class DDPGAgent(object):
"""Interacts with and learns from the environment."""
def __init__(self, id, state_size, action_size, seed, memory, num_agents,
hyperparameters: Mapping[str, float]):
"""Initialize a DDPG agent object.
Params
======
id (int): agent's id
state_size (int): dimension of each state
action_size (int): dimension of each action
seed (int): random seed
memory (ReplayBuffer): replay buffer to store the experience of this agent
hyperparameters (dictionnary of str:): hyperparameters' values of the model. The expected parameters are:
- batch_size (int): minibatch size
- lr_actor (float): learning rate of the actor
- lr_critic (float): learning rate of the critic
- gamma (float): discount factor
- weight_decay (float): critic L2 weight decay
- tau (float): value for soft update of target parameters
- update_frequency (int): how much steps must be executed before starting learn
- n_learns (int): how many learning for update
"""
self.id = id
self.__name__ = 'DDPG'
self.state_size = state_size
self.action_size = action_size
self.gamma = hyperparameters['gamma']
self.batch_size = int(hyperparameters['batch_size'])
self.tau = hyperparameters['tau']
self.update_frequency = int(hyperparameters['update_frequency'])
self.n_learns = int(hyperparameters['n_learns'])
# Actor Network (w/ Target Network)
self.actor_local = Actor(state_size, action_size, seed).to(device)
self.actor_target = Actor(state_size, action_size, seed).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=hyperparameters['lr_actor'])
# Critic Network (w/ Target Network)
self.critic_local = Critic(state_size, action_size, num_agents,
seed).to(device)
self.critic_target = Critic(state_size, action_size, num_agents,
seed).to(device)
self.critic_optimizer = optim.Adam(
self.critic_local.parameters(),
lr=hyperparameters['lr_critic'],
weight_decay=hyperparameters['weight_decay'])
# Noise process
self.noise = Ornstein(action_size)
# Replay memory
self.memory = memory
# Initialize the time step (for every update_frequency steps)
self.t_step = 0
def step(self, state, action, reward, next_state, done, other_states,
other_actions, other_next_states):
"""Save experience in replay memory, and use random sample from buffer to learn."""
self.memory.add(state, action, reward, next_state, done, other_states,
other_actions, other_next_states)
self.t_step = (self.t_step + 1) % self.update_frequency
if self.t_step == 0:
# Learn, if enough samples are available in memory
for _ in range(self.n_learns):
if len(self.memory) > self.batch_size:
experiences = self.memory.sample(self.batch_size)
self.learn(experiences, self.gamma)
def act(self, states, add_noise=True):
"""Returns actions for given state as per current policy."""
states = torch.from_numpy(states).float().to(device)
self.actor_local.eval()
with torch.no_grad():
action = self.actor_local(states).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, _, _, _, _, other_states, _, _ = experiences
self.update_critic(experiences, gamma)
self.update_actor(states, other_states)
self.update_target_networks()
def update_critic(self, experiences, gamma):
"""Update the critic network given the experiences"""
states, actions, rewards, next_states, dones, other_states, other_actions, other_next_states = experiences
all_states = torch.cat((states, other_states), dim=1).to(device)
all_actions = torch.cat((actions, other_actions), dim=1).to(device)
all_next_states = torch.cat((next_states, other_next_states),
dim=1).to(device)
local_all_next_actions = []
local_all_next_actions.append(self.actor_target(states))
local_all_next_actions.append(self.actor_target(other_states))
all_next_actions = torch.cat(local_all_next_actions, dim=1).to(device)
# ---------------------------- update critic ---------------------------- #
# Get predicted next-state actions and Q values from target models
Q_targets_next = self.critic_target(all_next_states, all_next_actions)
# Compute Q targets for current states (y_i)
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Compute critic loss
Q_expected = self.critic_local(all_states, all_actions)
critic_loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
def update_actor(self, states, other_states):
all_states = torch.cat((states, other_states), dim=1).to(device)
# ---------------------------- update actor ---------------------------- #
# Compute actor loss
actions_pred = self.actor_local(states)
other_actions_pred = self.actor_local(other_states)
other_actions_pred = other_actions_pred.detach()
actions_pred = torch.cat((actions_pred, other_actions_pred),
dim=1).to(device)
actor_loss = -self.critic_local(all_states, actions_pred).mean()
# Minimize the loss
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
def update_target_networks(self):
# ----------------------- 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)
class DDPGAgent:
def __init__(self,
env,
gamma,
tau,
buffer_maxlen,
critic_learning_rate,
actor_learning_rate,
max_action=1):
self.device = torch.device(
"cuda" if torch.cuda.is_available() else "cpu")
self.env = env
self.obs_dim = env.observation_space.shape[0]
self.action_dim = env.action_space.shape[0]
self.noise = OUNoise(env.action_space)
self.iter = 0.0
self.noisy = False
self.max_action = max_action
print(self.action_dim)
print(self.obs_dim)
# RL hyperparameters
self.env = env
self.gamma = gamma
self.tau = tau
# Initialize critic and actorr networks
self.critic = Critic(self.obs_dim, self.action_dim).to(self.device)
self.critic_target = Critic(self.obs_dim,
self.action_dim).to(self.device)
self.actor = Actor(self.obs_dim, self.action_dim,
self.max_action).to(self.device)
self.actor_target = Actor(self.obs_dim,
self.action_dim).to(self.device)
# Copy target network paramters for critic
for target_param, param in zip(self.critic_target.parameters(),
self.critic.parameters()):
target_param.data.copy_(param.data)
# Set Optimization algorithms
self.critic_optimizer = optim.Adam(self.critic.parameters(),
lr=critic_learning_rate)
self.actor_optimizer = optim.Adam(self.actor.parameters(),
lr=actor_learning_rate)
self.replay_buffer = ExperienceReplayLog(buffer_maxlen)
def get_action(self, obs):
#print('obs;',obs)
if self.noisy == True:
state = torch.FloatTensor(obs).unsqueeze(0).to(self.device)
action = self.actor.forward(state)
action = action.squeeze(0).cpu().detach().numpy()
action = self.noise.get_action(action, self.iter)
self.iter = self.iter + 1
else:
state = torch.FloatTensor(obs).unsqueeze(0).to(self.device)
action = self.actor.forward(state)
action = action.squeeze(0).cpu().detach().numpy()
return action
def update(self, batch_size):
#Batch updates
states, actions, rewards, next_states, _ = self.replay_buffer.sample(
batch_size)
state_batch, action_batch, reward_batch, next_state_batch, masks = self.replay_buffer.sample(
batch_size)
state_batch = torch.FloatTensor(state_batch).to(self.device)
action_batch = torch.FloatTensor(action_batch).to(self.device)
reward_batch = torch.FloatTensor(reward_batch).to(self.device)
next_state_batch = torch.FloatTensor(next_state_batch).to(self.device)
masks = torch.FloatTensor(masks).to(self.device)
# Q info updates
curr_Q = self.critic.forward(state_batch, action_batch)
next_actions = self.actor_target.forward(next_state_batch)
next_Q = self.critic_target.forward(next_state_batch,
next_actions.detach())
expected_Q = reward_batch + self.gamma * next_Q
# Update Critic network
q_loss = F.mse_loss(curr_Q, expected_Q.detach())
self.critic_optimizer.zero_grad()
q_loss.backward()
self.critic_optimizer.step()
# Update Actor network
policy_loss = -self.critic.forward(
state_batch, self.actor.forward(state_batch)).mean()
self.actor_optimizer.zero_grad()
policy_loss.backward()
self.actor_optimizer.step()
# Update Actor and Critic target networks
for target_param, param in zip(self.actor_target.parameters(),
self.actor.parameters()):
target_param.data.copy_(param.data * self.tau + target_param.data *
(1.0 - self.tau))
for target_param, param in zip(self.critic_target.parameters(),
self.critic.parameters()):
target_param.data.copy_(param.data * self.tau + target_param.data *
(1.0 - self.tau))