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 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(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)