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 = 0.99 # discount factor
self.tau = 0.01 # for soft update of target parameters
# self.tau = 0.001 # for soft update of target parameters
# self.tau = 0.1 # for soft update of target parameters
def reset_episode(self):
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)
# Roll over 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]
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."""
# 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 DDPG():
"""Reinforcement Learning agent that learns using DDPG."""
def __init__(self, task):
print('loaded DDPG ')
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.action_range = self.action_high - self.action_low
# 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 = 0.99 # discount factor
self.tau = 0.01 # for soft update of target parameters
# Score tracker and learning parameters
self.best_w = None
self.best_score = -np.inf
self.noise_scale = 0.1
self.w = np.random.normal(size=(self.state_size, self.action_size), # weights for simple linear policy: state_space x action_space
scale=(self.action_range / (2 * self.state_size))) # start producing actions in a decent range
# Episode variables
# self.reset_episode()
#load weight if existing
def reset_episode(self):
self.total_reward = 0.0
self.count = 0
self.noise_scale = 0.1
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)
self.total_reward += reward
self.count += 1
# 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
if done:
self.score_update()
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]
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."""
# 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)
def score_update(self):
# Learn by random policy search, using a reward-based score
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
self.best_w = self.w
self.noise_scale = max(0.5 * self.noise_scale, 0.01)
else:
self.w = self.best_w
self.noise_scale = min(2.0 * self.noise_scale, 3.2)
self.w = self.w + self.noise_scale * np.random.normal(size=self.w.shape) # equal noise in all directions
class DDPG():
"""Reinforcement Learning agent that learns using DDPG.
Params
======
task (Task): Instance of the Task class which reports the environment to this agent
log (Log): Reference to the log utility.
"""
def __init__(self, task, log):
self.task = task
#Add log utility
self.log = log
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 0.99
self.tau = 0.001 # for soft update of target parameters 0.01
# score
self.total_reward = 0.0
self.count = 0
self.score = 0
self.best_score = -np.inf
def reset_episode(self):
self.noise.reset()
state = self.task.reset()
self.last_state = state
self.total_reward = 0.0
self.count = 0
return state
def step(self, action, reward, next_state, done):
# Save experience / reward
self.memory.add(self.last_state, action, reward, next_state, done)
self.total_reward += reward
self.count += 1
self.log.write('DDPG.step len(self.memory)=' + str(len(self.memory)) +
' total_reward=' + str(total_reward) + ' count=' +
str(count))
# Learn, if enough samples are available in memory
if len(self.memory) > self.batch_size:
experiences = self.memory.sample()
self.log.write('DDPG.step len(self.memory) > self.batch_size' +
str(len(self.memory)) + '>' + str(self.batch_size))
print(experiences)
with open('experience.csv', 'w') as csvfile:
writer = csv.writer(csvfile)
writer.writerow(experiences)
self.learn(experiences)
#print(self.critic_local.model.get_weights()[0][0])
#print(self.critic_target.model.get_weights()[0][0])
#for lay in self.critic_local.model.layers:
# if lay.name == 'q_values':
# print(lay.name + ': ' + str(lay.get_weights()[0][5]))
# Roll over 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]
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."""
# Convert experience tuples to separate arrays for each element (states, actions, rewards, etc.)
self.log.write('DDPG.learn experiences=' + str(len(experiences)))
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)
self.score = self.total_reward / float(
self.count) if self.count else 0.0
self.best_score = max(self.best_score, self.score)
#if self.score > self.best_score:
# self.best_score = self.score
def soft_update(self, local_model, target_model):
"""Soft update model parameters. As opposed to a fixed Q-targets method."""
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)
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.action_range = self.action_high - self.action_low
# 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 = 0.4
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 = 0.995 # discount factor
self.tau = 0.9 # for soft update of target parameters
# Score tracker and learning parameters
self.best_w = None
self.best_score = -np.inf
self.noise_scale = 0.1
self.w = np.random.normal(
size=(
self.state_size, self.action_size
), # weights for simple linear policy: state_space x action_space
scale=(self.action_range / (2 * self.state_size)
)) # start producing actions in a decent range
# Episode variables
# self.reset_episode()
#load weight if existing
try:
self.actor_local.model.load_weights('actormodel.h5')
self.actor_target.model.load_weights('actormodel.h5')
self.critic_local.model.load_weights('criticmodel.h5')
self.critic_target.model.load_weights('criticmodel.h5')
print('Weight load successfully')
except:
print("Cannot find the weight")
class 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
#Policy Model & Value Model
self.actorLocal = Actor(self.state_size, self.action_size,
self.action_low, self.action_high)
self.criticLocal = Critic(self.state_size, self.action_size)
self.actorTarget = Actor(self.state_size, self.action_size,
self.action_low, self.action_high)
self.criticTarget = Critic(self.state_size, self.action_size)
#Initializing target model with local model params
self.criticTarget.model.set_weights(
self.criticLocal.model.get_weights())
self.actorTarget.model.set_weights(self.actorLocal.model.get_weights())
#Replay Buffer
self.buffer_size = 100000
self.batch_size = 64
self.memory = ReplayBuffer(self.buffer_size, self.batch_size)
self.noise = OUNoise(self.action_size, 0, 0.1, 0.25)
self.discountGamma = 0.9
def reset_episode(self):
self.noise.reset()
state = self.task.reset()
self.last_state = state
return state
def step(self, action, reward, next_state, done):
self.memory.add(self.last_state, action, reward, next_state, done)
if len(self.memory) > self.batch_size:
exp = self.memory.sample()
self.learn(exp)
self.last_state = next_state
def act(self, state):
state = np.reshape(state, [-1, self.state_size])
action = self.actorLocal.model.predict(state)[0]
return list(action + self.noise.sample())
def learn(self, exp):
"""
https://docs.scipy.org/doc/numpy/reference/generated/numpy.vstack.html
Vertical Stacking of arrays
This took a long time to get in place :). Thanks to some other references in github too for examples.
"""
state = np.vstack([ex.state for ex in exp if ex is not None])
action = np.array([ex.action for ex in exp
if ex is not None]).reshape(-1, self.action_size)
reward = np.array([ex.reward for ex in exp
if ex is not None]).reshape(-1, 1)
done = np.array([ex.done for ex in exp
if ex is not None]).astype(np.uint8).reshape(-1, 1)
next_states = np.vstack(
[ex.next_state for ex in exp if ex is not None])
actions_next = self.actorTarget.model.predict_on_batch(next_states)
QTargets_next = self.criticTarget.model.predict_on_batch(
[next_states, actions_next])
Q_targets = reward + self.discountGamma * QTargets_next * (1 - done)
self.criticLocal.model.train_on_batch(x=[state, action], y=Q_targets)
actionGradients = np.reshape(
self.criticLocal.get_action_gradients([state, action, 0]),
(-1, self.action_size))
self.actorLocal.train_fn([state, actionGradients, 1])
# Soft-update target models
self.criticTarget.model.set_weights(
0.01 * np.array(self.criticLocal.model.get_weights()) +
(1 - 0.01) * np.array(self.criticTarget.model.get_weights()))
self.actorTarget.model.set_weights(
0.01 * np.array(self.actorLocal.model.get_weights()) +
(1 - 0.01) * np.array(self.actorTarget.model.get_weights()))