class Agent():
"""Interacts with and learns from the environment."""
def __init__(self,
state_size,
action_size,
random_seed,
fc1_units,
fc2_units,
weighted=False,
individual=False):
"""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)
self.epsilon = EPSILON_MAX
# Actor Network (w/ Target Network)
if weighted:
self.actor_local = Weight_adapter(state_size,
action_size).to(device)
self.actor_target = Weight_adapter(state_size,
action_size).to(device)
elif individual:
self.actor_local = IndividualModel(state_size, action_size,
random_seed,
fc1_units).to(device)
self.actor_target = IndividualModel(state_size, action_size,
random_seed,
fc1_units).to(device)
else:
self.actor_local = Actor(state_size, action_size, random_seed,
fc1_units, fc2_units).to(device)
self.actor_target = Actor(state_size, action_size, random_seed,
fc1_units, 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).to(device)
self.critic_target = Critic(state_size, action_size,
random_seed).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
# Noise process
self.noise = OUNoise(action_size,
random_seed,
mu=0,
theta=0.15,
sigma=0.2)
# 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_update(self.actor_target, self.actor_local)
self.hard_update(self.critic_target, self.critic_local)
self.t_step = 0
def step(self, state, action, reward, next_state, done, timestep):
"""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)
if len(self.memory) > LEARN_START:
# 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:
for _ in range(UPDATES_PER_STEP):
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()
#print(action)
self.actor_local.train()
if add_noise:
tem_noise = self.noise.sample()
action += self.epsilon * tem_noise
# print(tem_noise, np.clip(action, -1, 1))
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 + ? * 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)
Q_targets_next = self.critic_target(next_states, actions_next)
# 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(states, actions)
critic_loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
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 ---------------------------- #
# Compute actor loss
actions_pred = self.actor_local(states)
actor_loss = -self.critic_local(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, TAU)
self.soft_update(self.actor_local, self.actor_target, TAU)
# ---------------------------- update noise ---------------------------- #
if self.epsilon - EPSILON_DECAY > EPSILON_MIN:
self.epsilon -= EPSILON_DECAY
else:
self.epsilon = EPSILON_MIN
self.noise.reset()
def soft_update(self, local_model, target_model, tau):
"""Soft update model parameters.
?_target = t*?_local + (1 - t)*?_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_update(self, target, source):
for target_param, param in zip(target.parameters(),
source.parameters()):
target_param.data.copy_(param.data)
class Agent():
"""
The Agent interact with and learn from the Environment
"""
def __init__(self, state_size, action_size, num_agents, random_seed):
"""
Initialize Agent Object.
----------------------------------------
Parameters
----------------------------------------
state_size (int): dimension of each state
action_size (int): dimension of each action
num_agents (int): number of agents
random_seed(int): random seed
"""
self.state_size = state_size
self.action_size = action_size
self.num_agents = num_agents
self.seed = random.seed(random_seed)
self.eps = EPS_START
self.eps_decay = 1 / (EPS_EP_END * LEARN_NUM)
self.timestep = 0
# <--------------- Actor 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 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=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
# <--------------- Noise --------------->
self.noise = OUNoise((num_agents, action_size), random_seed)
# <----------- Replay Memory ------------->
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE,
random_seed)
def step(self, state, action, reward, next_state, done, agent_number):
"""
Save Experience in Replay Memory and select randomly from the buffer to learn """
self.timestep += 1
# <-------Save Experience --------->
self.memory.add(state, action, reward, next_state, done)
# <-------- Learn at given interval, if enough samples are available in the memory ----------->
if len(self.memory) > BATCH_SIZE and self.timestep % LEARN_EVERY == 0:
for _ in range(LEARN_NUM):
experiences = self.memory.sample()
self.learn(experiences, GAMMA, agent_number)
# <-------Obtain Action------------>
def act(self, states, add_noise):
"""
Returns Actions for both agents given their respective states and based on the current Policy
"""
states = torch.from_numpy(states).float().to(device)
actions = np.zeros((self.num_agents, self.action_size))
self.actor_local.eval()
with torch.no_grad():
# Obtain action for each agent and concatenate them
for agent_num, state in enumerate(states):
action = self.actor_local(state).cpu().data.numpy()
actions[agent_num, :] = action
self.actor_local.train()
#add noise to actions
if add_noise:
actions += self.eps * self.noise.sample()
actions = np.clip(actions, -1, 1)
return actions
def reset(self):
self.noise.reset()
def learn(self, experiences, gamma, agent_number):
"""
Update the policy and Value Parameters using given batch of experience tuples
Q_targets = r + y * critic_target(next_state, actor_target(next_state))
actor_target(state) --- Action
critic_target(state, action) --- Q-Value
-----------------------------------------
Parameters
-----------------------------------------
experiences (Tuple[torch.Tensor]) -- tuple(s,a,r,s',done)
gamma (float) -- discount factor
"""
states, actions, rewards, next_states, dones = experiences
# <----------------------- Update the Critic -------------------->
#Get predicted next-state actions and Q-values from target models
actions_next = self.actor_target(next_states)
# Construct Next actions vector relative to the agent
if agent_number == 0:
actions_next = torch.cat((actions_next, actions[:, 2:]), dim=1)
else:
actions_next = torch.cat((actions[:, :2], actions_next), dim=1)
# Compute Q tarets for current states (y_i)
Q_targets_next = self.critic_target(next_states, actions_next)
Q_targets = rewards + (gamma * Q_targets_next * (1 - dones))
# Compute Critic Loss
Q_expected = self.critic_local(states, actions)
critic_loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the Loss
self.critic_optimizer.zero_grad()
critic_loss.backward()
torch.nn.utils.clip_grad_norm_(self.critic_local.parameters(), 1)
self.critic_optimizer.step()
# <----------------------- Update the Actor -------------------->
# Compute Loss
actions_pred = self.actor_local(states)
# Construct action prediction Vector relative to reach agent
if agent_number == 0:
actions_pred = torch.cat((actions_pred, actions[:, 2:]), dim=1)
else:
actions_pred = torch.cat((actions[:, :2], actions_pred), dim=1)
# Compute Actor Loss
actor_loss = -self.critic_local(states, actions_pred).mean()
# Minimize the Loss
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# <----------------------- Update the Target Networks -------------------->
self.soft_update(self.critic_local, self.critic_target, TAU)
self.soft_update(self.actor_local, self.actor_target, TAU)
# <----------------------- Update the noise -------------------->
self.eps -= self.eps_decay
self.eps = max(self.eps, EPS_FINAL)
self.noise.reset()
# <----------------------- Perform Soft Update -------------------->
def soft_update(self, local_model, target_model, tau):
"""
Soft Update model parameters
θ_target = τ*θ_local + (1 - τ)*θ_target
---------------------------
Parameters
---------------------------
local_model: Weights will be copied fron this pytorch model
target_model: weights will be copied to this pytorch model
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 DDPG:
def __init__(self,
env,
tau=1e-3,
gamma=0.99,
batch_size=64,
depsilon=50000):
self.num_states = env.observation_space.shape[0]
self.num_actions = env.action_space.shape[0]
self.policy = Actor(self.num_states, self.num_actions).train()
self.policy_target = Actor(self.num_states, self.num_actions).eval()
self.hard_update(self.policy, self.policy_target)
self.critic = Critic(self.num_states, self.num_actions).train()
self.critic_target = Critic(self.num_states, self.num_actions).eval()
self.hard_update(self.critic, self.critic_target)
self.critic_loss = nn.MSELoss()
self.batch_size = batch_size
self.gamma = gamma
self.tau = tau
self.epsilon = 1.0
self.depsilon = 1.0 / float(depsilon)
self.opt_critic = torch.optim.Adam(self.critic.parameters(), lr=1e-3)
self.opt_policy = torch.optim.Adam(self.policy.parameters(), lr=1e-4)
self.policy.cuda()
self.policy_target.cuda()
self.critic.cuda()
self.critic_target.cuda()
def train(self, buffer):
b_state, b_action, b_reward, b_state_next, b_term = buffer.sample(
self.batch_size)
with torch.no_grad():
action_target = self.policy_target(b_state_next)
Q_prime = self.critic_target(b_state_next, action_target)
self.opt_critic.zero_grad()
Q = self.critic(b_state, b_action)
L_critic = self.critic_loss(
Q, b_reward + self.gamma * Q_prime * (1.0 - b_term))
L_critic.backward()
self.opt_critic.step()
self.opt_policy.zero_grad()
action = self.policy(b_state)
L_Q = -1.0 * self.critic(b_state, action).mean()
L_Q.backward()
self.opt_policy.step()
self.soft_update(self.critic, self.critic_target)
self.soft_update(self.policy, self.policy_target)
return L_critic.item(), L_Q.item()
def get_entropy(self, buffer, m=5, n=100):
# b_state, b_action, b_reward, b_state_next, b_term = buffer.sample(n)
b_angle = torch.rand(n) * np.pi * 2.0
b_speed = 2.0 * (torch.rand(n) - 0.5) * 8.0
b_state = torch.stack(
[torch.cos(b_angle),
torch.sin(b_angle), b_speed], dim=1).to(device='cuda',
dtype=torch.float32)
coef = torch.zeros(n, dtype=b_state.dtype, device=b_state.device)
with torch.no_grad():
action = self.policy(b_state)
X, ind = torch.sort(action, dim=0)
for i in range(n):
if i < m:
c = 1
a = X[i + m]
b = X[0]
elif i >= m and i < n - m:
c = 2
a = X[i + m]
b = X[i - m]
else:
c = 1
a = X[n - 1]
b = X[i - m]
coef[i] = float(n) * float(c) / float(m) * (a - b + 1E-5)
S = torch.log(coef).mean()
return S.item()
def get_value(self, state, action):
with torch.no_grad():
return self.critic(state, action).item()
def select_action(self, state, random_process):
with torch.no_grad():
action = self.policy(state)
noise = max(self.epsilon, 0.0) * random_process.sample()
self.epsilon -= self.depsilon
action += torch.from_numpy(noise).to(device=action.device,
dtype=action.dtype)
action = torch.clamp(action, -1, 1)
return action
def random_action(self):
m = Uniform(torch.tensor([-1.0 for i in range(self.num_actions)]),
torch.tensor([1.0 for i in range(self.num_actions)]))
return m.sample()
def soft_update(self, src, dst):
with torch.no_grad():
for src_param, dst_param in zip(src.parameters(),
dst.parameters()):
dst_param.copy_(self.tau * src_param +
(1.0 - self.tau) * dst_param)
def hard_update(self, src, dst):
with torch.no_grad():
for src_param, dst_param in zip(src.parameters(),
dst.parameters()):
dst_param.copy_(src_param.clone())
def load_weights(self, path):
self.actor.load_state_dict(torch.load('{}/actor.pkl'.format(path)))
self.critic.load_state_dict(torch.load('{}/critic.pkl'.format(path)))
def save_model(self, path):
torch.save(self.actor.state_dict(), '{}/actor.pkl'.format(output))
torch.save(self.critic.state_dict(), '{}/critic.pkl'.format(output))
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 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=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
self.hard_copy_weights(self.actor_target, self.actor_local)
self.hard_copy_weights(self.critic_target, self.critic_local)
# Noise process
self.noise = OUNoise(action_size, random_seed)
# Replay memory
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE,
random_seed)
def hard_copy_weights(self, target, source):
""" copy weights from source to target network (part of initialization)"""
for target_param, param in zip(target.parameters(),
source.parameters()):
target_param.data.copy_(param.data)
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, 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 += 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)
Q_targets_next = self.critic_target(next_states, actions_next)
# 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(states, actions)
critic_loss = F.mse_loss(Q_expected, Q_targets)
# Minimize the loss
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 ---------------------------- #
# Compute actor loss
actions_pred = self.actor_local(states)
actor_loss = -self.critic_local(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, 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)
class Agent():
def __init__(self, state_size, action_size, random_seed, num_agents=1):
self.state_size = state_size
self.action_size = action_size
self.seed = random.seed(random_seed)
self.num_agents = num_agents
#Raw and Targer Actor 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)
##copying the weights of the raw network to the target network
for target, local in zip(self.actor_target.parameters(),
self.actor_local.parameters()):
target.data.copy_(local.data)
#Raw and Target CRITIC Network
self.critic_local = Critic(state_size * num_agents,
action_size * num_agents,
random_seed).to(device)
self.critic_target = Critic(state_size * num_agents,
action_size * num_agents,
random_seed).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
##copying the weights of the raw network to the target network
for target, local in zip(self.critic_target.parameters(),
self.critic_local.parameters()):
target.data.copy_(local.data)
##Creating Noise Process
self.noise = OrnUhlNoise(action_size, random_seed)
###Replay Memory; in MADDPG, the replay buffer is common to all agents
self.memory = ReplayBuffer(action_size, BUFFER_SIZE, BATCH_SIZE,
random_seed)
def step(self, state, action, reward, next_state, done):
"""Shared Memory; to save experiences in replay memory, and use random sample from buffer to learn"""
#check if this is accurately implemented in training the MADDPG agent
self.memory.add(state, action, reward, next_state, done)
#start learning if the buffer size is full
if len(self.memory) > BATCH_SIZE:
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, state, noise=0.0):
"""uses current ploicy to output the next action"""
''' Please understand the below code snippet in detail '''
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_OU_NOISE:
action += self.noise.sample() * noise
return np.clip(action, -1, 1)
def reset(self):
self.noise.reset()
def learn(self, experiences, gamma):
''' only used in traininf DDPG agent, not for MADDPG'''
#updates policy and value params using a givrn batch of experience tuples
states, actions, rewards, next_states, dones = experiences
#################update critic################################
next_actions = self.actor_target(next_states)
next_Q_targets = self.critic_target(next_states, next_actions)
Q_targets = rewards + (gamma * next_Q_targets *
(1 - dones)) #Q targets for current states
Q_Expected = self.critic_local(states, actions)
critic_loss = F.mse_loss(Q_Expected, Q_targets)
#minimizing the loss
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
###############update actor##################################
##computing actions_loss
actions_pred = self.actor_local(states)
actor_loss = -self.critic_local(states, actions_pred).mean()
##minimizing 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, TAU)
self.soft_update(self.actor_local, self.actor_target, TAU)
def soft_update(self, local_model, target_model, tau):
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)