(states).gather(1, actions)
return ps_local
def _forward_targets(self, rewards, next_states, dones):
"""
Use Double Q-Net Algorithm
Returns
======
ps_target (torch.tensor)
"""
ps_actions = self.qnetwork_local.forward(next_states).detach().max(
dim=1)[1].view(-1, 1)
ps_target = rewards + self.gamma * (1 - dones) * self.qnetwork_target.forward(next_states).detach().\
gather(1, ps_actions)
return ps_target
class Agent():
"""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.Variable]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = experiences
# get targets by doing a forward pass of the next states in the target network
self.qnetwork_target.eval()
with torch.no_grad():
Q_targets_next = torch.max(self.qnetwork_target.forward(next_states), dim=1, keepdim=True)[0]
# distinguish the cases in which next states are terminal and those which are not
# for the first case the targets are only the one-step rewards
Q_targets = rewards + (GAMMA * Q_targets_next * (1 - dones))
# get outputs by forward pass of states in the local network
# Note: our qnetwork for a given state all action values for that state.
# However, for each state we know what action to do, so we gather all corresponding action values
self.qnetwork_local.train()
Q_expected = self.qnetwork_local.forward(states).gather(1, actions)
# compute the mean squared error of the Bellman Eq.
loss = F.mse_loss(Q_expected, Q_targets)
# clear gradients buffer from previous iteration
self.optimizer.zero_grad()
# backprop error through local network
loss.backward()
# update weights of local network by taking one SGD step
self.optimizer.step()
# update target network by copying the latest weights of the locat 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 = tau*θ_local + (1 - tau)*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."""
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.q_optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
# Policy Network
self.policy_network_local = PolicyNetwork(state_size, action_size,
seed).to(device)
self.policy_network_target = PolicyNetwork(state_size, action_size,
seed).to(device)
self.policy_optimizer = optim.Adam(
self.policy_network_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
# Action selection
self.noise_scale = START_NOISE_SCALE
def step(self, states, actions, rewards, next_states, dones):
# With multiple arms we need to save each experience separately in the replay
# buffer
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:
for _ in range(20):
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state):
"""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().to(device)
self.qnetwork_local.eval()
self.policy_network_local.eval()
with torch.no_grad():
action = self.policy_network_local(state).cpu().data.numpy()
self.qnetwork_local.train()
self.policy_network_local.train()
# Add noise to the policy that decays to 0 over time to encourage exploration
noise = np.random.normal(loc=0,
scale=self.noise_scale,
size=(1, self.action_size))
action += noise
self.noise_scale *= NOISE_DECAY
return np.clip(action, a_min=-1, a_max=1)
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
# Update the Q-network
argmax_a_next = self.policy_network_target.forward(next_states)
best_next_Q = self.qnetwork_target.forward(next_states, argmax_a_next)
Q_target = rewards + gamma * best_next_Q * (1 - dones)
Q_current = self.qnetwork_local.forward(states, actions)
self.q_optimizer.zero_grad()
criterion = torch.nn.MSELoss()
loss = criterion(Q_current, Q_target.detach())
loss.backward()
torch.nn.utils.clip_grad_norm_(self.qnetwork_local.parameters(), 1)
self.q_optimizer.step()
# Update the policy network
argmax_a = self.policy_network_local.forward(states)
action_values = self.qnetwork_local.forward(states, argmax_a)
self.policy_optimizer.zero_grad()
loss = -action_values.mean(
) # Negative b/c we're doing gradient ascent
loss.backward()
torch.nn.utils.clip_grad_norm_(self.policy_network_local.parameters(),
1)
self.policy_optimizer.step()
# ------------------- update target network ------------------- #
self.soft_update(self.qnetwork_local, self.qnetwork_target, TAU)
self.soft_update(self.policy_network_local, self.policy_network_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)
class Agent:
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, seed, td_target_type="DQN"):
"""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)
assert td_target_type in {"DQN", "Double DQN"}
self.td_target_type = td_target_type
# 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=0.0):
"""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
criterion = torch.nn.MSELoss()
optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
optimizer.zero_grad()
if self.td_target_type == "DQN":
# compute the Q target using the Q-target network
best_next_Q = (
self.qnetwork_target.forward(next_states)
.detach()
.max(1)[0]
.unsqueeze(1)
)
elif self.td_target_type == "Double DQN":
# select best action using current network
best_next_actions = (
self.qnetwork_local.forward(next_states)
.detach()
.max(1)[1]
.reshape(-1, 1)
)
# Use the target network to evaluate the best actions
best_next_Q = (
self.qnetwork_target.forward(next_states)
.detach()
.gather(1, best_next_actions)
)
Q_target = rewards + gamma * best_next_Q * (1 - dones)
Q_current = self.qnetwork_local.forward(states).gather(1, actions)
loss = criterion(Q_current, Q_target)
loss.backward()
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
)