act_input = tf.keras.Input((num_actions, ), name='actions_input')
if MODEL == 'DRL':
# Create an environment
ENV = Environment(input_train, y_train)
# Actor-Critic
# Train
critic_model = Network(num_actions, embeddings_matrix, maxlen=maxlen)
_ = critic_model(inputs=[inputs, act_input])
critic_model.compile(loss='mse', optimizer='adam')
# Decision
actor_q_model = tf.keras.Model(
inputs=critic_model.input,
outputs=critic_model.get_layer('q_outputs').output)
if MODE == 'train':
def train(samples):
if len(samples) < BATCH_SIZE:
return
samples = np.array(samples)
states, actions, old_q, rewards, next_states = zip(*samples)
states, actions, old_q, rewards = np.array(states), np.array(actions).reshape(-1, 1),\
np.array(old_q).reshape(-1, 1), np.array(rewards).reshape(-1, 1)
actions_one_hot = tf.keras.utils.to_categorical(
actions, num_actions)
q_estimate = (1 - ALPHA) * old_q + ALPHA * rewards.reshape(
