def __init__(self, task):
self.task = task
self.state_size = task.state_size
self.action_size = task.action_size
self.action_low = task.action_low
self.action_high = task.action_high
self.MU = 0
self.THETA = 0.15
self.SIGMA = 0.10
self.GAMMA = 0.99
self.TAU = 0.001
self.BATCHS = 256
self.MAX_REWARD = -999999999
self.actor_local = Actor(self.state_size, self.action_size,
self.action_low, self.action_high)
self.actor_target = Actor(self.state_size, self.action_size,
self.action_low, self.action_high)
self.actor_target.model.set_weights(
self.actor_local.model.get_weights())
self.critic_local = Critic(self.state_size, self.action_size)
self.critic_target = Critic(self.state_size, self.action_size)
self.critic_target.model.set_weights(
self.critic_local.model.get_weights())
self.noiseObj = Noise(self.action_size, self.MU, self.THETA,
self.SIGMA)
self.replayObj = Replay(self.BATCHS)
def __init__(self, task):
self.task = task
self.state_size = task.state_size
self.action_size = task.action_size
self.action_low = task.action_low
self.action_high = task.action_high
# Actor (Policy) Model
self.actor_local = Actor(self.state_size,
self.action_size,
self.action_low,
self.action_high)
self.actor_target = Actor(self.state_size,
self.action_size,
self.action_low,
self.action_high)
# Critic (Value) Model
self.critic_local = Critic(self.state_size, self.action_size)
self.critic_target = Critic(self.state_size, self.action_size)
# Initialize target model parameters with local model parameters
self.critic_target.model.set_weights(
self.critic_local.model.get_weights())
self.actor_target.model.set_weights(
self.actor_local.model.get_weights())
# Noise process
self.exploration_mu = 0.0 #0
self.exploration_theta = 0.125 # 0.14 | 0.1
self.exploration_sigma = 0.0009 # 0.001 | 0.2 | 0.001
self.noise = OUNoise(self.action_size,
self.exploration_mu,
self.exploration_theta,
self.exploration_sigma)
# Replay memory
self.buffer_size = 100000
self.batch_size = 64
self.memory = ReplayBuffer(self.buffer_size,
self.batch_size)
# Algorithm parameters
self.gamma = 0.998 # 0.99 | 0.9 | discount factor
self.tau = 0.099 # 0.001| 0.01 | 0.1 | 0.05 | for soft update of target parameters
# Score tracker
self.best_score = -np.inf
self.score = 0
def __init__(self,state_size, action_size,single_rotor_control=False, buffer_size=int(1e5), batch_size=128,
gamma=0.98, tau=1e-3, lr_actor=1e-4,lr_critic=1e-3, random_seed=0):
"""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
# rotor control mode
self.single_rotor_control = single_rotor_control
if self.single_rotor_control:
action_size = 1
# define hyperparameters
self.buffer_size = buffer_size
self.batch_size = batch_size
self.gamma = gamma
self.tau = tau
self.lr_actor = lr_actor
self.lr_critic = lr_critic
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)
# Noise process
self.noise = OUNoise(size=action_size, seed=random_seed)
# Replay memory
self.memory = ReplayBuffer(buffer_size, random_seed)
# counter for time steps
self.time_step = 0
def __init__(self, task, gym=False):
self.task = task
if gym:
self.state_size = np.prod(task.observation_space.shape)
self.action_size = np.prod(task.action_space.shape)
self.action_low = task.action_space.low
self.action_high = task.action_space.high
else:
self.state_size = task.state_size
self.action_size = task.action_size
self.action_low = task.action_low
self.action_high = task.action_high
self.alpha = task.alpha
self.beta = task.beta
# Actor (Policy) Model
self.actor_local = Actor(self.state_size, self.action_size, self.action_low, self.action_high, self.alpha)
self.actor_target = Actor(self.state_size, self.action_size, self.action_low, self.action_high, self.alpha)
# Critic (Value) Model
self.critic_local = Critic(self.state_size, self.action_size, self.beta)
self.critic_target = Critic(self.state_size, self.action_size, self.beta)
# Initialize target model parameters with local model parameters
self.critic_target.model.set_weights(self.critic_local.model.get_weights())
self.actor_target.model.set_weights(self.actor_local.model.get_weights())
# Noise process
self.exploration_mu = 0 # changed from 0
self.exploration_theta = 0.15 # changed from 0.15
self.exploration_sigma = 3.2 # changed from 0.2
self.noise = OUNoise(self.action_size, self.exploration_mu, self.exploration_theta, self.exploration_sigma)
# Replay memory
self.buffer_size = 1000 # was 100000 originally
self.batch_size = 64
self.memory = ReplayBuffer(self.buffer_size, self.batch_size)
# Algorithm parameters
self.gamma = 0.99 # discount factor, 0.99
self.tau = .001 # for soft update of target parameters, 0.01 originally
def __init__(self, task):
self.task = task
self.state_size = task.state_size
self.action_size = task.action_size
self.action_low = task.action_low
self.action_high = task.action_high
# Actor (Policy) Model
self.actor_local = Actor(self.state_size, self.action_size,
self.action_low, self.action_high)
self.actor_target = Actor(self.state_size, self.action_size,
self.action_low, self.action_high)
# Critic (Value) Model
self.critic_local = Critic(self.state_size, self.action_size)
self.critic_target = Critic(self.state_size, self.action_size)
# Initialize target model parameters with local model parameters
self.critic_target.model.set_weights(
self.critic_local.model.get_weights())
self.actor_target.model.set_weights(
self.actor_local.model.get_weights())
# Noise process
self.exploration_mu = 0.0
self.exploration_theta = 0.15 #0.15
self.exploration_sigma = 0.25 #0.2
self.noise = OUNoise(self.action_size, self.exploration_mu,
self.exploration_theta, self.exploration_sigma)
# Replay memory
self.buffer_size = 100000
self.batch_size = 16
self.memory = ReplayBuffer(self.buffer_size, self.batch_size)
# Algorithm parameters
self.gamma = 0.99 # discount factor
self.tau = 0.01 # for soft update of target parameters
class DDPG():
"""Reinforcement Learning agent that learns using DDPG."""
def __init__(self, task):
self.task = task
self.state_size = task.state_size
self.action_size = task.action_size
self.action_low = task.action_low
self.action_high = task.action_high
# Actor (Policy) Model
self.actor_local = Actor(self.state_size, self.action_size, self.action_low, self.action_high)
self.actor_target = Actor(self.state_size, self.action_size, self.action_low, self.action_high)
# Critic (Value) Model
self.critic_local = Critic(self.state_size, self.action_size)
self.critic_target = Critic(self.state_size, self.action_size)
# Initialize target model parameters with local model parameters
self.critic_target.model.set_weights(self.critic_local.model.get_weights())
self.actor_target.model.set_weights(self.actor_local.model.get_weights())
# Noise process
self.exploration_mu = 0
self.exploration_theta = 0.15
self.exploration_sigma = 0.2
self.noise = OUNoise(self.action_size, self.exploration_mu, self.exploration_theta, self.exploration_sigma)
# Replay memory
self.buffer_size = 100000
self.batch_size = 64
self.memory = ReplayBuffer(self.buffer_size, self.batch_size)
# Algorithm parameters
self.gamma = 0.9 # discount factor
self.tau = 0.001 # for soft update of target parameters
self.total_reward = 0
self.count = 0
self.best_score = -np.inf
self.score = 0
def reset_episode(self):
self.total_reward = 0
self.count = 0
self.score = 0
self.noise.reset()
state = self.task.reset()
self.last_state = state
return state
def step(self, action, reward, next_state, done):
self.total_reward += reward
self.count += 1
# Save experience / reward
self.memory.add(self.last_state, action, reward, next_state, done)
# Learn, if enough samples are available in memory
if len(self.memory) > self.batch_size:
experiences = self.memory.sample()
self.learn(experiences)
# Roll over last state and action
self.last_state = next_state
def act(self, states):
"""Returns actions for given state(s) as per current policy."""
states = np.reshape(states, [-1, self.state_size])
action = self.actor_local.model.predict(states)[0]
return list(action + self.noise.sample()) # add some noise for exploration
def learn(self, experiences):
"""Update policy and value parameters using given batch of experience tuples."""
self.score = self.total_reward / float(self.count) if self.count else 0.0
if self.score > self.best_score:
self.best_score = self.score
# Convert experience tuples to separate arrays for each element (states, actions, rewards, etc.)
states = np.vstack([e.state for e in experiences if e is not None])
actions = np.array([e.action for e in experiences if e is not None]).astype(np.float32).reshape(-1, self.action_size)
rewards = np.array([e.reward for e in experiences if e is not None]).astype(np.float32).reshape(-1, 1)
dones = np.array([e.done for e in experiences if e is not None]).astype(np.uint8).reshape(-1, 1)
next_states = np.vstack([e.next_state for e in experiences if e is not None])
# Get predicted next-state actions and Q values from target models
# Q_targets_next = critic_target(next_state, actor_target(next_state))
actions_next = self.actor_target.model.predict_on_batch(next_states)
Q_targets_next = self.critic_target.model.predict_on_batch([next_states, actions_next])
# Compute Q targets for current states and train critic model (local)
Q_targets = rewards + self.gamma * Q_targets_next * (1 - dones)
self.critic_local.model.train_on_batch(x=[states, actions], y=Q_targets)
# Train actor model (local)
action_gradients = np.reshape(self.critic_local.get_action_gradients([states, actions, 0]), (-1, self.action_size))
self.actor_local.train_fn([states, action_gradients, 1]) # custom training function
# Soft-update target models
self.soft_update(self.critic_local.model, self.critic_target.model)
self.soft_update(self.actor_local.model, self.actor_target.model)
def soft_update(self, local_model, target_model):
"""Soft update model parameters."""
local_weights = np.array(local_model.get_weights())
target_weights = np.array(target_model.get_weights())
assert len(local_weights) == len(target_weights), "Local and target model parameters must have the same size"
new_weights = self.tau * local_weights + (1 - self.tau) * target_weights
target_model.set_weights(new_weights)
class Agent():
"""Interacts with and learns from the environment."""
def __init__(self,state_size, action_size,single_rotor_control=False, buffer_size=int(1e5), batch_size=128,
gamma=0.98, tau=1e-3, lr_actor=1e-4,lr_critic=1e-3, random_seed=0):
"""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
# rotor control mode
self.single_rotor_control = single_rotor_control
if self.single_rotor_control:
action_size = 1
# define hyperparameters
self.buffer_size = buffer_size
self.batch_size = batch_size
self.gamma = gamma
self.tau = tau
self.lr_actor = lr_actor
self.lr_critic = lr_critic
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)
# Noise process
self.noise = OUNoise(size=action_size, seed=random_seed)
# Replay memory
self.memory = ReplayBuffer(buffer_size, random_seed)
# counter for time steps
self.time_step = 0
#self.soft_update(self.critic_local, self.critic_target, 1.0)
#self.soft_update(self.actor_local, self.actor_target, 1.0)
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
if self.single_rotor_control:
self.memory.add(state, action[0], reward, next_state, done)
else:
self.memory.add(state, action, reward, next_state, done)
# Learn, if enough samples are available in memory
if len(self.memory) > self.batch_size:
experiences = self.memory.sample(self.batch_size)
self.learn(experiences, self.gamma)
# increase time step count
self.time_step+=1
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():
actions = self.actor_local(states.unsqueeze(0)).cpu().data.numpy()[0]
self.actor_local.train()
if add_noise:
#noise = np.repeat(,self.action_size)
actions += self.noise.sample()
if self.single_rotor_control:
actions = np.repeat(actions,self.action_size)
return np.clip(actions, -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, idxs, is_weights = 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 Q value predictions
Q_expected = self.critic_local(states, actions)
# compute td error
td_error = Q_targets - Q_expected
# update td error in Replay buffer
self.memory.update_priorities(idxs,td_error.detach().cpu().numpy().squeeze())
# compute critic loss
critic_loss = ((is_weights*td_error)**2).mean()
# Minimize the loss
self.critic_optimizer.zero_grad()
critic_loss.backward()
self.critic_optimizer.step()
# ---------------------------- update actor ---------------------------- #
# Compute actor loss
actions_pred = self.actor_local(states)
actor_loss = -(is_weights * self.critic_local(states, actions_pred).squeeze()).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, 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 Agent():
def __init__(self, task):
self.task = task
self.state_size = task.state_size
self.action_size = task.action_size
self.action_low = task.action_low
self.action_high = task.action_high
# Actor Policy Model
self.actor_local = Actor(self.state_size, self.action_size,
self.action_low, self.action_high)
self.actor_target = Actor(self.state_size, self.action_size,
self.action_low, self.action_high)
# Critic Value Model
self.critic_local = Critic(self.state_size, self.action_size)
self.critic_target = Critic(self.state_size, self.action_size)
self.critic_target.model.set_weights(
self.critic_local.model.get_weights())
self.actor_target.model.set_weights(
self.actor_local.model.get_weights())
# Noise process
self.exploration_mu = 0.0
self.exploration_theta = 0.15 #0.15
self.exploration_sigma = 0.25 #0.2
self.noise = OUNoise(self.action_size, self.exploration_mu,
self.exploration_theta, self.exploration_sigma)
#Replay Memory
self.buffer_size = 100000
self.batch_size = 16
self.memory = ReplayBuffer(self.buffer_size, self.batch_size)
# Algorithm parameters
self.gamma = 0.99
self.tau = 0.01
def reset_episode(self):
#reset of the episode
self.noise.reset()
state = self.task.reset()
self.last_state = state
return state
def step(self, action, reward, next_state, done):
#Save experience/reward
self.memory.add(self.last_state, action, reward, next_state, done)
# Learn if enough samples are available in memory
if len(self.memory) > self.batch_size:
experiences = self.memory.sample()
self.learn(experiences)
#last state and action
self.last_state = next_state
def act(self, state):
#Returns actions for given state(s) as per current policy
state = np.reshape(state, [-1, self.state_size])
action = self.actor_local.model.predict(state)[0]
#add some noise for exploration
return list(action + self.noise.sample())
def learn(self, experiences):
#Update policy and value parameters using given batch of experience tuples
# Convert experience tuples to separate arrays for each element (states, actions, rewards, etc.)
states = np.vstack([e.state for e in experiences if e is not None])
actions = np.array([e.action for e in experiences
if e is not None]).astype(np.float32).reshape(
-1, self.action_size)
rewards = np.array([e.reward for e in experiences if e is not None
]).astype(np.float32).reshape(-1, 1)
dones = np.array([e.done for e in experiences
if e is not None]).astype(np.uint8).reshape(-1, 1)
next_states = np.vstack(
[e.next_state for e in experiences if e is not None])
#Get predicted next-state actions and Q values from target models
#Q_targets_next = critic_target(next_state, actor_target(next_state))
actions_next = self.actor_target.model.predict_on_batch(next_states)
Q_targets_next = self.critic_target.model.predict_on_batch(
[next_states, actions_next])
#Compute Q targets for current states and train critic model (local)
Q_targets = rewards + self.gamma * Q_targets_next * (1 - dones)
self.critic_local.model.train_on_batch(x=[states, actions],
y=Q_targets)
#Train actor model
action_gradients = np.reshape(
self.critic_local.get_action_gradients([states, actions, 0]),
(-1, self.action_size))
#Training function
self.actor_local.train_fn([states, action_gradients, 1])
# Soft-update target models
self.soft_update(self.critic_local.model, self.critic_target.model)
self.soft_update(self.actor_local.model, self.actor_target.model)
def soft_update(self, local_model, target_model):
local_weights = np.array(local_model.get_weights())
target_weights = np.array(target_model.get_weights())
#Local model parameters and target model parameters should have the same size
assert len(local_weights) == len(target_weights)
new_weights = self.tau * local_weights + (1 -
self.tau) * target_weights
target_model.set_weights(new_weights)
class Agent(object):
def __init__(self, task):
self.task = task
self.state_size = task.state_size
self.action_size = task.action_size
self.action_low = task.action_low
self.action_high = task.action_high
#Parames
self.MU = 0
self.THETA = 0.15
self.SIGMA = 0.10
self.GAMMA = 0.99
self.TAU = 0.001
self.BATCHS = 256
self.MAX_REWARD = -999999999
#Actor Model
self.actor_local = Actor(self.state_size, self.action_size,
self.action_low, self.action_high)
self.actor_target = Actor(self.state_size, self.action_size,
self.action_low, self.action_high)
#init
self.actor_target.model.set_weights(
self.actor_local.model.get_weights())
#Critic Model
self.critic_local = Critic(self.state_size, self.action_size)
self.critic_target = Critic(self.state_size, self.action_size)
#init
self.critic_target.model.set_weights(
self.critic_local.model.get_weights())
#Noise process
self.noiseObj = Noise(self.action_size, self.MU, self.THETA,
self.SIGMA)
#Replay memory
self.replayObj = Replay(self.BATCHS)
def reset_episode(self):
self.count = 0
self.total_reward = 0
self.noiseObj.reset()
state = self.task.reset()
self.last_state = state
return (state)
def step(self, action, reward, next_state, done):
self.replayObj.add(self.last_state, action, reward, next_state, done)
self.total_reward += reward
self.count += 1
if self.total_reward > self.MAX_REWARD:
self.MAX_REWARD = self.total_reward
if len(self.replayObj) > self.BATCHS:
experiences = self.replayObj.sample()
self.learn(experiences)
self.last_state = next_state
def act(self, states):
action = self.actor_local.model.predict(
np.reshape(states, [-1, self.state_size]))[0]
return (list(action + self.noiseObj.sample()))
def learn(self, experiences):
states = np.array([e.state for e in experiences if e is not None])
actions = np.array([e.action for e in experiences
if e is not None]).reshape(-1, self.action_size)
rewards = np.array([e.reward for e in experiences
if e is not None]).reshape(-1, 1)
dones = np.array([e.done for e in experiences
if e is not None]).reshape(-1, 1)
next_state = np.array(
[e.next_state for e in experiences if e is not None])
#获取预测next_state的actions 和目标模型的Q值
next_actions = self.actor_target.model.predict_on_batch(next_state)
next_Q_targets = self.critic_target.model.predict_on_batch(
[next_state, next_actions])
Q_targets = rewards + self.GAMMA * next_Q_targets * (1 - dones)
self.critic_local.model.train_on_batch(x=[states, actions],
y=Q_targets)
#训练actor 模型
action_gradients = np.reshape(
self.critic_local.get_action_gradients([states, actions, 0]),
(-1, self.action_size))
self.actor_local.train_fn([states, action_gradients, 1])
self.update(self.critic_local.model, self.critic_target.model)
self.update(self.actor_local.model, self.actor_target.model)
def update(self, local_model, target_model):
local_weights = np.array(local_model.get_weights())
target_weights = np.array(target_model.get_weights())
new_weights = self.TAU * local_weights + (1 -
self.TAU) * target_weights
target_model.set_weights(new_weights)
def __init__(self,
state_size,
action_size,
num_agents,
device,
seed=23520,
GRADIENT_CLIP=1,
ACTIVATION=F.relu,
BOOTSTRAP_SIZE=5,
GAMMA=0.99,
TAU=1e-3,
LR_CRITIC=5e-4,
LR_ACTOR=5e-4,
UPDATE_EVERY=1,
TRANSFER_EVERY=2,
UPDATE_LOOP=10,
ADD_NOISE_EVERY=5,
WEIGHT_DECAY=0,
MEMORY_SIZE=5e4,
BATCH_SIZE=64):
"""Initialize an Agent object.
Params
======
state_size : dimension of each state
action_size : dimension of each action
num_agents : number of running agents
device: cpu or cuda:0 if available
-----These are hyperparameters----
BOOTSTRAP_SIZE : How far ahead to bootstrap
GAMMA : Discount factor
TAU : Parameter for performing soft updates of target parameters
LR_CRITIC, LR_ACTOR : Learning rate of the networks
UPDATE_EVERY : How often to update the networks
TRANSFER_EVERY : How often to transfer the weights from local to target
UPDATE_LOOP : Number of iterations for network update
ADD_NOISE_EVERY : How often to add noise to favor exploration
WEIGHT_DECAY : L2 weight decay for critic optimizer
GRADIENT_CLIP : Limit of gradient to be clipped, to avoid exploding gradient issue
"""
# Actor networks
self.actor_local = Actor(state_size, action_size).to(device)
self.actor_target = Actor(state_size, action_size).to(device)
self.actor_optim = optim.Adam(self.actor_local.parameters(),
lr=LR_ACTOR,
weight_decay=WEIGHT_DECAY)
hard_update(self.actor_local, self.actor_target)
#critic networks
self.critic_local = Critic(state_size * 2, action_size).to(device)
self.critic_target = Critic(state_size * 2, action_size).to(device)
self.critic_optim = optim.Adam(self.critic_local.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
hard_update(self.critic_local, self.critic_target)
self.device = device
self.num_agents = num_agents
# Noise : using simple noise instead of OUNoise
self.noise = [
SimpleNoise(action_size, scale=1) for i in range(num_agents)
]
# Replay memory
self.memory = ReplayBuffer(action_size, device, int(MEMORY_SIZE),
BATCH_SIZE, seed)
# Initialize time steps (for updating every UPDATE_EVERY steps)
self.u_step = 0
self.n_step = 0
#keeping hyperparameters within the instance
self.BOOTSTRAP_SIZE = BOOTSTRAP_SIZE
self.GAMMA = GAMMA
self.TAU = TAU
self.LR_CRITIC = LR_CRITIC
self.LR_ACTOR = LR_ACTOR
self.UPDATE_EVERY = UPDATE_EVERY
self.TRANSFER_EVERY = TRANSFER_EVERY
self.UPDATE_LOOP = UPDATE_LOOP
self.ADD_NOISE_EVERY = ADD_NOISE_EVERY
self.GRADIENT_CLIP = GRADIENT_CLIP
# initialize these variables to store the information of the n-previous timestep that are necessary to apply the bootstrap_size
self.rewards = deque(maxlen=BOOTSTRAP_SIZE)
self.states = deque(maxlen=BOOTSTRAP_SIZE)
self.actions = deque(maxlen=BOOTSTRAP_SIZE)
self.gammas = np.array([[GAMMA**i for j in range(num_agents)]
for i in range(BOOTSTRAP_SIZE)])
self.loss_function = torch.nn.SmoothL1Loss()
class MADDPG():
"""
Class definition of MADDPG agent. Interacts with and learns from the environment
Comprises of a pair of Actor-Critic network ad implements centralized training and decentralized exeution (learn function)
"""
def __init__(self,
state_size,
action_size,
num_agents,
device,
seed=23520,
GRADIENT_CLIP=1,
ACTIVATION=F.relu,
BOOTSTRAP_SIZE=5,
GAMMA=0.99,
TAU=1e-3,
LR_CRITIC=5e-4,
LR_ACTOR=5e-4,
UPDATE_EVERY=1,
TRANSFER_EVERY=2,
UPDATE_LOOP=10,
ADD_NOISE_EVERY=5,
WEIGHT_DECAY=0,
MEMORY_SIZE=5e4,
BATCH_SIZE=64):
"""Initialize an Agent object.
Params
======
state_size : dimension of each state
action_size : dimension of each action
num_agents : number of running agents
device: cpu or cuda:0 if available
-----These are hyperparameters----
BOOTSTRAP_SIZE : How far ahead to bootstrap
GAMMA : Discount factor
TAU : Parameter for performing soft updates of target parameters
LR_CRITIC, LR_ACTOR : Learning rate of the networks
UPDATE_EVERY : How often to update the networks
TRANSFER_EVERY : How often to transfer the weights from local to target
UPDATE_LOOP : Number of iterations for network update
ADD_NOISE_EVERY : How often to add noise to favor exploration
WEIGHT_DECAY : L2 weight decay for critic optimizer
GRADIENT_CLIP : Limit of gradient to be clipped, to avoid exploding gradient issue
"""
# Actor networks
self.actor_local = Actor(state_size, action_size).to(device)
self.actor_target = Actor(state_size, action_size).to(device)
self.actor_optim = optim.Adam(self.actor_local.parameters(),
lr=LR_ACTOR,
weight_decay=WEIGHT_DECAY)
hard_update(self.actor_local, self.actor_target)
#critic networks
self.critic_local = Critic(state_size * 2, action_size).to(device)
self.critic_target = Critic(state_size * 2, action_size).to(device)
self.critic_optim = optim.Adam(self.critic_local.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY)
hard_update(self.critic_local, self.critic_target)
self.device = device
self.num_agents = num_agents
# Noise : using simple noise instead of OUNoise
self.noise = [
SimpleNoise(action_size, scale=1) for i in range(num_agents)
]
# Replay memory
self.memory = ReplayBuffer(action_size, device, int(MEMORY_SIZE),
BATCH_SIZE, seed)
# Initialize time steps (for updating every UPDATE_EVERY steps)
self.u_step = 0
self.n_step = 0
#keeping hyperparameters within the instance
self.BOOTSTRAP_SIZE = BOOTSTRAP_SIZE
self.GAMMA = GAMMA
self.TAU = TAU
self.LR_CRITIC = LR_CRITIC
self.LR_ACTOR = LR_ACTOR
self.UPDATE_EVERY = UPDATE_EVERY
self.TRANSFER_EVERY = TRANSFER_EVERY
self.UPDATE_LOOP = UPDATE_LOOP
self.ADD_NOISE_EVERY = ADD_NOISE_EVERY
self.GRADIENT_CLIP = GRADIENT_CLIP
# initialize these variables to store the information of the n-previous timestep that are necessary to apply the bootstrap_size
self.rewards = deque(maxlen=BOOTSTRAP_SIZE)
self.states = deque(maxlen=BOOTSTRAP_SIZE)
self.actions = deque(maxlen=BOOTSTRAP_SIZE)
self.gammas = np.array([[GAMMA**i for j in range(num_agents)]
for i in range(BOOTSTRAP_SIZE)])
self.loss_function = torch.nn.SmoothL1Loss()
def reset(self):
if self.noise:
for n in self.noise:
n.reset()
def set_noise(self, noise):
self.noise = noise
def save(self, filename):
torch.save(self.actor_local.state_dict(),
"{}_actor_local.pth".format(filename))
torch.save(self.actor_target.state_dict(),
"{}_actor_target.pth".format(filename))
torch.save(self.critic_local.state_dict(),
"{}_critic_local.pth".format(filename))
torch.save(self.critic_target.state_dict(),
"{}_critic_target.pth".format(filename))
def load(self, path):
self.actor_local.load_state_dict(torch.load(path + "_actor_local.pth"))
self.actor_target.load_state_dict(
torch.load(path + "_actor_target.pth"))
self.critic_local.load_state_dict(
torch.load(path + "_critic_local.pth"))
self.critic_target.load_state_dict(
torch.load(path + "_critic_target.pth"))
def act(self, states, noise=0.0):
"""
Returns actions of each actor for given states.
Params
======
state : current states
add_noise: Introduce some noise in agent's action or not. During training, this is necessary to promote the exploration but should not be used during validation
"""
actions = None
self.n_step = (self.n_step + 1) % self.ADD_NOISE_EVERY
with torch.no_grad():
self.actor_local.eval()
states = torch.from_numpy(states).float().unsqueeze(0).to(
self.device)
actions = self.actor_local(states).squeeze().cpu().data.numpy()
self.actor_local.train()
if self.n_step == 0:
for i in range(len(actions)):
actions[i] += noise * self.noise[i].sample()
return actions
def step(self, states, actions, rewards, next_states, dones):
"""
Take a step for the current episode
1. Save the experience
2. Bootstrap the rewards
3. If update conditions are statisfied, perform learning on required number of loops
"""
# Save experience in replay memory
self.rewards.append(rewards)
self.states.append(states)
self.actions.append(actions)
if len(self.rewards) == self.BOOTSTRAP_SIZE:
# get the bootstrapped sum of rewards per agents
reward = np.sum(self.rewards * self.gammas, axis=-2)
self.memory.add(self.states[0], self.actions[0], reward,
next_states, dones)
if np.any(dones):
self.rewards.clear()
self.states.clear()
self.actions.clear()
# Learn every UPDATE_EVERY timesteps
self.u_step = (self.u_step + 1) % self.UPDATE_EVERY
t_step = 0
if len(self.memory) > self.memory.batch_size and self.u_step == 0:
for _ in range(self.UPDATE_LOOP):
self.learn()
# transfer the weights as specified
t_step = (t_step + 1) % self.TRANSFER_EVERY
if t_step == 0:
soft_update(self.actor_local, self.actor_target, self.TAU)
soft_update(self.critic_local, self.critic_target,
self.TAU)
def transform_states(self, states):
"""
Transforms states to full states so that both agents can see each others state via the critic network
"""
batch_size = states.shape[0]
state_size = states.shape[-1]
num_agents = states.shape[-2]
transformed_states = torch.zeros(
(batch_size, num_agents, state_size * num_agents)).to(self.device)
for i in range(num_agents):
start = 0
for j in range(num_agents):
transformed_states[:, i, start:start + state_size] += states[:,
j]
start += state_size
return transformed_states
def learn(self):
"""
Update the network parameters using the experiences. The algorithm is described in detail in readme
Params
======
experiences : List of (s, a, r, s', done) tuples
"""
# sample the memory to disrupt the internal correlation
states, actions, rewards, next_states, dones = self.memory.sample()
full_states = self.transform_states(states)
# The critic should estimate the value of the states to be equal to rewards plus
# the estimation of the next_states value according to the critic_target and actor_target
with torch.no_grad():
self.actor_target.eval()
self.critic_target.eval()
# obtain next actions as given by the target network and get transformed states for the critic
next_actions = self.actor_target(next_states)
next_full_states = self.transform_states(next_states)
# calculate Q value using transformed next states and next actions, basically predict what the next value is from target's perspective
q_next = self.critic_target(next_full_states,
next_actions).squeeze(-1)
# calculate the target's value
targeted_value = rewards + (
self.GAMMA**self.BOOTSTRAP_SIZE) * q_next * (1 - dones)
current_value = self.critic_local(full_states, actions).squeeze(-1)
loss = self.loss_function(current_value, targeted_value)
# During the optimization, the critic tells how much the value is off from the action value and adjusts the network towards. Basically, the critic takes the actions predicted by the actor, and tells how good or bad they are by calculating its Q-value
# calculate the loss of the critic network and backpropagate
self.critic_optim.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.critic_local.parameters(),
self.GRADIENT_CLIP)
self.critic_optim.step()
# optimize the actor by having the critic evaluating the value of the actor's decision
self.actor_optim.zero_grad()
actions_pred = self.actor_local(states)
mean = self.critic_local(full_states, actions_pred).mean()
(-mean).backward()
self.actor_optim.step()