def __init__(self, state_size, action_size, hidden_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.hidden_size = hidden_size
self.seed = random.seed(seed)
# Q-Network
self.qnetwork_local = QNetwork(state_size, action_size, hidden_size,
seed).to(device)
self.qnetwork_target = QNetwork(state_size, action_size, hidden_size,
seed).to(device)
#self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
self.optimizer = optim.RMSprop(self.qnetwork_local.parameters(),
lr=LR,
momentum=0.95)
# 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
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, hidden_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.hidden_size = hidden_size
self.seed = random.seed(seed)
# Q-Network
self.qnetwork_local = QNetwork(state_size, action_size, hidden_size,
seed).to(device)
self.qnetwork_target = QNetwork(state_size, action_size, hidden_size,
seed).to(device)
#self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
self.optimizer = optim.RMSprop(self.qnetwork_local.parameters(),
lr=LR,
momentum=0.95)
# 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, idxs, ws = self.memory.sample()
self.learn(experiences, idxs, ws, 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, idxs, ws, 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
## TODO: compute and minimize the loss
next_action_values_local = self.qnetwork_local(states).gather(
1, actions)
# Only change proposed for Double DQN: Get maximizing future actions from local network and get their
# corresponding values from target network. Compare then these to the local taken actions.
local_max_actions = self.qnetwork_local(next_states).detach().max(
1)[1].unsqueeze(1)
next_action_values_target = self.qnetwork_target(
next_states).detach().gather(1, local_max_actions)
'''
print(next_action_values_local.shape)
print(next_action_values_local[0][:])
print(next_action_values_local.gather(1, actions).shape)
print(actions[0][0])
print(next_action_values_local.gather(1, actions)[0][0])
'''
y = rewards + (gamma * next_action_values_target * (1 - dones))
# Local network will be actualized, target one is used as ground truth
ws = torch.from_numpy(ws.astype(float)).float().to(device)
loss = F.mse_loss(ws * next_action_values_local, ws * y)
errors = np.abs(y.cpu().data.numpy() -
next_action_values_local.cpu().data.numpy())
self.memory.memory.update_batch(idxs, errors)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# ------------------- update target network ------------------- #
# Copy from local to target network parameters
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)
def adjust_learning_rate(self, episode, val):
print("adjusting learning rate!")
for param_group in self.optimizer.param_groups:
param_group['lr'] = val
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
hidden_layers = [128, 64]
self.qnetwork_local = QNetwork(state_size, action_size, hidden_layers,
seed).to(device)
self.qnetwork_target = QNetwork(state_size, action_size, hidden_layers,
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
## TODO: compute and minimize the loss
"*** YOUR CODE HERE ***"
max_actions = self.qnetwork_local.forward(next_states).detach().max(
1)[1].unsqueeze(1)
output_target = self.qnetwork_target.forward(next_states).gather(
1, max_actions)
td_target = rewards + gamma * (output_target * (1 - dones))
output_local = self.qnetwork_local(states).gather(1, actions)
loss = F.mse_loss(output_local, td_target)
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)