class Agent():
def __init__(self,
alpha,
beta,
input_dims,
tau,
n_actions,
gamma=0.99,
max_size=50000,
fc1_dims=400,
fc2_dims=300,
batch_size=32):
self.gamma = gamma
self.tau = tau
self.batch_size = batch_size
self.alpha = alpha
self.beta = beta
self.memory = ReplayBuffer(max_size, input_dims, n_actions)
self.noise = OUActionNoise(mu=np.zeros(n_actions))
self.actor = ActorNetwork(alpha,
input_dims,
fc1_dims,
fc2_dims,
n_actions=n_actions,
name='actor')
self.critic = CriticNetwork(beta,
input_dims,
fc1_dims,
fc2_dims,
n_actions=n_actions,
name='critic')
self.target_actor = ActorNetwork(alpha,
input_dims,
fc1_dims,
fc2_dims,
n_actions=n_actions,
name='target_actor')
self.target_critic = CriticNetwork(beta,
input_dims,
fc1_dims,
fc2_dims,
n_actions=n_actions,
name='target_critic')
self.update_network_parameters(
tau=1) # for the first time target_actor and actor are same
def choose_action(self, observation):
self.actor.eval(
) # we are setting our actor network to eval mode because we have batch normalization layer
# and we dont want to calculate statistics for that layer at this step
state = T.tensor([observation], dtype=T.float).to(self.actor.device)
mu = self.actor.forward(state).to(self.actor.device)
mu_prime = mu + T.tensor(self.noise(), dtype=T.float).to(
self.actor.device)
self.actor.train()
return mu_prime.cpu().detach().numpy()[0]
def remember(self, state, action, reward, state_, done):
self.memory.store_transition(state, action, reward, state_, done)
def save_models(self):
self.actor.save_checkpoint()
self.target_actor.save_checkpoint()
self.critic.save_checkpoint()
self.target_critic.save_checkpoint()
def load_models(self):
self.actor.load_checkpoint()
self.target_actor.load_checkpoint()
self.critic.load_checkpoint()
self.target_critic.load_checkpoint()
def learn(self):
if self.memory.mem_cntr < self.batch_size:
return
states, actions, rewards, states_, done = self.memory.sample_buffer(
self.batch_size)
states = T.tensor(states, dtype=T.float).to(self.actor.device)
states_ = T.tensor(states_, dtype=T.float).to(self.actor.device)
actions = T.tensor(actions, dtype=T.float).to(self.actor.device)
rewards = T.tensor(rewards, dtype=T.float).to(self.actor.device)
done = T.tensor(done).to(self.actor.device)
target_actions = self.target_actor.forward(states_)
critic_value_ = self.target_critic.forward(states_, target_actions)
critic_value = self.critic.forward(states, actions)
critic_value_[done] = 0.0
critic_value_ = critic_value_.view(-1)
target = rewards + self.gamma * critic_value_
target = target.view(self.batch_size, 1)
self.critic.optimizer.zero_grad()
critic_loss = F.mse_loss(target, critic_value)
critic_loss.backward()
self.critic.optimizer.step()
self.actor.optimizer.zero_grad()
actor_loss = -self.critic.forward(states, self.actor.forward(states))
actor_loss = T.mean(actor_loss)
actor_loss.backward()
self.actor.optimizer.step()
self.update_network_parameters(
) # sending tau None so that local tau variable there takes the value of class tau variable
def update_network_parameters(self, tau=None):
if tau is None:
tau = self.tau
actor_params = self.actor.named_parameters()
critic_params = self.critic.named_parameters()
target_actor_params = self.target_actor.named_parameters()
target_critic_params = self.target_critic.named_parameters()
critic_state_dict = dict(critic_params)
actor_state_dict = dict(actor_params)
target_critic_state_dict = dict(target_critic_params)
target_actor_state_dict = dict(target_actor_params)
for name in critic_state_dict:
critic_state_dict[name] = tau * critic_state_dict[name].clone() + (
1 - tau) * target_critic_state_dict[name].clone()
for name in actor_state_dict:
actor_state_dict[name] = tau * actor_state_dict[name].clone() + (
1 - tau) * target_actor_state_dict[name].clone()
self.target_critic.load_state_dict(critic_state_dict)
self.target_actor.load_state_dict(actor_state_dict)
class DdpgAgent:
"""
A Deep Deterministic Policy Gradient Agent.
Interacts with and learns from the environment.
"""
def __init__(self, num_agents, state_size, action_size, random_seed):
"""
Initialize an Agent object.
Params
======
num_agents (int): number of agents observed at the same time. multiple agents are handled within the class.
state_size (int): dimension of each state
action_size (int): dimension of each action
random_seed (int): random seed
"""
if random_seed is not None:
random.seed(random_seed)
np.random.seed(random_seed)
self.t_step = 0 # A counter that increases each time the "step" function is executed
self.state_size = state_size
self.action_size = action_size
# Actor Network (w/ Target Network)
self.actor_local = ActorNetwork(state_size,
action_size,
USE_BATCH_NORM,
random_seed,
fc1_units=FC1_UNITS,
fc2_units=FC2_UNITS,
fc3_units=FC3_UNITS).to(device)
self.actor_target = ActorNetwork(state_size,
action_size,
USE_BATCH_NORM,
random_seed,
fc1_units=FC1_UNITS,
fc2_units=FC2_UNITS,
fc3_units=FC3_UNITS).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=LR_ACTOR,
weight_decay=WEIGHT_DECAY_ACTOR)
# self.actor_optimizer = optim.RMSprop(self.actor_local.parameters(), lr=LR_ACTOR,
# weight_decay=WEIGHT_DECAY_ACTOR) # Also solves it, but Adam quicker
# Critic Network (w/ Target Network)
self.critic_local = CriticNetwork(state_size,
action_size,
USE_BATCH_NORM,
random_seed,
fc1_units=FC1_UNITS,
fc2_units=FC2_UNITS,
fc3_units=FC3_UNITS).to(device)
self.critic_target = CriticNetwork(state_size,
action_size,
USE_BATCH_NORM,
random_seed,
fc1_units=FC1_UNITS,
fc2_units=FC2_UNITS,
fc3_units=FC3_UNITS).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=LR_CRITIC,
weight_decay=WEIGHT_DECAY_CRITIC)
# self.critic_optimizer = optim.RMSprop(self.critic_local.parameters(), lr=LR_CRITIC,
# weight_decay=WEIGHT_DECAY_CRITIC) # Also solves it, but Adam quicker
# Make sure target is initiated with the same weight as the local network
self.soft_update(self.actor_local, self.actor_target, 1)
self.soft_update(self.critic_local, self.critic_target, 1)
# Setting default modes for the networks
# Target networks do not need to train, so always eval()
# Local networks, in training mode, unless altered in code - eg when acting.
self.actor_local.train()
self.actor_target.eval()
self.critic_local.train()
self.critic_target.eval()
# Action Noise process (encouraging exploration during training)
# Could consider parameter noise in future as a potentially better alternative / addition
if ACTION_NOISE_METHOD == 'initial':
self.noise = InitialOrnsteinUhlenbeckActionNoise(
shape=(num_agents, action_size),
random_seed=random_seed,
x0=0,
mu=0,
theta=NOISE_THETA,
sigma=NOISE_SIGMA)
elif ACTION_NOISE_METHOD == 'adjusted':
self.noise = AdjustedOrnsteinUhlenbeckActionNoise(
shape=(num_agents, action_size),
random_seed=random_seed,
x0=0,
mu=0,
sigma=NOISE_SIGMA,
theta=NOISE_THETA,
dt=NOISE_DT,
sigma_delta=NOISE_SIGMA_DELTA,
)
else:
raise ValueError('Unknown action noise method: ' +
ACTION_NOISE_METHOD)
# Replay memory
self.memory = ReplayBuffer(
buffer_size=REPLAY_BUFFER_SIZE,
batch_size=BATCH_SIZE,
sampling_method=REPLAY_BUFFER_SAMPLING_METHOD,
random_seed=random_seed)
def step(self, states, actions, rewards, next_states, dones):
"""Save experience in replay memory, and use random sample from buffer to learn."""
self.t_step += 1
# Save experience / reward
self.memory.add(states, actions, rewards, next_states, dones)
# Learn, if enough samples are available in memory, every UPDATE_EVERY steps
if self.t_step % UPDATE_EVERY == 0:
if len(self.memory) > BATCH_SIZE:
experiences = self.memory.sample()
self.learn(experiences, GAMMA)
def act(self, states, add_action_noise=False):
"""Returns actions for given state as per current policy."""
states = torch.from_numpy(states).float().to(device)
self.actor_local.eval(
) # train state is set right before actual training
with torch.no_grad(
): # All calcs here with no_grad, but many examples didn't do this. Weirdly, this is slower..
return np.clip(
self.actor_local(states).cpu().data.numpy() +
(self.noise.sample() if add_action_noise else 0), -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): reward discount factor
"""
states, actions, rewards, next_states, dones = experiences
self.actor_local.train(
) # critic_local is always in train state, but actor_local goes into eval with acting
# 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()
if CLIP_GRADIENT_CRITIC:
torch.nn.utils.clip_grad_norm_(self.critic_local.parameters(), 1)
self.critic_optimizer.step()
# 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()
if CLIP_GRADIENT_ACTOR:
torch.nn.utils.clip_grad_norm_(self.actor_local.parameters(), 1)
self.actor_optimizer.step()
# Soft-Update of 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 target model parameters from local 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(object):
def __init__(self,
alpha,
beta,
input_dims,
action_bound,
tau,
env,
gamma=0.99,
n_actions=2,
max_size=1000000,
layer1_size=400,
layer2_size=300,
batch_size=64):
self.gamma = gamma
self.tau = tau
self.memory = ReplayBuffer(max_size, input_dims, n_actions)
self.batch_size = batch_size
self.action_bound = action_bound
self.actor = ActorNetwork(alpha,
input_dims,
layer1_size,
layer2_size,
n_actions=n_actions,
name='Actor')
self.critic = CriticNetwork(beta,
input_dims,
layer1_size,
layer2_size,
n_actions=n_actions,
name='Critic')
self.target_actor = ActorNetwork(alpha,
input_dims,
layer1_size,
layer2_size,
n_actions=n_actions,
name='TargetActor')
self.target_critic = CriticNetwork(beta,
input_dims,
layer1_size,
layer2_size,
n_actions=n_actions,
name='TargetCritic')
self.noise = OUActionNoise(mu=np.zeros(n_actions))
self.update_network_parameters(tau=1)
def choose_action(self, observation):
self.actor.eval()
observation = T.tensor(observation,
dtype=T.float).to(self.actor.device)
mu = self.actor.forward(observation).to(self.actor.device)
mu_prime = mu + T.tensor(self.noise(), dtype=T.float).to(
self.actor.device)
self.actor.train()
return (mu_prime * T.tensor(self.action_bound)).cpu().detach().numpy()
def remember(self, state, action, reward, new_state, done):
self.memory.store_transition(state, action, reward, new_state, done)
def learn(self):
if self.memory.mem_cntr < self.batch_size:
return
state, action, reward, new_state, done = \
self.memory.sample_buffer(self.batch_size)
reward = T.tensor(reward, dtype=T.float).to(self.critic.device)
done = T.tensor(done).to(self.critic.device)
new_state = T.tensor(new_state, dtype=T.float).to(self.critic.device)
action = T.tensor(action, dtype=T.float).to(self.critic.device)
state = T.tensor(state, dtype=T.float).to(self.critic.device)
self.target_actor.eval()
self.target_critic.eval()
self.critic.eval()
target_actions = self.target_actor.forward(new_state)
critic_value_ = self.target_critic.forward(new_state, target_actions)
critic_value = self.critic.forward(state, action)
target = []
for j in range(self.batch_size):
target.append(reward[j] + self.gamma * critic_value_[j] * done[j])
target = T.tensor(target).to(self.critic.device)
target = target.view(self.batch_size, 1)
self.critic.train()
self.critic.optimizer.zero_grad()
critic_loss = F.mse_loss(target, critic_value)
critic_loss.backward()
self.critic.optimizer.step()
self.critic.eval()
self.actor.optimizer.zero_grad()
mu = self.actor.forward(state)
self.actor.train()
actor_loss = -self.critic.forward(state, mu)
actor_loss = T.mean(actor_loss)
actor_loss.backward()
self.actor.optimizer.step()
self.update_network_parameters()
def update_network_parameters(self, tau=None):
if tau is None:
tau = self.tau
actor_params = self.actor.named_parameters()
critic_params = self.critic.named_parameters()
target_actor_params = self.target_actor.named_parameters()
target_critic_params = self.target_critic.named_parameters()
critic_state_dict = dict(critic_params)
actor_state_dict = dict(actor_params)
target_critic_dict = dict(target_critic_params)
target_actor_dict = dict(target_actor_params)
for name in critic_state_dict:
critic_state_dict[name] = tau*critic_state_dict[name].clone() + \
(1-tau)*target_critic_dict[name].clone()
self.target_critic.load_state_dict(critic_state_dict)
for name in actor_state_dict:
actor_state_dict[name] = tau*actor_state_dict[name].clone() + \
(1-tau)*target_actor_dict[name].clone()
self.target_actor.load_state_dict(actor_state_dict)
"""
#Verify that the copy assignment worked correctly
target_actor_params = self.target_actor.named_parameters()
target_critic_params = self.target_critic.named_parameters()
critic_state_dict = dict(target_critic_params)
actor_state_dict = dict(target_actor_params)
print('\nActor Networks', tau)
for name, param in self.actor.named_parameters():
print(name, T.equal(param, actor_state_dict[name]))
print('\nCritic Networks', tau)
for name, param in self.critic.named_parameters():
print(name, T.equal(param, critic_state_dict[name]))
input()
"""
def save_models(self):
self.actor.save_checkpoint()
self.target_actor.save_checkpoint()
self.critic.save_checkpoint()
self.target_critic.save_checkpoint()
def load_models(self):
self.actor.load_checkpoint()
self.target_actor.load_checkpoint()
self.critic.load_checkpoint()
self.target_critic.load_checkpoint()
def check_actor_params(self):
current_actor_params = self.actor.named_parameters()
current_actor_dict = dict(current_actor_params)
original_actor_dict = dict(self.original_actor.named_parameters())
original_critic_dict = dict(self.original_critic.named_parameters())
current_critic_params = self.critic.named_parameters()
current_critic_dict = dict(current_critic_params)
print('Checking Actor parameters')
for param in current_actor_dict:
print(
param,
T.equal(original_actor_dict[param], current_actor_dict[param]))
print('Checking critic parameters')
for param in current_critic_dict:
print(
param,
T.equal(original_critic_dict[param],
current_critic_dict[param]))
input()
class Agent:
""" This class represents the reinforcement learning agent """
def __init__(self,
state_size: int,
action_size: int,
gamma: float = 0.99,
lr_actor: float = 0.001,
lr_critic: float = 0.003,
weight_decay: float = 0.0001,
tau: float = 0.001,
buffer_size: int = 100000,
batch_size: int = 64):
"""
:param state_size: how many states does the agent get as input (input size of neural networks)
:param action_size: from how many actions can the agent choose
:param gamma: discount factor
:param lr_actor: learning rate of the actor network
:param lr_critic: learning rate of the critic network
:param weight_decay:
:param tau: soft update parameter
:param buffer_size: size of replay buffer
:param batch_size: size of learning batch (mini-batch)
"""
self.tau = tau
self.gamma = gamma
self.batch_size = batch_size
self.actor_local = ActorNetwork(state_size, action_size).to(device)
self.actor_target = ActorNetwork(state_size, action_size).to(device)
self.actor_optimizer = optim.Adam(self.actor_local.parameters(),
lr=lr_actor)
print(self.actor_local)
self.critic_local = CriticNetwork(state_size, action_size).to(device)
self.critic_target = CriticNetwork(state_size, action_size).to(device)
self.critic_optimizer = optim.Adam(self.critic_local.parameters(),
lr=lr_critic,
weight_decay=weight_decay)
print(self.critic_local)
self.hard_update(self.actor_local, self.actor_target)
self.hard_update(self.critic_local, self.critic_target)
self.memory = ReplayBuffer(action_size, buffer_size, batch_size)
# this would probably also work with Gaussian noise instead of Ornstein-Uhlenbeck process
self.noise = OUNoise(action_size)
def step(self, experience: tuple):
"""
:param experience: tuple consisting of (state, action, reward, next_state, done)
:return:
"""
self.memory.add(*experience)
if len(self.memory) > self.batch_size:
experiences = self.memory.sample()
self.learn(experiences)
def act(self, state, add_noise: bool = True):
""" Actor uses the policy to act given a state """
state = torch.from_numpy(state).float().to(device)
self.actor_local.eval()
with torch.no_grad():
action = self.actor_local.forward(state).cpu().data.numpy()
self.actor_local.train()
if add_noise:
action += self.noise.sample()
return np.clip(action, -1, 1)
def learn(self, experiences):
# Q_targets = r + γ * critic_target(next_state, actor_target(next_state))
# the actor_target returns the next action, this next action is then used (with the state) to estimate
# the Q-value with the critic_target network
states, actions, rewards, next_states, dones = experiences
# region Update Critic
actions_next = self.actor_target.forward(next_states)
q_expected = self.critic_local.forward(states, actions)
q_targets_next = self.critic_target.forward(next_states, actions_next)
q_targets = rewards + (self.gamma * q_targets_next * (1 - dones))
# minimize the loss
critic_loss = F.mse_loss(q_expected, q_targets)
self.critic_optimizer.zero_grad()
critic_loss.backward()
torch.nn.utils.clip_grad_norm_(self.critic_local.parameters(), 1)
self.critic_optimizer.step()
# endregion Update Critic
# region Update actor
# Compute actor loss
actions_predictions = self.actor_local.forward(states)
actor_loss = -self.critic_local.forward(states,
actions_predictions).mean()
# Minimize actor loss
self.actor_optimizer.zero_grad()
actor_loss.backward()
self.actor_optimizer.step()
# endregion Update actor
# region update target network
self.soft_update(self.critic_local, self.critic_target)
self.soft_update(self.actor_local, self.actor_target)
# endregion update target network
def soft_update(self, local_model, target_model):
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)
def hard_update(self, local_model, target_model):
"""Copy the weights and biases from the local to the target network"""
for target_param, param in zip(target_model.parameters(),
local_model.parameters()):
target_param.data.copy_(param.data)
def reset(self):
self.noise.reset()