def run(self):
memory = ReplayMemory(self.memory_capacity,
transition_type=DQNTransition)
optimizer = tf.keras.optimizers.Adam(lr=self.learning_rate)
loss_fn = tf.keras.losses.Huber()
while self.done_queue.qsize() < self.n_envs:
while self.memory_queue.qsize() > 0:
sample = self.memory_queue.get()
memory.push(*sample)
if len(memory) < self.batch_size:
continue
transitions = memory.sample(self.batch_size)
batch = DQNTransition(*zip(*transitions))
state_batch = tf.convert_to_tensor(batch.state)
action_batch = tf.convert_to_tensor(batch.action)
reward_batch = tf.convert_to_tensor(batch.reward, dtype=tf.float32)
next_state_batch = tf.convert_to_tensor(batch.next_state)
done_batch = tf.convert_to_tensor(batch.done)
with tf.GradientTape() as tape:
state_action_values = tf.math.reduce_sum(
self.model.policy_network(state_batch) *
tf.one_hot(action_batch, self.model.n_actions),
axis=1)
next_state_values = tf.where(
done_batch, tf.zeros(self.batch_size),
tf.math.reduce_max(
self.model.target_network(next_state_batch), axis=1))
expected_state_action_values = reward_batch + \
(self.discount_factor * next_state_values)
loss_value = loss_fn(expected_state_action_values,
state_action_values)
variables = self.model.policy_network.trainable_variables
gradients = tape.gradient(loss_value, variables)
optimizer.apply_gradients(zip(gradients, variables))
self.model_update_queue.put(self.model)
def run(self):
episode = 0
for i in range(self.n_envs):
self.model_update_queue.put(self.model.policy_network.to_json())
memory = ReplayMemory(self.memory_capacity,
transition_type=DQNTransition)
optimizer = tf.keras.optimizers.Adam(lr=self.learning_rate)
loss_fn = tf.keras.losses.Huber()
while self.done_queue.qsize() < self.n_envs:
while self.sync_queue.qsize() < self.n_envs:
while self.memory_queue.qsize() > 0:
sample = self.memory_queue.get()
memory.push(*sample)
self.empty_queue(self.sync_queue)
transitions = memory.sample(self.batch_size)
batch = DQNTransition(*zip(*transitions))
state_batch = tf.convert_to_tensor(batch.state)
action_batch = tf.convert_to_tensor(batch.action)
reward_batch = tf.convert_to_tensor(batch.reward, dtype=tf.float32)
next_state_batch = tf.convert_to_tensor(batch.next_state)
done_batch = tf.convert_to_tensor(batch.done)
with tf.GradientTape() as tape:
state_action_values = tf.math.reduce_sum(
self.model.policy_network(state_batch) *
tf.one_hot(action_batch, self.model.n_actions),
axis=1)
next_state_values = tf.where(
done_batch, tf.zeros(self.batch_size),
tf.math.reduce_max(
self.model.target_network(next_state_batch), axis=1))
expected_state_action_values = reward_batch + \
(self.discount_factor * next_state_values)
loss_value = loss_fn(expected_state_action_values,
state_action_values)
variables = self.model.policy_network.trainable_variables
gradients = tape.gradient(loss_value, variables)
optimizer.apply_gradients(zip(gradients, variables))
self.model.update_target_network()
json_model = self.model.policy_network.to_json()
for n in range(self.n_envs):
self.model_update_queue.put(json_model)
for n in range(self.n_envs * 2):
self.sync_queue.put(1)
episode += 1
if episode == self.n_episodes:
break
while self.model_update_queue.qsize() > 0:
time.sleep(.1)
print("")
self.empty_queue(self.sync_queue)
self.empty_queue(self.model_update_queue)
time.sleep(2)
return self.model