def learn(self):
# Sample a batch from the replay buffer
transitions = self.replay.sample()
states, full_state, actions, rewards, next_states, next_full_state, dones = transpose_to_tensor(
transitions, self.config.device)
### Update online critic model ###
# Compute actions for next states with the target actor model
with torch.no_grad(): # don't use gradients for target
target_next_actions = [
self.target_actor(next_states[:, i, :])
for i in range(self.config.num_agents)
]
target_next_actions = torch.cat(target_next_actions, dim=1)
# Compute Q values for the next states and next actions with the target critic model
with torch.no_grad(): # don't use gradients for target
target_next_qs = self.target_critic(
next_full_state.to(self.config.device),
target_next_actions.to(self.config.device))
# Compute Q values for the current states and actions
target_qs = rewards.sum(
1, keepdim=True) + self.config.discount * target_next_qs * (
1 - dones.max(1, keepdim=True)[0])
# Compute Q values for the current states and actions with the online critic model
actions = actions.view(actions.shape[0], -1)
online_qs = self.online_critic(full_state.to(self.config.device),
actions.to(self.config.device))
# Compute and minimize the online critic loss
online_critic_loss = F.mse_loss(online_qs, target_qs.detach())
self.online_critic_opt.zero_grad()
online_critic_loss.backward()
torch.nn.utils.clip_grad_norm_(self.online_critic.parameters(), 1)
self.online_critic_opt.step()
### Update online actor model ###
# Compute actions for the current states with the online actor model
online_actions = [
self.online_actor(states[:, i, :])
for i in range(self.config.num_agents)
]
online_actions = torch.cat(online_actions, dim=1)
# Compute the online actor loss with the online critic model
online_actor_loss = -self.online_critic(
full_state.to(self.config.device),
online_actions.to(self.config.device)).mean()
# Minimize the online critic loss
self.online_actor_opt.zero_grad()
online_actor_loss.backward()
self.online_actor_opt.step()
### Update target critic and actor models ###
soft_update(self.target_actor, self.online_actor,
self.config.target_mix)
soft_update(self.target_critic, self.online_critic,
self.config.target_mix)
def optimize(self):
"""
Samples a random batch from replay memory and performs optimization
:return:
"""
s1, a1, r1, s2 = self.ram.sample(self.args.batch_size)
s1 = Variable(torch.from_numpy(s1))
a1 = Variable(torch.from_numpy(a1))
r1 = Variable(torch.from_numpy(r1))
s2 = Variable(torch.from_numpy(s2))
# ---------------------- optimize critic ----------------------
# Use target actor exploitation policy here for loss evaluation
a2 = self.target_actor.forward(s2).detach()
next_val = torch.squeeze(self.target_critic.forward(s2, a2).detach())
# y_exp = r + gamma*Q'( s2, pi'(s2))
y_expected = r1 + self.args.gamma * next_val
# y_pred = Q( s1, a1)
y_predicted = torch.squeeze(self.critic.forward(s1, a1))
# compute critic loss, and update the critic
loss_critic = F.smooth_l1_loss(y_predicted, y_expected)
self.critic_optimizer.zero_grad()
loss_critic.backward()
self.critic_optimizer.step()
# ---------------------- optimize actor ----------------------
pred_a1 = self.actor.forward(s1)
loss_actor = -1 * torch.sum(self.critic.forward(s1, pred_a1))
self.actor_optimizer.zero_grad()
loss_actor.backward()
self.actor_optimizer.step()
utils.soft_update(self.target_actor, self.actor, self.args.tau)
utils.soft_update(self.target_critic, self.critic, self.args.tau)
def optimize(self):
s1, a1, r1, s2, done = self.ram.sample(self.batch_size)
s1 = Variable(torch.tensor(s1)).cuda()
a1 = Variable(torch.tensor(a1, dtype=torch.int64)).cuda()
r1 = Variable(torch.tensor(r1)).cuda()
s2 = Variable(torch.tensor(s2)).cuda()
self.optimizer.zero_grad()
# optimize
#
action_predict = torch.argmax(self.learning_net.forward(s2), dim=1)
r_predict = torch.squeeze(
self.target_net.forward(s2).gather(1, action_predict.view(-1, 1)))
r_predict = self.gamma * r_predict
y_j = r1 + r_predict
y_j = self.done_state_value(r1, y_j, done)
# r_ : Q(s_j, a_j)
r_ = self.learning_net.forward(s1)
r_ = torch.squeeze(r_.gather(1, a1.view(-1, 1)))
# loss: (y_j - Q(s_j, a_j))^2
loss = self.loss_f(y_j, r_)
loss.backward()
self.optimizer.step()
utils.soft_update(self.target_net, self.learning_net, self.tau)
self.iter += 1
return loss.cpu()
def learn(self):
self.learning_steps += 1
if self.learning_steps % self.target_update_interval == 0:
soft_update(self.critic_target, self.critic, self.tau)
if self.per:
# batch with indices and priority weights
batch, indices, weights = \
self.memory.sample(self.batch_size)
else:
batch = self.memory.sample(self.batch_size)
# set priority weights to 1 when we don't use PER.
weights = 1.
q1_loss, q2_loss, errors, mean_q1, mean_q2 =\
self.calc_critic_loss(batch, weights)
policy_loss, entropies = self.calc_policy_loss(batch, weights)
update_params(self.q1_optim, self.critic.Q1, q1_loss, self.grad_clip)
update_params(self.q2_optim, self.critic.Q2, q2_loss, self.grad_clip)
update_params(self.policy_optim, self.policy, policy_loss,
self.grad_clip)
if self.entropy_tuning:
entropy_loss = self.calc_entropy_loss(entropies, weights)
update_params(self.alpha_optim, None, entropy_loss)
self.alpha = self.log_alpha.exp()
if self.per:
# update priority weights
self.memory.update_priority(indices, errors.cpu().numpy())
def update(self, batch, update_actor=True):
"""Updates parameters of TD3 actor and critic given samples from the batch.
Args:
batch: A list of timesteps from environment.
update_actor: a boolean variable, whether to perform a policy update.
"""
obs = contrib_eager_python_tfe.Variable(
np.stack(batch.obs).astype('float32'))
action = contrib_eager_python_tfe.Variable(
np.stack(batch.action).astype('float32'))
next_obs = contrib_eager_python_tfe.Variable(
np.stack(batch.next_obs).astype('float32'))
mask = contrib_eager_python_tfe.Variable(
np.stack(batch.mask).astype('float32'))
if self.get_reward is not None:
reward = self.get_reward(obs, action, next_obs)
else:
reward = contrib_eager_python_tfe.Variable(
np.stack(batch.reward).astype('float32'))
if self.use_td3:
self._update_critic_td3(obs, action, next_obs, reward, mask)
else:
self._update_critic_ddpg(obs, action, next_obs, reward, mask)
if self.critic_step.numpy() % self.policy_update_freq == 0:
if update_actor:
self._update_actor(obs, mask)
soft_update(self.actor.variables, self.actor_target.variables,
self.tau)
soft_update(self.critic.variables, self.critic_target.variables,
self.tau)
def optimize(self):
"""
Samples a random batch from replay memory and performs optimization
:return:
"""
s1, a1, r1, s2 = self.ram.sample(BATCH_SIZE)
s1 = Variable(torch.from_numpy(s1))
a1 = Variable(torch.from_numpy(a1))
r1 = Variable(torch.from_numpy(r1))
s2 = Variable(torch.from_numpy(s2))
# ---------------------- optimize critic ----------------------
a2 = self.target_actor.forward(s2).detach()
next_val = torch.squeeze(self.target_critic.forward(s2, a2).detach())
y_expected = r1 + GAMMA * next_val
y_predicted = torch.squeeze(self.critic.forward(s1, a1))
loss_critic = F.smooth_l1_loss(y_predicted, y_expected)
self.critic_optimizer.zero_grad()
loss_critic.backward()
self.critic_optimizer.step()
# ---------------------- optimize actor ----------------------
pred_a1 = self.actor.forward(s1)
loss_actor = -1 * torch.sum(self.critic.forward(s1, pred_a1))
self.actor_optimizer.zero_grad()
loss_actor.backward()
self.actor_optimizer.step()
utils.soft_update(self.target_actor, self.actor, TAU)
utils.soft_update(self.target_critic, self.critic, TAU)
def learn(self, times=1):
for i in range(times):
if self.buffer.len < self.batch_size:
return
s, a, r, s_ = self.buffer.sample(self.batch_size)
s = Variable(torch.from_numpy(s).float())
a = Variable(torch.from_numpy(a).float())
r = Variable(torch.from_numpy(r).float())
s_ = Variable(torch.from_numpy(s_).float())
#print (s)
# for Critic
#a_ = self.target_actor.forward(s_).detach().data.numpy()
a_ = self.target_actor.forward(s_).detach()
next_Q = self.target_critic.forward(s_, a_).detach()
y = r + self.gamma * next_Q
y_pred = self.critic.forward(s, a)
loss_critic = F.smooth_l1_loss(y_pred, y)
self.critic_optimizer.zero_grad()
loss_critic.backward()
self.critic_optimizer.step()
# for Actor
#pred_a = self.actor.forward(s).data.numpy()
pred_a = self.actor.forward(s)
loss_actor = -1 * torch.sum(self.critic.forward(s, pred_a))
self.actor_optimizer.zero_grad()
loss_actor.backward()
self.actor_optimizer.step()
utils.soft_update(self.target_actor, self.actor, self.tau)
utils.soft_update(self.target_critic, self.critic, self.tau)
def optimize(self):
s1, a1, r1, s2, done = self.ram.sample(self.batch_size)
s1 = Variable(torch.tensor(s1)).cuda()
a1 = Variable(torch.tensor(a1, dtype=torch.int64)).cuda()
r1 = Variable(torch.tensor(r1)).cuda()
s2 = Variable(torch.tensor(s2)).cuda()
self.optimizer.zero_grad()
# optimize
r_predict = self.gamma * torch.max(self.target_net.forward(s2), dim=1).values
y_j = r1 + r_predict
y_j = self.done_state_value(r1, y_j, done)
# r_ : Q(s_j, a_j)
r_ = self.learning_net.forward(s1)
# mask = F.one_hot(torch.squeeze(a1), num_classes=self.act_dim)
# mask = torch.tensor(mask.clone().detach(), dtype=torch.uint8).cuda()
# r_ = torch.masked_select(r_, mask)
r_ = torch.squeeze(r_.gather(1, a1.view(-1,1)))
# loss: (y_j - Q(s_j, a_j))^2
loss = self.loss_f(y_j, r_)
loss.backward()
self.optimizer.step()
utils.soft_update(self.target_net, self.learning_net, self.tau)
self.iter += 1
return loss.cpu()
def update_policy(self):
state_batch, action_batch, reward_batch, next_state_batch, terminal_batch = self.memory.sample_batch(self.batch_size)
state = to_tensor(np.array(state_batch), device=device)
action = to_tensor(np.array(action_batch), device=device)
next_state = to_tensor(np.array(next_state_batch), device=device)
# compute target Q value
next_q_value = self.critic_target([next_state, self.actor_target(next_state)])
target_q_value = to_tensor(reward_batch, device=device) \
+ self.discount * to_tensor((1 - terminal_batch.astype(np.float)), device=device) * next_q_value
# Critic and Actor update
self.critic.zero_grad()
with torch.set_grad_enabled(True):
q_values = self.critic([state, action])
critic_loss = criterion(q_values, target_q_value.detach())
critic_loss.backward()
self.critic_optim.step()
self.actor.zero_grad()
with torch.set_grad_enabled(True):
policy_loss = -self.critic([state.detach(), self.actor(state)]).mean()
policy_loss.backward()
self.actor_optim.step()
# Target update
soft_update(self.actor_target, self.actor, self.tau)
soft_update(self.critic_target, self.critic, self.tau)
return to_numpy(-policy_loss), to_numpy(critic_loss), to_numpy(q_values.mean())
def learn(self, experiences, all_curr_pred_actions, all_next_pred_actions):
agent_idx_device = torch.tensor(self.agent_idx).to(self.device)
states, actions, rewards, next_states, dones = experiences
rewards = rewards.index_select(1, agent_idx_device)
dones = dones.index_select(1, agent_idx_device)
# ---------------------------- update critic ---------------------------- #
# Get predicted next-state actions and Q values from target models
batch_size = next_states.shape[0]
actions_next = torch.cat(all_next_pred_actions, dim=1).to(self.device)
next_states = next_states.reshape(batch_size, -1)
with torch.no_grad():
Q_targets_next = self.critic_target(next_states, actions_next)
# Compute Q targets for current states (y_i)
Q_targets = rewards + (self.gamma * Q_targets_next * (1 - dones))
# Compute critic loss
states = states.reshape(batch_size, -1)
actions = actions.reshape(batch_size, -1)
Q_expected = self.critic_local(states, actions)
critic_loss = F.mse_loss(Q_expected, Q_targets.detach())
# Minimize the loss
self.critic_optim.zero_grad()
critic_loss.backward()
self.critic_optim.step()
# ---------------------------- update actor ---------------------------- #
# Compute actor loss
self.actor_optim.zero_grad()
predicted_actions = torch.cat([action if idx == self.agent_idx \
else action.detach()
for idx, action in enumerate(all_curr_pred_actions)],
dim=1).to(self.device)
actor_loss = -self.critic_local(states, predicted_actions).mean()
# minimize loss
actor_loss.backward()
self.actor_optim.step()
al = actor_loss.cpu().detach().item()
cl = critic_loss.cpu().detach().item()
if self.tensorboard_writer is not None:
self.tensorboard_writer.add_scalar("agent{}/actor_loss".format(self.agent_idx), al, self.iteration)
self.tensorboard_writer.add_scalar("agent{}/critic_loss".format(self.agent_idx), cl, self.iteration)
self.tensorboard_writer.file_writer.flush()
self.iteration += 1
# ----------------------- update target networks ----------------------- #
soft_update(self.critic_target, self.critic_local, self.tau)
soft_update(self.actor_target, self.actor_local, self.tau)
def update(self):
if len(self.replay_buffer) < self.batch_size:
return
states, actions, rewards, next_states, dones = self.replay_buffer.sample(
batch_size=self.batch_size, device='cpu')
#===============================Critic Update===============================
with torch.no_grad():
target = rewards + GAMMA * (1 - dones) * self.Q_target(
(next_states, self.P_target(next_states)))
Q = self.Q_online((states, actions))
td_error = self.loss_td(target, Q)
self.q_optimizer.zero_grad()
td_error.backward()
self.q_optimizer.step()
#===============================Actor Update===============================
q = self.Q_online((states, self.P_online(states)))
loss_a = -torch.mean(q)
self.p_optimizer.zero_grad()
loss_a.backward()
self.p_optimizer.step()
#===============================Target Update===============================
soft_update(self.Q_target, self.Q_online, tau=1e-2)
soft_update(self.P_target, self.P_online, tau=1e-2)
def update_all_targets(self):
"""
Update all target networks (called after normal updates have been
performed for each agent)
"""
soft_update(self.critic, self.target_critic, self.tau)
soft_update(self.policy, self.target_policy, self.tau)
def update(self):
if len(self.memory) < self.BATCH_SIZE:
return
# get training batch
transitions = self.memory.sample(self.BATCH_SIZE)
batch = Transition(*zip(*transitions))
state_batch = torch.cat(batch.state)
action_batch = torch.cat(batch.action)
reward_batch = torch.cat(batch.reward).unsqueeze(1)
next_state = torch.cat(batch.next_state)
# update value network
state_action = torch.cat((state_batch, action_batch), dim=1)
state_action_value = self.value_network(state_action)
next_action = self.action_target_network(next_state).detach()
next_state_action = torch.cat((next_state, next_action), dim=1)
next_state_action_value = self.value_target_network(
next_state_action).detach()
expected_state_action_value = (self.DISCOUNT *
next_state_action_value) + reward_batch
value_loss = self.criterion(state_action_value,
expected_state_action_value)
self.value_optimizer.zero_grad()
value_loss.backward()
self.value_optimizer.step()
# update action network
optim_action = self.action_network(state_batch)
optim_state_action = torch.cat((state_batch, optim_action), dim=1)
action_loss = -self.value_network(optim_state_action)
action_loss = action_loss.mean()
self.action_optimizer.zero_grad()
action_loss.backward()
self.action_optimizer.step()
# update target network
soft_update(self.value_target_network, self.value_network, 0.01)
soft_update(self.action_target_network, self.action_network, 0.01)
def update_parameters(self, memory, batch_size, updates):
# Sample a batch from memory
state_batch, action_batch, reward_batch, next_state_batch, mask_batch = memory.sample(batch_size=batch_size)
state_batch = torch.FloatTensor(state_batch).to(self.device)
next_state_batch = torch.FloatTensor(next_state_batch).to(self.device)
action_batch = torch.FloatTensor(action_batch).to(self.device)
reward_batch = torch.FloatTensor(reward_batch).to(self.device).unsqueeze(1)
mask_batch = torch.FloatTensor(mask_batch).to(self.device).unsqueeze(1)
with torch.no_grad():
next_state_action, next_state_log_pi, _ = self.policy.sample(next_state_batch)
qf1_next_target, qf2_next_target = self.critic_target(next_state_batch, next_state_action)
min_qf_next_target = torch.min(qf1_next_target, qf2_next_target) - self.alpha * next_state_log_pi
next_q_value = reward_batch + mask_batch * self.gamma * (min_qf_next_target)
qf1, qf2 = self.critic(state_batch, action_batch) # Two Q-functions to mitigate positive bias in the policy improvement step
qf1_loss = F.mse_loss(qf1, next_q_value) # JQ = 𝔼(st,at)~D[0.5(Q1(st,at) - r(st,at) - γ(𝔼st+1~p[V(st+1)]))^2]
qf2_loss = F.mse_loss(qf2, next_q_value) # JQ = 𝔼(st,at)~D[0.5(Q1(st,at) - r(st,at) - γ(𝔼st+1~p[V(st+1)]))^2]
pi, log_pi, _ = self.policy.sample(state_batch)
qf1_pi, qf2_pi = self.critic(state_batch, pi)
min_qf_pi = torch.min(qf1_pi, qf2_pi)
policy_loss = ((self.alpha * log_pi) - min_qf_pi).mean() # Jπ = 𝔼st∼D,εt∼N[α * logπ(f(εt;st)|st) − Q(st,f(εt;st))]
self.critic_optim.zero_grad()
qf1_loss.backward()
self.critic_optim.step()
self.critic_optim.zero_grad()
qf2_loss.backward()
self.critic_optim.step()
self.policy_optim.zero_grad()
policy_loss.backward()
self.policy_optim.step()
if self.automatic_entropy_tuning:
alpha_loss = -(self.log_alpha * (log_pi + self.target_entropy).detach()).mean()
self.alpha_optim.zero_grad()
alpha_loss.backward()
self.alpha_optim.step()
self.alpha = self.log_alpha.exp()
alpha_tlogs = self.alpha.clone() # For TensorboardX logs
else:
alpha_loss = torch.tensor(0.).to(self.device)
alpha_tlogs = torch.tensor(self.alpha) # For TensorboardX logs
if updates % self.target_update_interval == 0:
soft_update(self.critic_target, self.critic, self.tau)
return qf1_loss.item(), qf2_loss.item(), policy_loss.item(), alpha_loss.item(), alpha_tlogs.item()
def learn(self):
self.learning_steps += 1
if self.learning_steps % self.target_update_interval == 0:
soft_update(self.critic_target, self.critic, self.tau)
# Update the latent model.
self.learn_latent()
# Update policy and critic.
self.learn_sac()
def train_critic(self):
loss = self.td_critic_loss(self.critic)
self.critic_optimizer.zero_grad()
self.critic_upd_steps += 1
writer.add_scalar("critic loss", loss, self.critic_upd_steps)
loss.backward()
self.critic_optimizer.step()
soft_update(self.critic_target, self.critic, args.tau)
return loss
def update_parameters(self, batch):
obs, act, rew, done, obs_next = batch
obs = torch.FloatTensor(obs).to(self.device)
act = torch.FloatTensor(act).to(self.device)
rew = torch.FloatTensor(rew).unsqueeze(-1).to(self.device)
done = torch.BoolTensor(done).unsqueeze(-1).to(self.device)
obs_next = torch.FloatTensor(obs_next).to(self.device)
with torch.no_grad():
next_v = self.value_target(obs_next).masked_fill(done, 0.)
q_targ = rew + self.gamma * next_v
self.critic_optim.zero_grad()
q1, q2 = self.critic(obs, act)
critic_loss = (q1 - q_targ).pow(2.).mul(0.5) + (
q2 - q_targ).pow(2.).mul(0.5)
critic_loss = critic_loss.mean()
critic_loss.backward()
self.critic_optim.step()
with torch.no_grad():
critic_loss = (torch.min(q1, q2) - q_targ).pow(2).mul(0.5).mean()
self.policy_optim.zero_grad()
policy = TanhGaussian(*self.policy(obs))
action = policy.sample()
log_pi = policy.log_prob(action, param_grad=False).sum(dim=-1,
keepdim=True)
action_value = torch.min(*self.critic(obs, action))
with torch.no_grad():
v_targ = action_value - self.alpha * log_pi
self.value_optim.zero_grad()
v = self.value(obs)
value_loss = (v - v_targ).pow(2.).mul(0.5)
value_loss = value_loss.mean()
value_loss.backward()
self.value_optim.step()
policy_loss = self.alpha * log_pi - action_value
policy_loss = policy_loss.mean()
policy_loss.backward()
self.policy_optim.step()
soft_update(self.value_target, self.value, self.tau)
loss_info = {
'critic_loss': critic_loss.item(),
'policy_loss': policy_loss.item(),
'value_loss': value_loss.item(),
'policy_entropy': -log_pi.mean().item()
}
return loss_info
def update_parameters(self):
self.learning_steps += 1
# Update Q networks.
for i in range(self.Q_updates_per_step):
# Sample a batch from memory.
states, actions, rewards, next_states, masks = self.memory.sample(
batch_size=self.batch_size)
self.critic_optim.zero_grad()
qf1_loss, qf2_loss = self.calc_critic_loss(states, actions,
rewards, next_states,
masks)
qf1_loss.backward()
qf2_loss.backward()
clip_grad_norm_(self.critic_online.parameters(),
max_norm=self.max_grad_norm)
self.critic_optim.step()
policy_loss, kl, entropies, cross_entropies, policy_term = self.calc_policy_loss(
states, self.learning_steps)
# Update coefficents
if self.limit_kl:
self.alpha_optim.zero_grad()
alpha_loss = self.log_alpha * (
(self.target_kl - self.target_entropy) - cross_entropies)
alpha_loss.backward()
self.alpha_optim.step()
self.alpha_optim.zero_grad()
rho_loss = self.log_rho * (entropies - self.target_entropy)
rho_loss.backward()
self.rho_optim.step()
self.rho_optim.zero_grad()
# Update policy networks.
self.backup_policy()
self.policy_optim.zero_grad()
policy_loss.backward()
clip_grad_norm_(self.policy.parameters(), max_norm=self.max_grad_norm)
self.policy_optim.step()
# Soft update target critic network.
soft_update(self.critic_target, self.critic_online, self.tau)
# Log training information.
if self.learning_steps % self.log_interval == 0 and self.learning_steps > 100:
self.writer.add_scalar('loss/Q1',
qf1_loss.detach().item(),
self.learning_steps)
self.writer.add_scalar('loss/policy',
policy_loss.detach().item(),
self.learning_steps)
self.writer.add_scalar('stats/kl', kl, self.learning_steps)
self.writer.add_scalar('stats/entropies', entropies,
self.learning_steps)
self.writer.add_scalar('stats/cross_entropies', cross_entropies,
self.learning_steps)
def train(self):
state_batches, action_batches, reward_batches, next_state_batches, done_batches = self.get_batches(
)
state_batches = torch.Tensor(state_batches).to(self.device)
action_batches = torch.Tensor(action_batches).to(self.device)
reward_batches = torch.Tensor(reward_batches).reshape(
self.config.batch_size, self.n_agents, 1).to(self.device)
next_state_batches = torch.Tensor(next_state_batches).to(self.device)
done_batches = torch.Tensor(
(done_batches == False) * 1).reshape(self.config.batch_size,
self.n_agents,
1).to(self.device)
target_next_actions = self.policy_target.forward(next_state_batches)
target_next_q = self.critic_target.forward(next_state_batches,
target_next_actions)
main_q = self.critic(state_batches, action_batches)
'''
How to concat each agent's Q value?
'''
#target_next_q = target_next_q
#main_q = main_q.mean(dim=1)
'''
Reward Norm
'''
# reward_batches = (reward_batches - reward_batches.mean(dim=0)) / reward_batches.std(dim=0) / 1024
# Critic Loss
self.critic.zero_grad()
baselines = reward_batches + done_batches * self.config.gamma * target_next_q
loss_critic = torch.nn.MSELoss()(main_q, baselines.detach())
loss_critic.backward()
torch.nn.utils.clip_grad_norm_(self.critic.parameters(), 0.5)
self.critic_optimizer.step()
# Actor Loss
self.policy.zero_grad()
clear_action_batches = self.policy.forward(state_batches)
loss_actor = -self.critic.forward(state_batches,
clear_action_batches).mean()
loss_actor += (clear_action_batches**2).mean() * 1e-3
loss_actor.backward()
torch.nn.utils.clip_grad_norm_(self.policy.parameters(), 0.5)
self.policy_optimizer.step()
# This is for logging
self.c_loss = loss_critic.item()
self.a_loss = loss_actor.item()
soft_update(self.policy, self.policy_target, self.config.tau)
soft_update(self.critic, self.critic_target, self.config.tau)
def update_parameters(self, memory, batch_size, updates):
# Sample a batch from memory
state_batch, action_batch, reward_batch, next_state_batch, mask_batch = memory.sample(batch_size=batch_size)
state_batch = torch.FloatTensor(state_batch).to(self.device)
next_state_batch = torch.FloatTensor(next_state_batch).to(self.device)
action_batch = torch.FloatTensor(action_batch).to(self.device)
reward_batch = torch.FloatTensor(reward_batch).to(self.device).unsqueeze(1)
mask_batch = torch.FloatTensor(mask_batch).to(self.device).unsqueeze(1)
with torch.no_grad():
next_state_action, next_state_log_pi, _ = self.policy.sample(next_state_batch)
qf1_next_target, qf2_next_target = self.critic_target(next_state_batch, next_state_action)
min_qf_next_target = torch.min(qf1_next_target, qf2_next_target) - self.alpha * next_state_log_pi
next_q_value = reward_batch + mask_batch * self.gamma * (min_qf_next_target)
# Two Q-functions to mitigate positive bias in the policy improvement step
qf1, qf2 = self.critic(state_batch, action_batch)
qf1_loss = F.mse_loss(qf1, next_q_value)
qf2_loss = F.mse_loss(qf2, next_q_value)
pi, log_pi, _ = self.policy.sample(state_batch)
qf1_pi, qf2_pi = self.critic(state_batch, pi)
min_qf_pi = torch.min(qf1_pi, qf2_pi)
policy_loss = ((self.alpha * log_pi) - min_qf_pi).mean()
self.critic_optim.zero_grad()
qf1_loss.backward()
self.critic_optim.step()
self.critic_optim.zero_grad()
qf2_loss.backward()
self.critic_optim.step()
self.policy_optim.zero_grad()
policy_loss.backward()
self.policy_optim.step()
alpha_loss = -(self.log_alpha * (log_pi + self.target_entropy).detach()).mean()
self.alpha_optim.zero_grad()
alpha_loss.backward()
self.alpha_optim.step()
self.alpha = self.log_alpha.exp()
alpha_tlogs = self.alpha.clone() # For TensorboardX logs
soft_update(self.critic_target, self.critic, self.tau)
def update_parameters(self, batch):
obs, act, rew, done, obs_next = batch
obs = torch.FloatTensor(obs).to(self.device)
act = torch.LongTensor(act).unsqueeze(-1).to(self.device)
rew = torch.FloatTensor(rew).unsqueeze(-1).to(self.device)
done = torch.BoolTensor(done).unsqueeze(-1).to(self.device)
obs_next = torch.FloatTensor(obs_next).to(self.device)
with torch.no_grad():
next_policy = Cat(raw_base=self.policy(obs_next))
next_q = torch.min(*self.critic_target(obs_next))
next_eval = (next_policy.probs * next_q).sum(dim=-1, keepdim=True)
next_entr = -(next_policy.probs * next_policy.logits).sum(
dim=-1, keepdim=True)
next_v = (next_eval + self.alpha * next_entr).masked_fill(done, 0.)
q_targ = rew + self.gamma * next_v
self.critic_optim.zero_grad()
q1, q2 = self.critic(obs)
q_pred = torch.min(q1, q2).detach()
q1, q2 = q1.gather(dim=-1, index=act), q2.gather(dim=-1, index=act)
critic_loss = (q1 - q_targ).pow(2.).mul(0.5) + (
q2 - q_targ).pow(2.).mul(0.5)
critic_loss = critic_loss.mean()
critic_loss.backward()
self.critic_optim.step()
with torch.no_grad():
critic_loss = (torch.min(q1, q2) - q_targ).pow(2.).mul(0.5).mean()
self.policy_optim.zero_grad()
policy = Cat(raw_base=self.policy(obs))
policy_entr = -(policy.probs.detach() *
policy.logits).sum(dim=-1).mean()
policy_eval = (policy.probs * q_pred).sum(dim=-1).mean()
policy_loss = self.alpha * policy_entr - policy_eval
policy_loss.backward()
self.policy_optim.step()
soft_update(self.critic_target, self.critic, self.tau)
loss_info = {
'critic_loss': critic_loss.item(),
'policy_loss': policy_loss.item(),
'policy_entr': policy_entr.item()
}
return loss_info
def learn(self, experiences, gamma):
states, actions, rewards, next_states, dones = experiences
Q_targets_next = self.qnetwork_target(next_states).detach().max(
1)[0].unsqueeze(1)
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
Q_expected = self.qnetwork_local(states).gather(1, actions)
# Compute loss and backpropagate
loss = F.mse_loss(Q_expected, Q_targets)
self.optimizer.zero_grad()
loss.backward()
self.optimizer.step()
# Update target network
soft_update(self.qnetwork_local, self.qnetwork_target, self.tau)
def update_parameters(self, memory, batch_size, updates):
"""
:param memory:
:param batch_size:
:param updates:
:return:
"""
state_batch, action_batch, reward_batch, next_state_batch, mask_batch = memory.sample(
batch_size=batch_size)
states = torch.FloatTensor(state_batch).to(self.device)
next_states = torch.FloatTensor(next_state_batch).to(self.device)
actions = torch.FloatTensor(action_batch).to(self.device)
rewards = torch.FloatTensor(reward_batch).to(self.device).unsqueeze(1)
done = torch.FloatTensor(mask_batch).to(self.device).unsqueeze(1)
# UPDATE CRITIC #
# Get predicted next-state actions and Q values from target models
actions_next = self.target_policy_net(next_states)
q_targets_next = self.target_value_net(next_states,
actions_next.detach())
# Compute Q targets for current states (y_i)
q_targets = rewards + (self.gamma * q_targets_next * (1.0 - done))
# Compute critic loss
q_expected = self.critic_net(states, actions)
critic_loss = self.critic_loss(q_expected, q_targets)
# Minimize the loss
self.critic_opt.zero_grad()
critic_loss.backward()
self.critic_opt.step()
# UPDATE ACTOR #
# Compute actor loss
actions_pred = self.actor_net(states)
actor_loss = -self.critic_net(states, actions_pred).mean()
# Maximize the expected return
self.actor_opt.zero_grad()
actor_loss.backward()
self.actor_opt.step()
# UPDATE TARGET NETWORK #
if updates % self.target_update == 0:
soft_update(self.critic_net, self.target_value_net, self.tau)
soft_update(self.actor_net, self.target_policy_net, self.tau)
return actor_loss.item(), critic_loss.item()
def update(self):
if len(self.replay_buffer) < self.batch_size:
return
states, actions, rewards, next_states, dones = self.replay_buffer.sample(
batch_size=self.batch_size, device=self.device)
# discounted rewards
# rewards = torch.from_numpy(discount((rewards.view(rewards.shape[0])).cpu().numpy())).float().to(self.device)
### debug shape : ok
#===============================Critic Update===============================
self.Q_online.train()
Q = self.Q_online((states, actions))
with torch.no_grad(): # don't need backprop for target value
self.Q_target.eval()
self.P_target.eval()
target = rewards + self.gamma * (1 - dones) * self.Q_target(
(next_states, self.P_target(next_states)))
critic_loss_fn = torch.nn.MSELoss()
critic_loss = critic_loss_fn(Q, target).mean()
# update
self.q_optimizer.zero_grad()
critic_loss.backward()
self.q_optimizer.step()
# print("critic loss", critic_loss.item())
#===============================Actor Update===============================
# fix online_critic , update online_actor
self.Q_online.eval()
for p in self.Q_online.parameters():
p.requires_grad = False
for p in self.P_online.parameters():
p.requires_grad = True
policy_loss = -self.Q_online((states, self.P_online(states)))
policy_loss = policy_loss.mean()
self.p_optimizer.zero_grad()
policy_loss.backward()
self.p_optimizer.step()
# print("policy loss", policy_loss.item())
for p in self.Q_online.parameters():
p.requires_grad = True
#===============================Target Update===============================
soft_update(self.Q_target, self.Q_online, tau=1e-3)
soft_update(self.P_target, self.P_online, tau=1e-3)
self.eps -= EPSILON_DECAY
if self.eps <= 0:
self.eps = 0
def optimize(self):
"""
Samples a random batch from replay memory and performs optimization
:return:
"""
s1, a1, r1, s2 = self.ram.sample(BATCH_SIZE)
s1 = Variable(torch.from_numpy(s1))
a1 = Variable(torch.from_numpy(a1))
r1 = Variable(torch.from_numpy(r1))
s2 = Variable(torch.from_numpy(s2))
# ---------------------- optimize critic ----------------------
# Use target actor exploitation policy here for loss evaluation
a2 = self.target_actor.forward(s2).detach()
next_val = torch.squeeze(self.target_critic.forward(s2, a2).detach())
# y_exp = r + gamma*Q'( s2, pi'(s2))
y_expected = r1 + GAMMA * next_val
# y_pred = Q( s1, a1)
y_predicted = torch.squeeze(self.critic.forward(s1, a1))
# compute critic loss, and update the critic
loss_critic = F.smooth_l1_loss(y_predicted, y_expected)
self.critic_optimizer.zero_grad()
loss_critic.backward()
self.critic_optimizer.step()
# ---------------------- optimize actor ----------------------
pred_a1 = self.actor.forward(s1)
# compute actor loss and update actor
loss_actor = -1 * torch.sum(self.critic.forward(s1, pred_a1))
self.actor_optimizer.zero_grad()
loss_actor.backward()
self.actor_optimizer.step()
# updating target networks according to: y = TAU*x + (1 - TAU)*y
utils.soft_update(self.target_actor, self.actor, TAU)
utils.soft_update(self.target_critic, self.critic, TAU)
if self.iter % 100 == 0:
print('Iteration :- ', self.iter, ' Loss_actor :- ', loss_actor.data.numpy(),\
' Loss_critic :- ', loss_critic.data.numpy())
self.iter += 1
def optimize(self):
"""
Samples a random batch from replay memory and performs optimization
:return:
"""
if self.args.pri:
s1, a1, r1, s2, tree_idx, weights = self.ram.sample(
self.batch_size)
weights = torch.from_numpy(weights).to(self.device)
else:
s1, a1, r1, s2 = self.ram.sample(self.batch_size)
s1 = Variable(torch.from_numpy(s1)).to(self.device)
a1 = Variable(torch.from_numpy(a1)).to(self.device)
r1 = Variable(torch.from_numpy(r1)).to(self.device)
s2 = Variable(torch.from_numpy(s2)).to(self.device)
# ---------------------- optimize critic ----------------------
# Use target actor exploitation policy here for loss evaluation
a2 = self.target_actor.forward(s2).detach()
next_val = torch.squeeze(self.target_critic.forward(s2, a2).detach())
y_expected = r1 + self.gamma * next_val
y_predicted = torch.squeeze(self.critic.forward(s1, a1))
if self.args.pri:
td_error = torch.abs(y_predicted - y_expected)
loss_critic = torch.sum(weights * td_error**2)
self.ram.update_tree(tree_idx, td_error.detach().cpu().numpy())
else:
loss_critic = F.mse_loss(y_predicted, y_expected)
self.critic_optimizer.zero_grad()
loss_critic.backward()
self.critic_optimizer.step()
# ---------------------- optimize actor ----------------------
pred_a1 = self.actor.forward(s1)
loss_actor = -1 * torch.mean(self.critic.forward(s1, pred_a1))
self.actor_optimizer.zero_grad()
loss_actor.backward()
self.actor_optimizer.step()
# logging
self.writer.add_scalar('loss/critic', loss_critic, self.iter)
self.writer.add_scalar('loss/actor', loss_actor, self.iter)
utils.soft_update(self.AC_T, self.AC, self.tau)
self.iter += 1
def train(self):
state_batches, action_batches, reward_batches, next_state_batches, done_batches = self.get_batches(
)
state_batches = Variable(torch.Tensor(state_batches).to(self.device))
action_batches = Variable(
torch.Tensor(action_batches).reshape(-1, 1).to(self.device))
reward_batches = Variable(
torch.Tensor(reward_batches).reshape(-1, 1).to(self.device))
next_state_batches = Variable(
torch.Tensor(next_state_batches).to(self.device))
done_batches = Variable(
torch.Tensor(
(done_batches == False) * 1).reshape(-1, 1).to(self.device))
target_next_actions = self.actor_target.forward(
next_state_batches).detach()
target_next_q = self.critic_target.forward(
next_state_batches, target_next_actions).detach()
main_q = self.critic(state_batches, action_batches)
# Critic Loss
self.critic.zero_grad()
baselines = reward_batches + done_batches * self.config.gamma * target_next_q
loss_critic = torch.nn.MSELoss()(main_q, baselines)
loss_critic.backward()
self.critic_optimizer.step()
# Actor Loss
self.actor.zero_grad()
clear_action_batches = self.actor.forward(state_batches)
loss_actor = (
-self.critic.forward(state_batches, clear_action_batches)).mean()
loss_actor.backward()
self.actor_optimizer.step()
# This is for logging
self.c_loss = loss_critic.item()
self.a_loss = loss_actor.item()
soft_update(self.actor, self.actor_target, self.config.tau)
soft_update(self.critic, self.critic_target, self.config.tau)
def _learn_from_memory(self):
'''从记忆学习，更新两个网络的参数
'''
# 随机获取记忆里的Transmition
trans_pieces = self.sample(self.batch_size)
s0 = np.vstack([x.s0 for x in trans_pieces])
a0 = np.array([x.a0 for x in trans_pieces])
r1 = np.array([x.reward for x in trans_pieces])
# is_done = np.array([x.is_done for x in trans_pieces])
s1 = np.vstack([x.s1 for x in trans_pieces])
# 优化评论家网络参数
a1 = self.target_actor.forward(s1).detach()
next_val = torch.squeeze(self.target_critic.forward(s1, a1).detach())
# y_exp = r + gamma*Q'( s2, pi'(s2))
y_expected = torch.from_numpy(r1).type(
torch.FloatTensor) + self.gamma * next_val
y_expected = y_expected.type(torch.FloatTensor)
# y_pred = Q( s1, a1)
a0 = torch.from_numpy(a0) # 转换成Tensor
y_predicted = torch.squeeze(self.critic.forward(s0, a0))
# compute critic loss, and update the critic
loss_critic = F.smooth_l1_loss(y_predicted, y_expected)
self.critic_optimizer.zero_grad()
loss_critic.backward()
self.critic_optimizer.step()
# 优化演员网络参数，优化的目标是使得Q增大
pred_a0 = self.actor.forward(s0) # 直接使用a0会不收敛
#反向梯度下降(梯度上升)，以某状态的价值估计为策略目标函数
loss_actor = -1 * torch.sum(self.critic.forward(s0, pred_a0))
self.actor_optimizer.zero_grad()
loss_actor.backward()
self.actor_optimizer.step()
# 软更新参数
soft_update(self.target_actor, self.actor, self.tau)
soft_update(self.target_critic, self.critic, self.tau)
return (loss_critic.item(), loss_actor.item())
def _learn_from_memory(self):
'''从experience学习，更新两个网络的参数；
'''
# 随机获取记忆里的Transmition
trans_pieces = self.sample(self.batch_size)
s0 = np.vstack([x.s0 for x in trans_pieces])
a0 = np.array([x.a0 for x in trans_pieces])
r1 = np.array([x.reward for x in trans_pieces])
s1 = np.vstack([x.s1 for x in trans_pieces])
# 优化critic网络参数，最小化loss
a1 = self.target_actor.forward(s1).detach()
next_val = torch.squeeze(self.target_critic.forward(s1, a1).detach())
r1 = torch.from_numpy(r1)
# r1 = r1.type(torch.DoubleTensor)
next_val = next_val.type(torch.DoubleTensor)
y_expected = r1 + self.gamma * next_val
y_expected = y_expected.type(torch.FloatTensor)
a0 = torch.from_numpy(a0) # 转换成Tensor
y_predicted = torch.squeeze(self.critic.forward(s0, a0))
# 最小化loss，更新critic
loss_critic = F.smooth_l1_loss(y_predicted, y_expected)
self.critic_optimizer.zero_grad()
loss_critic.backward()
self.critic_optimizer.step()
# 优化actor网络参数，优化的目标是使得Q增大
pred_a0 = self.actor.forward(s0) # 直接使用a0会不收敛
#反向梯度下降(梯度上升)，以某状态的价值估计为策略目标函数
loss_actor = -1 * torch.sum(self.critic.forward(s0, pred_a0))
self.actor_optimizer.zero_grad()
loss_actor.backward()
self.actor_optimizer.step()
# 软更新参数，跟新target网络
soft_update(self.target_actor, self.actor, self.tau)
soft_update(self.target_critic, self.critic, self.tau)
return (loss_critic.item(), loss_actor.item())
def replay_buffer_training(self, sample, train_results, n):
s, a, r, t, stag = [sample[k] for k in ['s', 'a', 'r', 't', 'stag']]
self.train_mode()
with torch.no_grad():
pi_tag = self.pi_target(stag)
noise = (torch.randn_like(pi_tag) * self.policy_noise).clamp(
-self.noise_clip, self.noise_clip)
pi_tag = (pi_tag + noise).clamp(-1, 1)
q_target_1 = self.q_target_1(stag, pi_tag)
q_target_2 = self.q_target_2(stag, pi_tag)
q_target = torch.min(q_target_1, q_target_2)
g = r + (1 - t) * self.gamma**self.n_steps * q_target
qa = self.q_net_1(s, a)
loss_q = F.mse_loss(qa, g, reduction='mean')
self.optimizer_q_1.zero_grad()
loss_q.backward()
if self.clip_q:
nn.utils.clip_grad_norm(self.q_net_1.parameters(), self.clip_q)
self.optimizer_q_1.step()
qa = self.q_net_2(s, a)
loss_q = F.mse_loss(qa, g, reduction='mean')
self.optimizer_q_2.zero_grad()
loss_q.backward()
if self.clip_q:
nn.utils.clip_grad_norm(self.q_net_2.parameters(), self.clip_q)
self.optimizer_q_2.step()
if not n % self.td3_delayed_policy_update:
pi = self.pi_net(s)
v = self.q_net_1(s, pi)
loss_p = (-v).mean()
self.optimizer_p.zero_grad()
loss_p.backward()
if self.clip_p:
nn.utils.clip_grad_norm(self.pi_net.parameters(), self.clip_p)
self.optimizer_p.step()
train_results['scalar']['objective'].append(float(-loss_p))
soft_update(self.pi_net, self.pi_target, self.tau)
soft_update(self.q_net_1, self.q_target_1, self.tau)
soft_update(self.q_net_2, self.q_target_2, self.tau)
train_results['scalar']['loss_q'].append(float(loss_q))
return train_results