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
# Score tracker and learning parameters
#self.count = 0
#self.score = 0
#self.total_reward = 0
#self.best_w = None
#self.best_score = -np.inf
#self.noise_scale = 0.1
def reset_episode(self):
#self.count = 0
#self.score = 0
#self.total_reward = 0
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.count += 1
#self.total_reward += reward
# 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]
self.noise.sample()
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 Agent:
def __init__(self, env, sess):
# Environment
self.n_state = env.observation_space.shape[0]
self.n_action = env.action_space.shape[0]
# Neural Networks
self.sess = sess
self.actor = Actor(self.sess, self.n_state, self.n_action)
self.critic = Critic(self.sess, self.n_state, self.n_action)
# Replay Buffer
self.replay_buffer = ReplayBuffer(BUFFER_SIZE)
# Ornstein-Uhlenbeck Noise
self.exploration_noise = OUNoise(self.n_action)
def noise_action(self, state):
'''Get action with noise'''
return self.action(state) + self.exploration_noise.noise()
def action(self, state):
'''Get action from online actor'''
return self.actor.action(state)
def train(self):
'''Train Networks'''
# Draw sample from Replay Buffer
minibatch = self.replay_buffer.get_batch(BATCH_SIZE)
state_batch = np.asarray([d[0] for d in minibatch])
action_batch = np.asarray([d[1] for d in minibatch])
reward_batch = np.asarray([d[2] for d in minibatch])
next_state_batch = np.asarray([d[3] for d in minibatch])
done_batch = np.asarray([d[4] for d in minibatch])
# Train Critic
next_action_batch = self.actor.target_actions(next_state_batch)
target_q_value_batch = self.critic.target_q(next_state_batch,
next_action_batch)
# q = r if done else r+gamma*target_q
q_batch = reward_batch.reshape(
(BATCH_SIZE, 1)) + (1. - done_batch.reshape(
BATCH_SIZE, 1).astype(float)) * GAMMA * target_q_value_batch
self.critic.train(q_batch, state_batch, action_batch)
# Train Actor
action_batch_grads = self.actor.actions(state_batch)
q_grads_batch = self.critic.gradients(state_batch, action_batch_grads)
self.actor.train(q_grads_batch, state_batch)
# Slowly update Target Networks
self.actor.update_target()
self.critic.update_target()
def perceive(self, state, action, reward, next_state, done):
'''Add transition to replay buffer and train if there are sufficient amount of transitions'''
# Add samples
self.replay_buffer.add(state, action, reward, next_state, done)
# Train if there are sufficient number of samples
if self.replay_buffer.count() > REPLAY_START:
self.train()
# Reset the noise for next episode
if done:
self.exploration_noise.reset()