class DDPGAgent():
def __init__(self, state_dim, action_dim, random_seed):
self.state_dim = state_dim
self.action_dim = action_dim
self.seed = random.seed(random_seed)
# Actor network with its target network
self.actor_local = Actor(state_dim, action_dim, random_seed).to(device)
self.actor_target = Actor(state_dim, action_dim,
random_seed).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=LR_ACTOR)
# Critic network with its target network
self.critic_local = Critic(state_dim, action_dim,
random_seed).to(device)
self.critic_target = Critic(state_dim, action_dim,
random_seed).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
# Noise
self.noise = OUNoise(action_dim, random_seed)
self.epsilon = EPSILON
# Replay memory
self.memory = ReplayBuffer(action_dim, BUFFER_SIZE, BATCH_SIZE,
random_seed)
def step(self, state, action, reward, next_state, done, timestamp):
"""Save experience in replay memory, and use random sample from memory to learn."""
# Save experience
self.memory.add(state, action, reward, next_state, done)
# Learn (if there are enough samples in memory)
if len(self.memory) > BATCH_SIZE and timestamp % LEARN_EVERY == 0:
for _ in range(LEARN_NUMBER):
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state, add_noise=True):
"""Return actions for given state from 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() * self.epsilon
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."""
states, actions, rewards, next_states, dones = experiences
# UPDATE CRITIC #
actions_next = self.actor_target(next_states.to(device))
Q_targets_next = self.critic_target(next_states.to(device),
actions_next.to(device))
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
Q_expected = self.critic_local(states, actions)
critic_loss = F.mse_loss(Q_expected, Q_targets)
self.critic_optimizer.zero_grad()
critic_loss.backward()
clip_grad_norm_(self.critic_local.parameters(),
1) # Clip the gradient when update critic network
self.critic_optimizer.step()
# UPDATE ACTOR #
actions_pred = self.actor_local(states)
actor_loss = -self.critic_local(states, actions_pred).mean()
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# UPDATE TARGET NETWORKS #
self.soft_update(self.critic_local, self.critic_target, RHO)
self.soft_update(self.actor_local, self.actor_target, RHO)
# UPDATE EPSILON AND NOISE #
self.epsilon *= EPSILON_DECAY
self.noise.reset()
def soft_update(self, local_model, target_model, rho):
"""Soft update model parameters."""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(rho * target_param.data +
(1.0 - rho) * local_param.data)
class DDPGAgent:
def __init__(self,
output_dim,
input_dim,
name,
hidden=256,
lr_actor=1.0e-3,
lr_critic=1.0e-3,
tau=1.0e-2,
seed=10):
super(DDPGAgent, self).__init__()
self.seed = seed
self.actor = Actor(input_dim, hidden, output_dim, seed).to(device)
self.critic = Critic(input_dim=input_dim,
action_dim=output_dim,
hidden=hidden,
seed=seed,
output_dim=1).to(device)
self.target_actor = Actor(input_dim, hidden, output_dim,
seed).to(device)
self.target_critic = Critic(input_dim=input_dim,
action_dim=output_dim,
hidden=hidden,
seed=seed,
output_dim=1).to(device)
self.name = name
self.noise = OUNoise(output_dim, seed)
self.tau = tau
self.epsilon = EPSILON
self.gamma = GAMMA
self.clipgrad = CLIPGRAD
self.actor_optimizer = optim.Adam(self.actor.parameters(), lr=lr_actor)
self.critic_optimizer = optim.Adam(self.critic.parameters(),
lr=lr_critic,
weight_decay=0)
def act(self, state, add_noise=True):
"""Return actions for given state from current policy."""
state = torch.from_numpy(state).float().unsqueeze(0).to(
device) #.unsqueeze(0)
self.actor.eval()
with torch.no_grad():
action = self.actor(state).cpu().squeeze(0).data.numpy()
self.actor.train()
if add_noise:
action += self.noise.sample() * self.epsilon
return np.clip(action, -1, 1)
def learn(self, experiences):
"""Update policy and value parameters using given batch of experience tuples."""
states, actions, rewards, next_states, dones = experiences
# UPDATE CRITIC #
actions_next = self.target_actor(next_states.to(device))
Q_targets_next = self.target_critic(next_states.to(device),
actions_next.to(device))
Q_targets = rewards + (self.gamma * Q_targets_next * (1 - dones))
Q_expected = self.critic(states, actions)
critic_loss = F.mse_loss(Q_expected, Q_targets)
self.critic_optimizer.zero_grad()
critic_loss.backward()
clip_grad_norm_(self.critic.parameters(), self.clipgrad)
self.critic_optimizer.step()
# UPDATE ACTOR #
actions_pred = self.actor(states)
actor_loss = -self.critic(states, actions_pred).mean()
self.actor_optimizer.zero_grad()
actor_loss.backward()
#clip_grad_norm_(self.actor.parameters(), self.clipgrad)
self.actor_optimizer.step()
# UPDATE TARGET NETWORKS #
self.soft_update(self.critic, self.target_critic)
self.soft_update(self.actor, self.target_actor)
# UPDATE EPSILON AND NOISE #
self.epsilon *= EPSILON_DECAY
self.noise.reset()
def reset(self):
self.noise.reset()
def soft_update(self, local_model, target_model):
"""Soft update model parameters."""
for target_param, local_param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(self.tau * local_param.data +
(1.0 - self.tau) * target_param.data)