def __init__(self, actor_model, tgt_actor_model, critic_model, tgt_critic_model, action_limits,
actor_lr=1e-4, critic_lr=1e-3, critic_decay=1e-2, process=None, rb_size=1e6,
minibatch_size=64, tau=1e-3, gamma=0.99, warmup_episodes=None, logging=True):
super(DDPGAgent, self).__init__(warmup_episodes, logging)
self.actor = Actor(actor_model, critic_model, lr=actor_lr)
self.tgt_actor = Actor(tgt_actor_model, tgt_critic_model, lr=actor_lr)
self.tgt_actor.set_weights(self.actor.get_weights())
self.critic = Critic(critic_model, lr=critic_lr, decay=critic_decay)
self.tgt_critic = Critic(tgt_critic_model, lr=critic_lr, decay=critic_decay)
self.tgt_critic.set_weights(self.critic.get_weights())
self.action_limits = action_limits
self.process = process
self.buffer = ReplayBuffer(rb_size)
self.minibatch_size = minibatch_size
self.tau = tau
self.gamma = gamma
self.state_space = K.int_shape(critic_model.inputs[0])[1]
self.action_space = K.int_shape(critic_model.inputs[1])[1]
if process is None:
self.process = OrnsteinUhlenbeck(x0=np.zeros(self.action_space), theta=0.15, mu=0,
sigma=0.2)
else:
self.process = process
def __init__(self,
task,
exploration_mu=0,
exploration_theta=0.15,
exploration_sigma=0.2,
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 = exploration_mu
self.exploration_theta = exploration_theta
self.exploration_sigma = exploration_sigma
self.noise = OUNoise(self.action_size, self.exploration_mu,
self.exploration_theta, self.exploration_sigma)
# Replay memory
self.buffer_size = 100000
self.batch_size = 128 # 64
self.memory = ReplayBuffer(self.buffer_size, self.batch_size)
# Algorithm parameters
self.gamma = 0.85 # 0.99 # discount factor
self.tau = tau # for soft update of target parameters
##
self.total_reward = 0
self.best_score = -np.inf
self.score = 0
self.count = 0
def __init__(self, env, hparams):
n_action = len(env.action_space.high)
self.actor_main = Actor(n_action, hparams)
self.actor_target = Actor(n_action, hparams)
self.critic_main = Critic(hparams)
self.critic_target = Critic(hparams)
self.batch_size = 64
self.n_actions = len(env.action_space.high)
self.a_opt = tf.keras.optimizers.Adam(hparams['lr'])
# self.actor_target = tf.keras.optimizers.Adam(.001)
self.c_opt = tf.keras.optimizers.Adam(hparams['lr'])
# self.critic_target = tf.keras.optimizers.Adam(.002)
self.memory = RBuffer(1_00_000, env.observation_space.shape, len(env.action_space.high))
self.trainstep = 0
self.replace = 5
self.gamma = 0.99
self.min_action = env.action_space.low[0]
self.max_action = env.action_space.high[0]
class DDPGAgent(Agent):
def __init__(self, actor_model, tgt_actor_model, critic_model, tgt_critic_model, action_limits,
actor_lr=1e-4, critic_lr=1e-3, critic_decay=1e-2, process=None, rb_size=1e6,
minibatch_size=64, tau=1e-3, gamma=0.99, warmup_episodes=None, logging=True):
super(DDPGAgent, self).__init__(warmup_episodes, logging)
self.actor = Actor(actor_model, critic_model, lr=actor_lr)
self.tgt_actor = Actor(tgt_actor_model, tgt_critic_model, lr=actor_lr)
self.tgt_actor.set_weights(self.actor.get_weights())
self.critic = Critic(critic_model, lr=critic_lr, decay=critic_decay)
self.tgt_critic = Critic(tgt_critic_model, lr=critic_lr, decay=critic_decay)
self.tgt_critic.set_weights(self.critic.get_weights())
self.action_limits = action_limits
self.process = process
self.buffer = ReplayBuffer(rb_size)
self.minibatch_size = minibatch_size
self.tau = tau
self.gamma = gamma
self.state_space = K.int_shape(critic_model.inputs[0])[1]
self.action_space = K.int_shape(critic_model.inputs[1])[1]
if process is None:
self.process = OrnsteinUhlenbeck(x0=np.zeros(self.action_space), theta=0.15, mu=0,
sigma=0.2)
else:
self.process = process
def sense(self, s, a, r, s_new):
# print(self.state_space)
s = np.reshape(s, [-1, self.state_space])
s_new = np.reshape(s_new, [-1, self.state_space])
self.buffer.add((s, a, r, s_new))
def act(self, s):
s = np.reshape(s, [-1, self.state_space])
a = self.actor(s) #self.tgt_actor(s) # why acting with the target_ctor instead of the actor?
# Cache.
self.last_state = np.copy(s)
self.last_action = np.copy(a)
if self.learning_phase:
a = self.process(a)
#a += self.process() # for OU process
a = np.clip(a, self.action_limits[0], self.action_limits[1])
self.last_action_noisy = np.copy(a)
return a[0]
def train_step(self):
minibatch = self.buffer.sample(self.minibatch_size)
states = np.zeros([len(minibatch), self.state_space])
states_new = np.zeros([len(minibatch), self.state_space])
actions = np.zeros([len(minibatch), self.action_space])
r = np.zeros([len(minibatch), 1])
for i in range(len(minibatch)):
states[i], actions[i], r[i], states_new[i] = minibatch[i]
critic_out = self.critic(states_new, self.actor(states_new))
tgt_critic_out = self.tgt_critic(states_new, self.tgt_actor(states_new))
if self.logging:
log = [('s', self.last_state),
('a', self.last_action),
('a_noisy', self.last_action_noisy),
('q', self.critic(self.last_state, self.last_action)),
('q_tgt', self.tgt_critic(self.last_state, self.last_action)),
(('mse', np.mean(np.square(critic_out - tgt_critic_out))))]
self.add_log(log)
ys = r + self.gamma * tgt_critic_out
loss = self.critic.step(states, actions, ys) # update critic by minimizing the loss
self.actor.step(states) # update actor using the sampled policy gradient
# Soft weight update. (update the target networks)
critic_weights = self.critic.get_weights()
tgt_critic_weights = self.tgt_critic.get_weights()
actor_weights = self.actor.get_weights()
tgt_actor_weights = self.tgt_actor.get_weights()
for i in range(len(critic_weights)):
tgt_critic_weights[i] = (1 - self.tau) * tgt_critic_weights[i] + \
self.tau * critic_weights[i]
self.tgt_critic.set_weights(tgt_critic_weights)
for i in range(len(actor_weights)):
tgt_actor_weights[i] = (1 - self.tau) * tgt_actor_weights[i] + \
self.tau * actor_weights[i]
self.tgt_actor.set_weights(tgt_actor_weights)
return loss
def new_episode(self):
self.process.clear()
if self.logging:
self.logs.append({})
if len(self.logs) == 1:
self.logs[-1]['episode'] = 1 # Initial episode.
else:
self.logs[-1]['episode'] = self.logs[-2]['episode'] + 1
def save_weights(self, actor_suffix, critic_suffix):
self.actor.save_model_weights(actor_suffix)
self.critic.save_model_weights(critic_suffix)
class Agent():
"""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
self.best_score = -np.inf
self.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)
# 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
# Save experience / reward
self.total_reward += reward
self.count += 1
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)
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
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, hparams):
n_action = len(env.action_space.high)
self.actor_main = Actor(n_action, hparams)
self.actor_target = Actor(n_action, hparams)
self.critic_main = Critic(hparams)
self.critic_target = Critic(hparams)
self.batch_size = 64
self.n_actions = len(env.action_space.high)
self.a_opt = tf.keras.optimizers.Adam(hparams['lr'])
# self.actor_target = tf.keras.optimizers.Adam(.001)
self.c_opt = tf.keras.optimizers.Adam(hparams['lr'])
# self.critic_target = tf.keras.optimizers.Adam(.002)
self.memory = RBuffer(1_00_000, env.observation_space.shape, len(env.action_space.high))
self.trainstep = 0
self.replace = 5
self.gamma = 0.99
self.min_action = env.action_space.low[0]
self.max_action = env.action_space.high[0]
def act(self, state, evaluate=False):
state = tf.convert_to_tensor([state], dtype=tf.float32)
actions = self.actor_main(state)
if not evaluate:
actions += tf.random.normal(shape=[self.n_actions], mean=0.0, stddev=0.1)
actions = self.max_action * (tf.clip_by_value(actions, self.min_action, self.max_action))
# print(actions)
return actions[0]
def savexp(self, state, next_state, action, done, reward):
self.memory.storexp(state, next_state, action, done, reward)
def update_target(self):
self.actor_target.set_weights(self.actor_main.get_weights())
self.critic_target.set_weights(self.critic_main.get_weights())
def train(self):
if self.memory.cnt < self.batch_size:
return
states, next_states, rewards, actions, dones = self.memory.sample(self.batch_size)
states = tf.convert_to_tensor(states, dtype=tf.float32)
next_states = tf.convert_to_tensor(next_states, dtype=tf.float32)
rewards = tf.convert_to_tensor(rewards, dtype=tf.float32)
actions = tf.convert_to_tensor(actions, dtype=tf.float32)
# dones = tf.convert_to_tensor(dones, dtype= tf.bool)
with tf.GradientTape() as tape1, tf.GradientTape() as tape2:
target_actions = self.actor_target(next_states)
target_next_state_values = tf.squeeze(self.critic_target(next_states, target_actions), 1)
critic_value = tf.squeeze(self.critic_main(states, actions), 1)
target_values = rewards + self.gamma * target_next_state_values * dones
critic_loss = tf.keras.losses.MSE(target_values, critic_value)
new_policy_actions = self.actor_main(states)
actor_loss = -self.critic_main(states, new_policy_actions)
actor_loss = tf.math.reduce_mean(actor_loss)
grads1 = tape1.gradient(actor_loss, self.actor_main.trainable_variables)
grads2 = tape2.gradient(critic_loss, self.critic_main.trainable_variables)
self.a_opt.apply_gradients(zip(grads1, self.actor_main.trainable_variables))
self.c_opt.apply_gradients(zip(grads2, self.critic_main.trainable_variables))
if self.trainstep % self.replace == 0:
self.update_target()
self.trainstep += 1