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 PPOAgent:
def __init__(self, env, gamma, _avlambda, beta, delta, init_lr,
init_clip_range, n_updates, n_epochs, n_steps, n_mini_batch):
self.env = env
self.gamma = gamma
self._avlambda = _avlambda
self.beta = beta
self.delta = delta
self.init_lr = init_lr
self.init_clip_range = init_clip_range
self.lr = self.init_lr
self.init_clip_range = self.init_clip_range
self.n_updates = n_updates
self.epochs = n_epochs
self.n_steps = n_steps
self.n_mini_batch = n_mini_batch
self.n_episodes = 0
self.episodes_rewards = []
self.mini_batch_size = self.n_steps // self.n_mini_batch
assert (self.n_steps % self.n_mini_batch == 0)
self.current_state = None
self.input_dim = self.env.observation_space.shape[0]
self.actions_dim = self.env.action_space.n
self.actor = Actor(self.input_dim, self.actions_dim).to(device)
self.critic = Critic(self.input_dim).to(device)
self.optim_critic = optim.Adam(self.critic.parameters(),
lr=self.init_lr)
self.optim_actor = optim.Adam(self.actor.parameters(), lr=self.init_lr)
def act(self, state):
action = self.actor(state).sample()
return action
def act_greedy(self, state):
action = self.actor(state).probs.argmax(dim=-1)
return action
@staticmethod
def chunked_mask(mask):
end_pts = torch.where(mask)[0]
start = 0
for end in end_pts:
yield list(range(start, end + 1))
start = end + 1
if end_pts.nelement() == 0:
yield list(range(len(mask)))
elif end_pts[-1] != len(mask) - 1:
yield list(range(end_pts[-1] + 1, len(mask)))
def sample_trajectories(self):
""" Sample trajectories with current policy"""
rewards = np.zeros((self.n_steps, ), dtype=np.float32)
actions = np.zeros((self.n_steps, ), dtype=np.int32)
done = np.zeros((self.n_steps, ), dtype=np.bool)
states = np.zeros((self.n_steps, self.input_dim), dtype=np.float32)
log_pis = np.zeros((self.n_steps, ), dtype=np.float32)
values = np.zeros((self.n_steps, ), dtype=np.float32)
tmp_rewards = []
self.current_state = self.env.reset()
writer = SummaryWriter('runs/lunar/adaptative_kl/updates100')
for t in range(self.n_steps):
with torch.no_grad():
states[t] = self.current_state
state = torch.FloatTensor(self.current_state).to(device)
pi = self.actor(state)
values[t] = self.critic(state).cpu().numpy()
a = self.act(state)
actions[t] = a.cpu().numpy()
log_pis[t] = pi.log_prob(a).cpu().numpy()
new_state, r, end, info = self.env.step(actions[t])
rewards[t] = r / 100.
done[t] = False if info.get("TimeLimit.truncated") else end
tmp_rewards.append(r)
if end:
if info.get("TimeLimit.truncated"):
print(f"Episode {self.n_episodes} Reward: ",
np.sum(tmp_rewards), " TimeLimit.truncated !")
else:
print(f"Episode {self.n_episodes} Reward: ",
np.sum(tmp_rewards))
writer.add_scalar("Reward", np.sum(tmp_rewards),
self.n_episodes)
self.n_episodes += 1
self.episodes_rewards.append(np.sum(tmp_rewards))
if (self.n_episodes + 1) % 100 == 0:
print(
f"Total Episodes: {self.n_episodes + 1}, "
f"Mean Rewards of last 100 episodes: {np.mean(self.episodes_rewards[-100:]):.2f}"
)
tmp_rewards = []
new_state = self.env.reset()
self.current_state = new_state
states = torch.tensor(states, dtype=torch.float32, device=device)
actions = torch.tensor(actions, device=device)
values = torch.tensor(values, device=device)
rewards = torch.tensor(rewards, device=device)
done = torch.tensor(done, device=device)
log_pis = torch.tensor(log_pis, device=device)
advantages = self.compute_gae(rewards, values, done)
return states, actions, values, log_pis, advantages
def compute_gae(self, rewards, values, dones):
"""Compute Generalized Advantage Estimates"""
gaes = []
last_value = self.critic(
torch.FloatTensor(self.current_state).to(device))
next_values = torch.cat([values[1:], last_value.detach()])
for chunk in self.chunked_mask(dones):
T = rewards[chunk].size(0)
td_target = rewards[chunk] + self.gamma * (
1 - dones[chunk].int()) * next_values[chunk]
td_error = td_target - values[chunk]
discount = ((self.gamma *
self._avlambda)**torch.arange(T)).to(device)
gae = torch.tensor([(discount[:T - t] * td_error[t:]).sum()
for t in range(T)],
dtype=torch.float32).to(device)
gaes.append(gae)
return torch.cat(gaes)
def optimize_adaptative_kl(self, states, actions, values, log_pis,
advantages, learning_rate, clip_range):
old_pis = self.actor(states)
old_probs = old_pis.probs.detach()
for _ in tqdm(range(self.epochs)):
# shuffle for each epoch
indexes = torch.randperm(self.n_steps)
# for each mini batch
for start in range(0, self.n_steps, self.mini_batch_size):
# get mini batch
idx = indexes[start:start + self.mini_batch_size]
# compute loss
lambda_returns = values[idx] + advantages[idx]
A_old = (advantages[idx] - advantages[idx].mean()) / (
advantages[idx].std() + 1e-8)
new_pis = self.actor(states[idx])
new_values = self.critic(states[idx])
# policy loss (with entropy)
new_log_pis = new_pis.log_prob(actions[idx])
L = ((new_log_pis - log_pis[idx]).exp() * A_old).mean()
kl = kl_divergence(Categorical(old_probs[idx]), new_pis).mean()
entropy = new_pis.entropy().mean()
policy_loss = -(L - self.beta * kl + 0.01 * entropy)
# value loss
clipped_value = values[idx] + (new_values - values[idx]).clamp(
min=-clip_range, max=clip_range)
vf_loss = torch.max((new_values - lambda_returns)**2,
(clipped_value - lambda_returns)**2)
vf_loss = vf_loss.mean()
self.optim_actor.zero_grad()
self.optim_critic.zero_grad()
for pg in self.optim_actor.param_groups:
pg['lr'] = learning_rate
for pg in self.optim_critic.param_groups:
pg['lr'] = learning_rate
policy_loss.backward()
vf_loss.backward()
clip_grad_norm_(self.actor.parameters(), max_norm=0.5)
clip_grad_norm_(self.critic.parameters(), max_norm=0.5)
self.optim_actor.step()
self.optim_critic.step()
new_pis = self.actor(states)
kl = kl_divergence(Categorical(old_probs), new_pis).mean()
if kl >= 1.5 * self.delta:
self.beta *= 2
elif kl <= self.delta / 1.5:
self.beta /= 2
def optimize_clipped(self, states, actions, values, log_pis, advantages,
learning_rate, clip_range):
for _ in tqdm(range(self.epochs)):
# shuffle for each epoch
indexes = torch.randperm(self.n_steps)
# for each mini batch
for start in range(0, self.n_steps, self.mini_batch_size):
# get mini batch
idx = indexes[start:start + self.mini_batch_size]
# compute loss
lambda_returns = values[idx] + advantages[idx]
A_old = (advantages[idx] - advantages[idx].mean()) / (
advantages[idx].std() + 1e-8)
new_pis = self.actor(states[idx])
new_values = self.critic(states[idx])
# policy loss (with entropy)
new_log_pis = new_pis.log_prob(actions[idx])
ratio = (new_log_pis - log_pis[idx]).exp()
clipped_ratio = ratio.clamp(min=1.0 - clip_range,
max=1.0 + clip_range)
Lclip = torch.min(ratio * A_old, clipped_ratio * A_old)
Lclip = Lclip.mean()
entropy = new_pis.entropy().mean()
policy_loss = -(Lclip + 0.01 * entropy)
# value loss
clipped_value = values[idx] + (new_values - values[idx]).clamp(
min=-clip_range, max=clip_range)
vf_loss = torch.max((new_values - lambda_returns)**2,
(clipped_value - lambda_returns)**2)
vf_loss = vf_loss.mean()
self.optim_actor.zero_grad()
self.optim_critic.zero_grad()
for pg in self.optim_actor.param_groups:
pg['lr'] = learning_rate
for pg in self.optim_critic.param_groups:
pg['lr'] = learning_rate
policy_loss.backward()
vf_loss.backward()
clip_grad_norm_(self.actor.parameters(), max_norm=0.5)
clip_grad_norm_(self.critic.parameters(), max_norm=0.5)
self.optim_actor.step()
self.optim_critic.step()
def optimize_without(self, states, actions, values, log_pis, advantages,
learning_rate, clip_range):
old_pis = self.actor(states)
old_probs = old_pis.probs.detach()
for _ in tqdm(range(self.epochs)):
# shuffle for each epoch
indexes = torch.randperm(self.n_steps)
# for each mini batch
for start in range(0, self.n_steps, self.mini_batch_size):
# get mini batch
idx = indexes[start:start + self.mini_batch_size]
# compute loss
lambda_returns = values[idx] + advantages[idx]
A_old = (advantages[idx] - advantages[idx].mean()) / (
advantages[idx].std() + 1e-8)
new_pis = self.actor(states[idx])
new_values = self.critic(states[idx])
# policy loss (with entropy)
new_log_pis = new_pis.log_prob(actions[idx])
L = ((new_log_pis - log_pis[idx]).exp() * A_old).mean()
entropy = new_pis.entropy().mean()
policy_loss = -(L + 0.01 * entropy)
# value loss
clipped_value = values[idx] + (new_values - values[idx]).clamp(
min=-clip_range, max=clip_range)
vf_loss = torch.max((new_values - lambda_returns)**2,
(clipped_value - lambda_returns)**2)
vf_loss = vf_loss.mean()
self.optim_actor.zero_grad()
self.optim_critic.zero_grad()
for pg in self.optim_actor.param_groups:
pg['lr'] = learning_rate
for pg in self.optim_critic.param_groups:
pg['lr'] = learning_rate
policy_loss.backward()
vf_loss.backward()
clip_grad_norm_(self.actor.parameters(), max_norm=0.5)
clip_grad_norm_(self.critic.parameters(), max_norm=0.5)
self.optim_actor.step()
self.optim_critic.step()
new_pis = self.actor(states)
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)
class ActorCriticAgent(object):
"""Actor-Critic agent"""
def __init__(self, input_dim, n_actions, gamma, lr, beta, _lambda, _avlambda, target_freq):
self.gamma = gamma
self.lr = lr
self.beta = beta
self._lambda = _lambda
self._avlambda = _avlambda
self.input_dim = input_dim
self.n_actions = n_actions
self.actor = Actor(self.input_dim, self.n_actions).to(device)
self.critic = Critic(self.input_dim).to(device)
self.critic_target = deepcopy(self.critic)
self.actor_optim = torch.optim.Adam(self.actor.parameters(), lr=lr)
self.critic_optim = torch.optim.Adam(self.critic.parameters(), lr=lr)
self.transitions = []
self.target_freq = target_freq
self.learn_step = 0
self.event_count = 0
def act(self, state):
action = self.actor(state).sample()
return action.item()
def act_greedy(self, state):
action = self.actor(state).probs.argmax(dim=-1)
return action.item()
def store_transition(self, transition):
if len(self.transitions) == 0:
self.transitions.append([])
self.transitions[-1].append(transition)
@staticmethod
def _get_transitions(transitions):
states, actions, rewards, next_states, dones = [], [], [], [], []
for transition in transitions:
states.append(transition[0])
actions.append(transition[1])
rewards.append(transition[2] / 100.0)
next_states.append(transition[3])
dones.append(int(transition[4]))
return (
torch.stack(states),
torch.tensor(actions, dtype=torch.int64).to(device),
torch.FloatTensor(rewards).to(device),
torch.stack(next_states),
torch.IntTensor(dones).to(device),
)
def get_transitions(self, shuffle=True):
g_states, g_actions, g_rewards, g_next_states, g_dones, g_gaes = [], [], [], [], [], []
for transitions in self.transitions:
# get transitions for current trajectory
states, actions, rewards, next_states, dones = self._get_transitions(transitions)
# compute GAE for each trajectory
gaes = self.compute_gae(rewards, states, next_states, 1-dones)
g_states.append(states)
g_actions.append(actions)
g_rewards.append(rewards)
g_next_states.append(next_states)
g_dones.append(dones)
g_gaes.append(gaes)
self.transitions = []
if shuffle:
idx = torch.randperm(torch.cat(g_actions).size(0))
else:
idx = torch.arange(torch.cat(g_actions).size(0))
return (
torch.cat(g_states)[idx],
torch.cat(g_actions)[idx],
torch.cat(g_rewards)[idx],
torch.cat(g_next_states)[idx],
torch.cat(g_dones)[idx],
torch.cat(g_gaes)[idx],
)
def compute_gae(self, rewards, states, next_states, masks):
T = rewards.size(0)
values = self.critic(states).view(-1)
next_values = self.critic(next_states).view(-1)
td_target = rewards + self.gamma * masks * next_values
td_error = td_target - values
discount = ((self.gamma * self._avlambda) ** torch.arange(T)).to(device)
gae = torch.Tensor([(discount[:T - t] * td_error[t:]).sum() for t in range(T)]).to(device)
return gae.detach()
def optimize(self):
self.learn_step += 1
states, actions, rewards, next_states, dones, gaes = self.get_transitions()
lambda_returns = self.critic(states).view(-1) + gaes
# value loss
value_loss = F.mse_loss(self.critic(states).view(-1), lambda_returns.detach())
# policy loss
dist = self.actor(states)
log_probs = dist.log_prob(actions)
entropy = dist.entropy().mean()
policy_loss = - (log_probs * gaes.detach()).mean() - self.beta * entropy
# backpropagation
self.critic_optim.zero_grad()
value_loss.backward()
clip_grad_norm_(self.critic.parameters(), 0.1)
self.critic_optim.step()
self.actor_optim.zero_grad()
policy_loss.backward()
self.actor_optim.step()
clip_grad_norm_(self.actor.parameters(), 0.1)
return policy_loss, value_loss, gaes, entropy