class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, seed, model="QNetwork"):
"""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
if model == "QNetwork":
self.qnetwork_local = QNetwork(state_size, action_size,
seed).to(device)
self.qnetwork_target = QNetwork(state_size, action_size,
seed).to(device)
if model == "QNetworkConvolutional":
self.qnetwork_local = QNetworkConvolutional(
state_size, action_size, seed).to(device)
self.qnetwork_target = QNetworkConvolutional(
state_size, action_size, seed).to(device)
if model == "DuelingDQN":
self.qnetwork_local = DuelingDQN(state_size, action_size,
seed).to(device)
self.qnetwork_target = DuelingDQN(state_size, action_size,
seed).to(device)
print("Model: " + model)
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 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."""
def __init__(self, state_size, action_size, seed, max_t=1000):
"""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 = DuelingDQN(state_size, action_size,
seed).to(device)
self.qnetwork_target = DuelingDQN(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
self.prio_b = PRIO_B
self.b_step = 0
self.max_b_step = 2000
self.learnFirst = True
def step(self, state, action, reward, next_state, done):
# Save experience in replay memory
#self.memory.add(state, action, reward, next_state, done)
# Hassan : Save the experience in prioritized replay memory
self.memory.prio_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)
# Hassan : prioritized replay memory
self.b_step = self.b_step + 1
experiences, indices = self.memory.prio_sample()
self.learn(experiences, GAMMA, indices)
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 get_beta(self, t):
'''
Return the current exponent β based on its schedul. Linearly anneal β
from its initial value β0 to 1, at the end of learning.
:param t: integer. Current time step in the episode
:return current_beta: float. Current exponent beta
'''
#f_frac = min(float(t) / self.max_b_step, 1.0)
#current_beta = self.prio_b + f_frac * (1. - self.prio_b)
#current_beta = min(1,current_beta)
self.prio_b = min(1, self.prio_b + PRIO_B_INC)
return self.prio_b
def learn(self, experiences, gamma, indices):
"""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, probabilities = experiences
## TODO: compute and minimize the loss
"*** YOUR CODE HERE ***"
# Get max predicted Q values (for next states) from target model
# Hassan : Action is selected using greedy policy
#Q_targets_next = self.qnetwork_target(next_states).detach().max(1)[0].unsqueeze(1)
# Hassan : Double DQN
# Selecting actions which maximizes while taking w (qnetwork_local)
next_actions = self.qnetwork_local(next_states).detach().argmax(
dim=1).unsqueeze(1)
#next_actions_test = self.qnetwork_local(next_states).detach().max(1)[1].unsqueeze(1) # Hassan : from the example
#print(torch.sum(next_actions-next_actions_test)) # Hassan : no difference found
# Selecting q values of these actions using w' (qnetwork_target)
Q_targets_next = self.qnetwork_target(next_states).gather(
1, next_actions)
# Compute Q targets for current states
# Hassan : This is TD Target
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Get expected Q values from local model
# Hassan : This is current value
Q_expected = self.qnetwork_local(states).gather(1, actions)
#Hassan : Compute the td_error
td_error = Q_targets - Q_expected
#print(td_error.detach().numpy())
#self.prio_b = min(1, PRIO_B_INC+self.prio_b)
f_currbeta = self.get_beta(0)
#print(f_currbeta)
#f_currbeta = self.get_beta(self.b_step)
#print(self.b_step)
#print(t)
#print(self.prio_b)
weights_importance = probabilities.mul_(
self.memory.__len__()).pow_(-f_currbeta)
# Hassan : calculate max_weights_importance
#probabilities_min = min(self.memory.priorities)/self.memory.cum_priorities
probabilities_min = self.memory.min_priority / self.memory.cum_priorities
max_weights_importance = (probabilities_min *
self.memory.__len__())**(-f_currbeta)
# Hassan : divide the weights importance with the max_weights_importance
# Hassan : Improvement why not calculating the max_weights_importance = max(weights_importance)??
# Hassan : this will only calculating on the current list not the complete one
#print(weights_importance)
#print(weights_importance.max(0)[0])
#print(max_weights_importance)
#if self.learnFirst:
# self.learnFirst = False
#else :
# max_weights_importance = max_weights_importance[0]
weights_final = weights_importance.div_(max_weights_importance)
square_weighted_error = td_error.pow_(2).mul_(weights_final)
loss = square_weighted_error.mean()
# Hassan : after the observations observation from example, update was done after the weights calculation
if self.prio_b > 0.5:
self.memory.prio_update(indices,
td_error.detach().numpy(), PRIO_E, PRIO_A)
# 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 ------------------- #
# Hassan : Here not after C steps w is changed though cahnged slightly after every learn step
# Hassan : We can modify to change this after ever C steps
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."""
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 = DuelingDQN(state_size, action_size,
seed).to(device)
self.qnetwork_target = DuelingDQN(state_size, action_size,
seed).to(device)
self.optimizer = optim.Adam(self.qnetwork_local.parameters(), lr=LR)
self.priority_alpha = 0.0 #current best: 03
self.priority_beta_start = 0.4
self.priority_beta_frames = BUFFER_SIZE
# Replay memory
self.memory = PrioritizedReplayMemory(BUFFER_SIZE, self.priority_alpha,
self.priority_beta_start,
self.priority_beta_frames)
# 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.push((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 self.memory.storage_size() > BATCH_SIZE:
#print("storage == ", self.memory.storage_size())
experiences, idxes, weights = self.memory.sample(BATCH_SIZE)
self.learn(experiences, idxes, weights, 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, idxes, weights, 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 = zip(*experiences)
states = torch.from_numpy(
np.vstack([state for state in states
if state is not None])).float().to(device)
actions = torch.from_numpy(
np.vstack([action for action in actions
if action is not None])).long().to(device)
rewards = torch.from_numpy(
np.vstack([reward for reward in rewards
if reward is not None])).float().to(device)
next_states = torch.from_numpy(
np.vstack([
next_state for next_state in next_states
if next_state is not None
])).float().to(device)
dones = torch.from_numpy(
np.vstack([done for done in dones if done is not None
]).astype(np.uint8)).float().to(device)
# Get max predicted Q values (for next states) from target model
#print("state-action values:")
#print(self.qnetwork_target(next_states).detach())
#print(next_states)
next_target_Q = self.qnetwork_target.forward(next_states)
#print("next_target_Q == ", next_target_Q)
_, next_local_Q_index = torch.max(
self.qnetwork_local.forward(next_states), axis=1)
#Q_targets_next = self.qnetwork_target(next_states).detach().max(1)[0].unsqueeze(1)
Q_targets_next = next_target_Q[range(next_target_Q.shape[0]),
next_local_Q_index]
Q_targets_next1 = Q_targets_next.reshape((len(Q_targets_next), 1))
# Compute Q targets for current states
Q_targets = rewards + (gamma * Q_targets_next1 * (1 - dones))
# Get expected Q values from local model
Q_expected = self.qnetwork_local(states).gather(1, actions)
#print(Q_expected)
#print(Q_targets)
diff = Q_expected - Q_targets
#print(diff)
#diff = diff.mean()
#print(idxes)
#print(diff.detach().squeeze().abs().cpu().numpy().tolist())
#update the priority of the replay buffer
self.memory.update_priorities(
idxes,
diff.detach().squeeze().abs().cpu().numpy().tolist())
# Compute loss
loss = F.mse_loss(Q_expected, Q_targets) * weights
loss = loss.mean()
# 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."""
def __init__(self, state_size, action_size, seed,max_t=1000):
"""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 = DuelingDQN(state_size, action_size, seed).to(device)
self.qnetwork_target = DuelingDQN(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
self.prio_b = PRIO_B
self.b_step = 0
self.max_b_step = 2000
self.learnFirst = True
def step(self, state, action, reward, next_state, done):
# Hassan : Save the experience in prioritized replay memory
self.memory.prio_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:
self.b_step = self.b_step + 1
experiences, indices = self.memory.prio_sample()
self.learn(experiences, GAMMA, indices)
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 get_beta(self, t):
'''
Return the current exponent β based on its schedul. Linearly anneal β
from its initial value β0 to 1, at the end of learning.
:param t: integer. Current time step in the episode
:return current_beta: float. Current exponent beta
'''
#f_frac = min(float(t) / self.max_b_step, 1.0)
#current_beta = self.prio_b + f_frac * (1. - self.prio_b)
#current_beta = min(1,current_beta)
self.prio_b = min(1,self.prio_b + PRIO_B_INC)
return self.prio_b
def learn(self, experiences, gamma, indices):
"""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, probabilities = experiences
"*** YOUR CODE HERE ***"
# Double DQN implementation
# Selecting actions which maximizes while taking w (qnetwork_local)
next_actions = self.qnetwork_local(next_states).detach().argmax(dim=1).unsqueeze(1)
# evluate best actions using w' (qnetwork_target)
Q_targets_next = self.qnetwork_target(next_states).gather(1, next_actions)
# Compute Q targets for current states (TD Target)
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 the td_error
td_error = Q_targets - Q_expected
f_currbeta = self.get_beta(0)
# Prioritized experience replay : calculating the final weights for calculating loss function
weights_importance = probabilities.mul_(self.memory.__len__()).pow_(-f_currbeta)
probabilities_min = self.memory.min_priority/self.memory.cum_priorities
max_weights_importance = (probabilities_min * self.memory.__len__())**(-f_currbeta)
weights_final = weights_importance.div_(max_weights_importance)
# Compute mean squared weighted error
square_weighted_error = td_error.pow_(2).mul_(weights_final)
loss = square_weighted_error.mean()
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# Prioritized experience replay : updating the priority of experience tuple in replay buffer
self.memory.prio_update(indices,td_error.detach().numpy(),PRIO_E,PRIO_A)
# ------------------- 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."""
def __init__(self, config):
"""Initialize an Agent object"""
self.seed = random.seed(config["general"]["seed"])
self.config = config
# Q-Network
self.q = DuelingDQN(config).to(DEVICE)
self.q_target = DuelingDQN(config).to(DEVICE)
self.optimizer = optim.RMSprop(self.q.parameters(),
lr=config["agent"]["learning_rate"])
self.criterion = F.mse_loss
self.memory = ReplayBuffer(config)
# Initialize time step (for updating every UPDATE_EVERY steps)
self.t_step = 0
def save_experiences(self, state, action, reward, next_state, done):
"""Prepare and save experience in replay memory"""
reward = np.clip(reward, -1.0, 1.0)
self.memory.add(state, action, reward, next_state, done)
def _current_step_is_a_learning_step(self):
"""Check if the current step is an update step"""
self.t_step = (self.t_step + 1) % self.config["agent"]["update_rate"]
return self.t_step == 0
def _enough_samples_in_memory(self):
"""Check if minimum amount of samples are in memory"""
return len(self.memory) > self.config["train"]["batch_size"]
def epsilon_greedy_action_selection(self, action_values, eps):
"""Epsilon-greedy action selection"""
if random.random() > eps:
return np.argmax(action_values.cpu().data.numpy())
else:
return random.choice(
np.arange(self.config["general"]["action_size"]))
def act(self, state, eps=0.0):
"""Returns actions for given state as per current policy"""
state = torch.from_numpy(state).float().unsqueeze(0).to(DEVICE)
self.q.eval()
with torch.no_grad():
action_values = self.q(state)
self.q.train()
return self.epsilon_greedy_action_selection(action_values, eps)
def _calc_loss(self, states, actions, rewards, next_states, dones):
"""Calculates loss for a given experience batch"""
q_eval = self.q(states).gather(1, actions)
q_eval_next = self.q(next_states)
_, q_argmax = q_eval_next.detach().max(1)
q_next = self.q_target(next_states)
q_next = q_next.gather(1, q_argmax.unsqueeze(1))
q_target = rewards + (self.config["agent"]["gamma"] * q_next *
(1 - dones))
loss = self.criterion(q_eval, q_target)
return loss
def _update_weights(self, loss):
"""update the q network weights"""
torch.nn.utils.clip_grad.clip_grad_value_(self.q.parameters(), 1.0)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
def learn(self):
"""Update network using one sample of experience from memory"""
if self._current_step_is_a_learning_step(
) and self._enough_samples_in_memory():
states, actions, rewards, next_states, dones = self.memory.sample(
self.config["train"]["batch_size"])
loss = self._calc_loss(states, actions, rewards, next_states,
dones)
self._update_weights(loss)
self._soft_update(self.q, self.q_target)
def _soft_update(self, local_model, target_model):
"""Soft update target network parameters: θ_target = τ*θ_local + (1 - τ)*θ_target"""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(
self.config["agent"]["tau"] * local_param.data +
(1.0 - self.config["agent"]["tau"]) * target_param.data)
def save(self):
"""Save the network weights"""
helper.mkdir(
os.path.join(".", *self.config["general"]["checkpoint_dir"],
self.config["general"]["env_name"]))
current_date_time = helper.get_current_date_time()
current_date_time = current_date_time.replace(" ", "__").replace(
"/", "_").replace(":", "_")
torch.save(
self.q.state_dict(),
os.path.join(".", *self.config["general"]["checkpoint_dir"],
self.config["general"]["env_name"],
"ckpt_" + current_date_time))
def load(self):
"""Load latest available network weights"""
list_of_files = glob.glob(
os.path.join(".", *self.config["general"]["checkpoint_dir"],
self.config["general"]["env_name"], "*"))
latest_file = max(list_of_files, key=os.path.getctime)
self.q.load_state_dict(torch.load(latest_file))
self.q_target.load_state_dict(torch.load(latest_file))
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self,
state_size,
action_size,
seed,
gamma=GAMMA,
buffer_size=BUFFER_SIZE,
batch_size=BATCH_SIZE,
update_every=UPDATE_EVERY,
lr=LR,
tau=TAU
):
"""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.gamma = gamma
self.batch_size = batch_size
# Q-Network
self.device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
self.model_local = DuelingDQN(state_size, action_size, seed).to(self.device)
self.model_target = DuelingDQN(state_size, action_size, seed).to(self.device)
self.optimizer = optim.Adam(self.model_local.parameters(), lr=LR)
# Replay memory
self.memory = ReplayBuffer(
action_size=action_size,
buffer_size=BUFFER_SIZE,
batch_size=BATCH_SIZE,
seed=seed,
device=self.device
)
# 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) > self.batch_size:
experiences = self.memory.sample()
self.update(experiences)
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.FloatTensor(state).float().unsqueeze(0).to(self.device)
self.model_local.eval()
with torch.no_grad():
qvals = self.model_local.forward(state)
self.model_local.train()
# Epsilon-greedy action selection
if random.random() > eps:
action = np.argmax(qvals.cpu().detach().numpy())
return action
else:
return random.choice(np.arange(self.action_size))
def update(self, batch):
"""Update value parameters using given batch of experience tuples.
Params
======
batch (Tuple[torch.Tensor]): tuple of (s, a, r, s', done) tuples
gamma (float): discount factor
"""
states, actions, rewards, next_states, dones = batch
# Get expected Q values from local model
curr_Q = self.model_local.forward(states).gather(1, actions)
# curr_Q = curr_Q.squeeze(1)
# Get max predicted Q values (for next states) from target model
max_next_Q = self.model_target.forward(next_states).detach().max(1)[0].unsqueeze(1)
expected_Q = rewards + (self.gamma * max_next_Q * (1 - dones))
loss = F.mse_loss(curr_Q, expected_Q)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# update target model
self.update_target(self.model_local, self.model_target, TAU)
def update_target(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)