我正在尝试在tensorflow和keras中实现深度确定性梯度方法，但是，我似乎被卡住了。似乎没有学习发生，模型采取的行动似乎根本没有改变，并且应用于演员网络的梯度也非常小（在e ^ -5的数量级上）。我已经使用了另一个实现作为参考，并且这个实现与完全相同的超参数和网络体系结构运行良好（除了它使用tflearn实现并包括批量规范化层），让我相信我的代码中存在某个错误。也许有人可以发现它。 谢谢你的时间！
编辑：我认为表现糟糕的原因是评论家网络在行动方面的渐变正在消失。但是，我无法找出原因。也许我使用连接层是错误的？
class AIInterface(object):
def __init__(self, sim):
self.sim = sim
self.pedal_pos = 0
self.steering_pos = 0
self.sess = tf.Session()
self.learning_rate = 10e-4
self.BATCH_SIZE = 64
self.epsilon = .75 #amount of random exploration
self.epsilon_decay = .997
self.gamma = .99 #reward discount factor
self.tau = .00125 #target update factor
self.rewards = deque(maxlen=100000)
self.memory = deque(maxlen=100000)
# Actor stuff
self.actor_model, self.actor_var_in = self.initialize_actor()
self.target_actor, _ = self.initialize_actor()
self.actor_critic_grad = tf.placeholder(tf.float32, [None, 1])
self.actor_model_weights = self.actor_model.trainable_weights
with tf.name_scope("actor_gradients"):
self.actor_grads = tf.gradients(self.actor_model.output, self.actor_model_weights, -self.actor_critic_grad)
self.normalized_actor_grads = list(map(lambda x: tf.div(x, self.BATCH_SIZE), self.actor_grads))
grads = zip(self.normalized_actor_grads, self.actor_model_weights)
self.optimize = tf.train.AdamOptimizer(self.learning_rate).apply_gradients(grads)
# Critic stuff
self.critic_model, self.critic_var_in, self.critic_action_in = self.initialize_critic()
self.target_critic, _, _ = self.initialize_critic()
with tf.name_scope("CriticGrads"):
self.critic_grads = tf.gradients(self.critic_model.output, self.critic_action_in)
self.sess.run(tf.global_variables_initializer())
self.target_actor.set_weights(self.actor_model.get_weights())
self.target_critic.set_weights(self.critic_model.get_weights())
self.global_step = 0
def initialize_actor(self):
state_variable_input = Input(shape=(3, ))
init = TruncatedNormal(mean=0.0, stddev=0.02)
dense = Dense(128, activation="relu", kernel_initializer=init)(state_variable_input)
dense2 = Dense(128, activation="relu", kernel_initializer=init)(dense)
output = Dense(1, activation="tanh", kernel_initializer=RandomUniform(-3e-3, 3e-3))(dense2)
model = Model(inputs=state_variable_input,
outputs=output)
model.compile(optimizer="adam", loss="mse")
return model, state_variable_input
def initialize_critic(self):
state_variable_input = Input(shape=(3, ))
action_input = Input(shape=(1, ))
init = TruncatedNormal(mean=0.0, stddev=0.02)
dense_state = Dense(128, activation="relu", kernel_initializer=init)(state_variable_input)
merge = Concatenate()([dense_state, action_input])
dense2 = Dense(128, activation="relu", kernel_initializer=init)(merge)
output = Dense(1, activation="linear", kernel_initializer=RandomUniform(-3e-3, 3e-3))(dense2)
model = Model(inputs=[state_variable_input, action_input],
outputs=output)
model.compile(optimizer="adam", loss="mse")
return model, state_variable_input, action_input
def train(self):
if len(self.memory) < self.BATCH_SIZE:
return
samples = random.sample(self.memory, self.BATCH_SIZE)
samples = [np.concatenate(x) for x in zip(*samples)]
self.train_critic(samples)
self.train_actor(samples)
self.global_step += 1
def train_critic(self, samples):
cur_state_var, action, reward, new_state_var = samples
predicted_action = self.target_actor.predict([new_state_var])
future_reward = self.target_critic.predict([new_state_var, predicted_action])
Q = reward + self.gamma*future_reward
self.critic_model.train_on_batch([cur_state_var, action], Q)
def train_actor(self, samples):
cur_state_var, action, reward, new_state_var = samples
predicted_action = self.actor_model.predict([cur_state_var])
grads = self.sess.run([self.critic_grads], feed_dict={
self.critic_var_in: cur_state_var,
self.critic_action_in: predicted_action})
self.sess.run(self.optimize, feed_dict={
self.actor_var_in: cur_state_var,
self.actor_critic_grad: grads[0]
})
def update_actor_target(self):
actor_model_weights = self.actor_model.get_weights()
actor_target_weights = self.target_actor.get_weights()
for i in range(len(actor_target_weights)):
actor_target_weights[i] = self.tau * actor_model_weights[i] + (1-self.tau)*actor_target_weights[i]
self.target_actor.set_weights(actor_target_weights)
def update_critic_target(self):
critic_model_weights = self.critic_model.get_weights()
critic_target_weights = self.target_critic.get_weights()
for i in range(len(critic_target_weights)):
critic_target_weights[i] = self.tau * critic_model_weights[i] + (1-self.tau)*critic_target_weights[i]
self.target_critic.set_weights(critic_target_weights)
def update_model(self):
self.update_actor_target()
self.update_critic_target()
def act(self, cur_state_var, noise=None, env=None):
if env:
if np.random.random() < self.epsilon:
return env.action_space.sample()
else:
sh = cur_state_var.shape
action = self.actor_model.predict([cur_state_var], batch_size=1)[0]
return action
elif not noise:
if np.random.random() < self.epsilon:
return self.sample_action_space()
return self.actor_model.predict([cur_state_var], batch_size=1)[0]
else:
no = noise()
pred = self.actor_model.predict([cur_state_var], batch_size=1)[0]
return pred + no
def sample_action_space(self):
return np.array([random.uniform(-0.5, 0.5), random.uniform(-1.0, 1.0)]).reshape(2, )
def remember(self, cur_state_var, action, reward, new_state_var):
self.memory.append([cur_state_var, action, reward, new_state_var])