class DDPG():
"""Reinforcement Learning agent that learns using DDPG."""
def __init__(self,
task,
explore_mu=0,
explore_theta=0.15,
explore_sigma=0.2,
gamma=0.99,
tau=0.01):
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 = explore_mu # 0
self.exploration_theta = explore_theta # 0.15
self.exploration_sigma = explore_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 = gamma # 0.99 # discount factor
self.tau = gamma # 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_Agent:
""" RL agent using actor-critic method with DDPG concept (Deep Deterministic Policy Gradient) """
def __init__(self, task, alpha=0.0001, beta=0.001, gamma=0.9, tau=0.125 , mu=0.2,
theta=0.15, sigma=0.2, max_size=100000):
""" Initialize DDPG agent.
Params
======
- task: the task that shall be learned (for the quadcopter project: take-off or landing)
implemented is the take-off task which uses only the PhysicsSim pose information:
from the quadcopter position delivered by such physical basic class only the target height
(means the z value) is relevant; if its value is reached the episode termination condition
is valid
- alpha: actor learning rate
- beta: critic learning rate
- gamma: discount value, for termination function: final state with gamma(s)=0
- tau: used for smooth shifting from the local prediction models to the target models (soft update of targets)
- mu, theta, sigma: for creation of Ornstein-Uhlenbeck noise instance, part of exploration policy creation
- max_size: memory buffer size
"""
#
# set hyperparameters
#
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.gamma = gamma # algorithm discount value
self.tau = tau # algorithm soft update of targets
self.actor_learning_rate = alpha
self.critic_learning_rate = beta
#
# networks creation
#
# networks of actor and critic components
self.network_name = "DQNetwork"
self.target_name = "TargetNetwork"
self.actor = Model(self.task, "actor", self.network_name).get_model()
self.actor_target = Model(self.task, "actor", self.target_name).get_model()
self.critic = Model(self.task, "critic", self.network_name).get_model()
self.critic_target = Model(self.task, "critic", self.target_name).get_model()
#
# training resp. learning part
#
# set memory
self.buffer_size = max_size
self.batch_size = 64
self.memory = ReplayBuffer(self.buffer_size, self.batch_size)
# set noise
# Ornstein-Uhlenbeck noise process (as mentioned in the DDPG paper)
# for construction of exploration policy,
# by adding noise sampled from a noice process N to the actor policy
# in the paper supplementary as param values theta=0.15 and sigma=0.2 are used, shall be default values
self.exploration_mu = mu
self.exploration_theta = theta
self.exploration_sigma = sigma
self.noise = OUNoise(self.action_size, self.exploration_mu, self.exploration_theta, self.exploration_sigma)
def get_memory(self):
return self.memory
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):
""" Update the agent's knowledge, using the most recently sampled tuple.
Params
======
- state: the previous state of the environment
- action: the agent's previous choice of action
- reward: last reward received
- next_state: the current state of the environment
- done: whether the episode is complete (True or False)
"""
# 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.model.predict(state)[0]
return list(action + self.noise.get_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
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.model.train_on_batch(x=[states, actions], y=Q_targets)
# Train actor model (local)
action_gradients = np.reshape(self.critic.get_action_gradients([states, actions, 0]), (-1, self.action_size))
self.actor.train_func([states, action_gradients, 1]) # custom training function
# Soft-update target models
self.soft_update(self.critic.model, self.critic_target.model)
self.soft_update(self.actor.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 who learns using DDPG"""
def __init__(self, task):
"""Initialize models"""
self.env = task
self.state_size = task.state_size
self.action_size = task.action_size
self.action_high = task.action_high
self.action_low = task.action_low
# Initialize Actor (policy) models
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)
# Initialize Critic (value) models
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.actor_target.model.set_weights(
self.actor_local.model.get_weights())
self.critic_target.model.set_weights(
self.critic_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 buffer
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
def reset_episode(self, task):
"""Return state after reseting task"""
self.noise.reset()
state = task.reset()
self.last_state = state
return state
def step(self, action, reward, next_state, done):
# Add experience to memory
self.memory.add_experience(self.last_state, action, reward, next_state,
done)
# Learn is memory is larger than batch size
if len(self.memory) > self.batch_size:
experiences = self.memory.sample()
self.learn(experiences)
# Roll over state
self.last_state = next_state
def act(self, state):
"""Returns action using the policy network """
state = np.reshape(state, [-1, self.state_size])
action = self.actor_local.model.predict(state)[0]
return list(action + self.noise.sample())
def learn(self, experiences):
# Convert experience tuples to separate arrays for each element
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(next_states)
Q_targets_next = self.critic_target.model.predict(
[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])
# Soft-update target models
self.soft_update(self.actor_local.model, self.actor_target.model)
self.soft_update(self.critic_local.model, self.critic_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())
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 save_model(self, path):
self.actor_local.model.save_weights(path)
def load_model(self, path):
self.actor_local.model.load_weights(path)
def act_only(self, state):
state = np.reshape(state, [-1, self.state_size])
action = self.actor_local.model.predict(state)[0]
return list(action)