def __init__(self, state_size, action_size, random_seed=1):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
"""
# Store parameters
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.epsilon = EPSILON
self.lr_actor = LR_ACTOR
self.lr_critic = LR_CRITIC
self.lr_decay = WEIGHT_DECAY
# 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=self.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=self.lr_critic)
# Noise process
self.noise = OUNoise(action_size, random_seed)
self.timestep = 0
# Replay memory
self.memory = FifoMemory(BUFFER_SIZE, BATCH_SIZE)
# Short term memory contains only 1/100 of the complete memory and the most recent samples
self.memory_success = FifoMemory(int(BUFFER_SIZE), int(BATCH_SIZE))
self.memory_short = FifoMemory(5, 5)
def __init__(self, state_size, action_size, random_seed=1):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
"""
# Store parameters
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.epsilon = EPSILON
self.lr_actor = LR_ACTOR
self.lr_critic = LR_CRITIC
self.lr_decay = WEIGHT_DECAY
# 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=self.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=self.lr_critic)
# Noise process
self.noise = OUNoise(action_size, random_seed)
# Replay memory
self.memory = FifoMemory(BUFFER_SIZE, BATCH_SIZE)
# Success memory contains only the last 10 samples which led to a positive reward
self.memory_success = FifoMemory(int(BUFFER_SIZE), int(BATCH_SIZE))
# Rolling sample memory of last 10 samples
self.memory_short = FifoMemory(10, 10)
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 One Network (w/ Target Network)
self.critic_local_one = Critic(state_size, action_size,
random_seed).to(device)
self.critic_target_one = Critic(state_size, action_size,
random_seed).to(device)
self.critic_one_optimizer = optim.Adam(
self.critic_local_one.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
# Critic Two Network (w/ Target Network)
self.critic_local_two = Critic(state_size, action_size,
random_seed).to(device)
self.critic_target_two = Critic(state_size, action_size,
random_seed).to(device)
self.critic_two_optimizer = optim.Adam(
self.critic_local_two.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)
# Counter
self.t_step = 0
# learn_counter
self.learn_ctr = 0
def __init__(self, num_agents, state_size, action_size, random_seed,
actor_fc1_units, actor_fc2_units, critic_fcs1_units,
critic_fc2_units, buffer_size, batch_size, gamma, tau,
lr_actor, lr_critic, weight_decay, ou_mu, ou_theta, ou_sigma,
update_every_t_steps, num_of_updates):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
buffer_size (int) : replay buffer size
batch_size (int) : minibatch size
gamma (float) : discount factor
tau (float) : for soft update of target parameter
lr_actor (float) : learning rate of the actor
lr_critic (float) : learning rate of the critic
weight_decay (float) : L2 weight decay
ou_mu (float) : OUNoise mu
ou_theta (float) : OUNoise theta
ou_sigma (float) : OUNoise sigma
update_every_t_steps (int): timesteps between updates
num_of_updates (int): num of update passes when updating
"""
print(
"[AGENT INFO] DDPG constructor initialized parameters:\n num_agents={} \n state_size={} \n action_size={} \n random_seed={} \n actor_fc1_units={} \n actor_fc2_units={} \n critic_fcs1_units={} \n critic_fc2_units={} \n buffer_size={} \n batch_size={} \n gamma={} \n tau={} \n lr_actor={} \n lr_critic={} \n weight_decay={} \n ou_mu={}\n ou_theta={}\n ou_sigma={}\n update_every_t_steps={}\n num_of_updates={}\n"
.format(num_agents, state_size, action_size, random_seed,
actor_fc1_units, actor_fc2_units, critic_fcs1_units,
critic_fc2_units, buffer_size, batch_size, gamma, tau,
lr_actor, lr_critic, weight_decay, ou_mu, ou_theta,
ou_sigma, update_every_t_steps, num_of_updates))
self.num_agents = num_agents
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.actor_fc1_units = actor_fc1_units
self.actor_fc2_units = actor_fc2_units
self.critic_fcs1_units = critic_fcs1_units
self.critic_fc2_units = critic_fc2_units
self.buffer_size = buffer_size
self.batch_size = batch_size
self.gamma = gamma
self.tau = tau
self.lr_actor = lr_actor
self.lr_critic = lr_critic
self.weight_decay = weight_decay
self.ou_mu = ou_mu
self.ou_theta = ou_theta
self.ou_sigma = ou_sigma
self.update_every_t_steps = update_every_t_steps
self.num_of_updates = num_of_updates
# Actor Network (w/ Target Network)
self.actor_local = Actor(state_size, action_size, random_seed,
actor_fc1_units, actor_fc2_units).to(device)
self.actor_target = Actor(state_size, action_size, random_seed,
actor_fc1_units, actor_fc2_units).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,
critic_fcs1_units,
critic_fc2_units).to(device)
self.critic_target = Critic(state_size, action_size, random_seed,
critic_fcs1_units,
critic_fc2_units).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=lr_critic,
weight_decay=self.weight_decay)
# Noise process
self.noise = OUNoise(action_size,
random_seed,
mu=self.ou_mu,
theta=self.ou_theta,
sigma=self.ou_sigma)
# Replay memory
self.memory = ReplayBuffer(action_size, buffer_size, batch_size,
random_seed)
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self, num_agents, state_size, action_size, random_seed,
actor_fc1_units, actor_fc2_units, critic_fcs1_units,
critic_fc2_units, buffer_size, batch_size, gamma, tau,
lr_actor, lr_critic, weight_decay, ou_mu, ou_theta, ou_sigma,
update_every_t_steps, num_of_updates):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
buffer_size (int) : replay buffer size
batch_size (int) : minibatch size
gamma (float) : discount factor
tau (float) : for soft update of target parameter
lr_actor (float) : learning rate of the actor
lr_critic (float) : learning rate of the critic
weight_decay (float) : L2 weight decay
ou_mu (float) : OUNoise mu
ou_theta (float) : OUNoise theta
ou_sigma (float) : OUNoise sigma
update_every_t_steps (int): timesteps between updates
num_of_updates (int): num of update passes when updating
"""
print(
"[AGENT INFO] DDPG constructor initialized parameters:\n num_agents={} \n state_size={} \n action_size={} \n random_seed={} \n actor_fc1_units={} \n actor_fc2_units={} \n critic_fcs1_units={} \n critic_fc2_units={} \n buffer_size={} \n batch_size={} \n gamma={} \n tau={} \n lr_actor={} \n lr_critic={} \n weight_decay={} \n ou_mu={}\n ou_theta={}\n ou_sigma={}\n update_every_t_steps={}\n num_of_updates={}\n"
.format(num_agents, state_size, action_size, random_seed,
actor_fc1_units, actor_fc2_units, critic_fcs1_units,
critic_fc2_units, buffer_size, batch_size, gamma, tau,
lr_actor, lr_critic, weight_decay, ou_mu, ou_theta,
ou_sigma, update_every_t_steps, num_of_updates))
self.num_agents = num_agents
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.actor_fc1_units = actor_fc1_units
self.actor_fc2_units = actor_fc2_units
self.critic_fcs1_units = critic_fcs1_units
self.critic_fc2_units = critic_fc2_units
self.buffer_size = buffer_size
self.batch_size = batch_size
self.gamma = gamma
self.tau = tau
self.lr_actor = lr_actor
self.lr_critic = lr_critic
self.weight_decay = weight_decay
self.ou_mu = ou_mu
self.ou_theta = ou_theta
self.ou_sigma = ou_sigma
self.update_every_t_steps = update_every_t_steps
self.num_of_updates = num_of_updates
# Actor Network (w/ Target Network)
self.actor_local = Actor(state_size, action_size, random_seed,
actor_fc1_units, actor_fc2_units).to(device)
self.actor_target = Actor(state_size, action_size, random_seed,
actor_fc1_units, actor_fc2_units).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,
critic_fcs1_units,
critic_fc2_units).to(device)
self.critic_target = Critic(state_size, action_size, random_seed,
critic_fcs1_units,
critic_fc2_units).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=lr_critic,
weight_decay=self.weight_decay)
# Noise process
self.noise = OUNoise(action_size,
random_seed,
mu=self.ou_mu,
theta=self.ou_theta,
sigma=self.ou_sigma)
# Replay memory
self.memory = ReplayBuffer(action_size, buffer_size, batch_size,
random_seed)
# Make sure target is with the same weight as the source
#self.hard_copy(self.actor_target, self.actor_local)
#self.hard_copy(self.critic_target, self.critic_local)
def step(self, states, actions, rewards, next_states, dones, timestep):
"""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, if enough samples are available in memory
if len(
self.memory
) > self.batch_size and timestep % self.update_every_t_steps == 0:
for _ in range(self.num_of_updates):
experiences = self.memory.sample()
self.learn(experiences, self.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)
Q1_targets_next, Q2_targets_next = self.critic_target(
next_states, actions_next)
Q_targets_next = torch.min(Q1_targets_next, Q2_targets_next)
# Compute Q targets for current states (y_i)
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Compute critic loss
Q1_expected, Q2_expected = self.critic_local(states, actions)
critic_loss = F.mse_loss(Q1_expected, Q_targets) + F.mse_loss(
Q2_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.Q1(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, self.tau)
self.soft_update(self.actor_local, self.actor_target, self.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 hard_copy(self, target, source):
for target_param, param in zip(target.parameters(),
source.parameters()):
target_param.data.copy_(param.data)
class TD3Agent():
"""Interacts with and learns from the environment."""
def __init__(self, state_size, action_size, random_seed=1):
"""Initialize an Agent object.
Params
======
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
"""
# Store parameters
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.epsilon = EPSILON
self.lr_actor = LR_ACTOR
self.lr_critic = LR_CRITIC
self.lr_decay = WEIGHT_DECAY
# 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=self.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=self.lr_critic)
# Noise process
self.noise = OUNoise(action_size, random_seed)
# Replay memory
self.memory = FifoMemory(BUFFER_SIZE, BATCH_SIZE)
# Success memory contains only the last 10 samples which led to a positive reward
self.memory_success = FifoMemory(int(BUFFER_SIZE), int(BATCH_SIZE))
# Rolling sample memory of last 10 samples
self.memory_short = FifoMemory(10, 10)
def update_model(self, state, action, reward, next_state, done):
self.step(state, action, reward, next_state, done)
def step(self, state, action, reward, next_state, done):
"""Save experience in replay memory, and use random sample from buffer to learn."""
reached = True
if len(self.memory_success) < BATCH_SIZE:
reached = False
self.memory.add(state, action, reward, next_state, done)
self.memory_short.add(state, action, reward, next_state, done)
# Fill the success memory in case this agents receives positive reward
if reward > 0.0:
for i in range(len(self.memory_short)):
self.memory_success.add(
self.memory_short.samples[i].state, \
self.memory_short.samples[i].action, \
self.memory_short.samples[i].reward, \
self.memory_short.samples[i].next_state, \
self.memory_short.samples[i].done)
self.memory_short.clear()
if reached == False and len(self.memory_success) > BATCH_SIZE:
print("Success memory ready for use!")
# Train with the complete replay memory
if len(self.memory) > BATCH_SIZE:
for i in range(LEARN_NUM_MEMORY):
experiences = self.memory.sample()
# delay update of the policy and only update every 2nd training
self.learn(experiences, 0 , GAMMA)
# Train with the success replay memory
if (len(self.memory_success) > self.memory_success.batch_size):
for i in range(LEARN_NUM_MEMORY_SUCCESS):
experiences_success = self.memory_success.sample()
self.learn(experiences_success, 0 ,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()
# TD3 --> Action noise regularisation
if add_noise:
action += self.epsilon * self.noise.sample()
# The range of noise is clipped in order to keep the target value
# close to the original action.
clipped_action = np.clip(action, -1, 1)
self.epsilon *= EPSILON_DECAY
return clipped_action
def reset(self):
self.noise.reset()
def learn(self, experiences, delay ,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 ---------------------------- #
# TD3 --> Using a pair of critic networks (The twin part of the title)
# Get predicted next-state actions and Q values from target models
actions_next = self.actor_target(next_states)
Q_targets_next1, Q_targets_next2 = self.critic_target(next_states, actions_next)
# TD3 --> Take the minimum of both critic in order to avoid overestimation
#Q_targets_next = torch.min(Q_targets_next1, Q_targets_next2)
Q_targets_next = Q_targets_next1
# Compute Q targets for current states (y_i)
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Compute critic loss
Q_expected1, Q_expected2 = self.critic_local(states, actions)
# compute critic loss [HOW MUCH OFF?] as sum of both loss from target
#critic_loss = F.mse_loss(Q_expected1, Q_targets)+F.mse_loss(Q_expected2, Q_targets)
critic_loss = F.mse_loss(Q_expected1, Q_targets)
# minimize loss [TRAIN]
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 ---------------------------- #
# TD3 --> Delayed updates of the actor = policy (The delayed part)
# Compute actor loss
if delay == 0:
actions_pred = self.actor_local(states)
# compute loss [HOW MUCH OFF?]
actor_loss = -self.critic_local.Q1(states, actions_pred).mean()
# minimize loss [TRAIN]
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)
# ---------------------------- update noise ---------------------------- #
self.noise.reset()
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, 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 One Network (w/ Target Network)
self.critic_local_one = Critic(state_size, action_size,
random_seed).to(device)
self.critic_target_one = Critic(state_size, action_size,
random_seed).to(device)
self.critic_one_optimizer = optim.Adam(
self.critic_local_one.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
# Critic Two Network (w/ Target Network)
self.critic_local_two = Critic(state_size, action_size,
random_seed).to(device)
self.critic_target_two = Critic(state_size, action_size,
random_seed).to(device)
self.critic_two_optimizer = optim.Adam(
self.critic_local_two.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)
# Counter
self.t_step = 0
# learn_counter
self.learn_ctr = 0
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 every UPDATE_EVERY time steps.
# self.t_step = (self.t_step + 1) % UPDATE_EVERY
#
# if self.t_step == 0:
# 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 += np.random.normal(0, 0.2, size=action.shape)
return np.clip(action, -1, 1)
# def act(self, state, add_noise=True):
# """Returns actions for given state as per current policy."""
# if self.t_step < WARMUP:
# action = np.random.normal(scale=0.1, size=(self.action_size))
# else:
# 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 += np.random.normal(0, 0.1, action.shape)
# #update counter
# self.t_step += 1
# return np.clip(action, -1, 1)
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)
noise = torch.randn_like(actions_next).mul(0.2)
noise = noise.clamp(-0.5, 0.5)
actions_next = (actions_next + noise).clamp(-1, 1)
# actions_next = self.actor_target(next_states)
# actions_next = actions_next + torch.clamp(torch.from_numpy(np.random.normal(loc=0, scale=0.2, size=actions_next.shape)).float().to(device), -0.5, 0.5)
# actions_next = torch.clamp(actions_next, self.min_size[0], self.max_size[0])
critic_one_target = self.critic_target_one(next_states, actions_next)
critic_two_target = self.critic_target_two(next_states, actions_next)
Q_targets_next = torch.min(critic_one_target, critic_two_target)
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
Q_targets.detach()
critic_one_expected = self.critic_local_one(states, actions)
critic_two_expected = self.critic_local_two(states, actions)
# Compute Q targets for current states (y_i)
# Compute both critics loss
# Minimize the loss
critic_one_loss = F.mse_loss(critic_one_expected, Q_targets)
critic_two_loss = F.mse_loss(critic_two_expected, Q_targets)
critic_loss = critic_one_loss + critic_two_loss
self.critic_one_optimizer.zero_grad()
self.critic_two_optimizer.zero_grad()
critic_loss.backward()
self.critic_one_optimizer.step()
self.critic_two_optimizer.step()
self.learn_ctr = (self.learn_ctr + 1) % UPDATE_EVERY
if self.learn_ctr != 0:
return
# ---------------------------- update actor ---------------------------- #
# Compute actor loss
actor_loss = -self.critic_local_one(states,
self.actor_local(states)).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_one, self.critic_target_one, TAU)
self.soft_update(self.critic_local_two, self.critic_target_two, 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)