python - Open AI Gym Cartpole 的策略梯度方法

Question

我是强化学习的初学者，正在尝试实现策略梯度方法来使用 Tensorflow 解决 Open AI Gym CartPole 任务。然而，我的代码似乎运行得非常慢；第一集的运行速度还可以，但是从第二集开始就很慢。为什么会这样，如何解决这个问题？

我的代码:

import tensorflow as tf
import numpy as np
import gym

env = gym.make('CartPole-v0')

class Policy:
    def __init__(self):
        self.input_layer_fake = tf.placeholder(tf.float32, [4,1])
        self.input_layer = tf.reshape(self.input_layer_fake, [1,4])
        self.dense1 = tf.layers.dense(inputs = self.input_layer, units = 4,
                                  activation = tf.nn.relu)
        self.logits = tf.layers.dense(inputs = self.dense1, units = 2,
                                  activation = tf.nn.relu)
    def predict(self, inputObservation):
        sess = tf.InteractiveSession()
        tf.global_variables_initializer().run()
        x = tf.reshape(inputObservation, [4,1]).eval()
        return (sess.run(self.logits, feed_dict = {self.input_layer_fake: x}))

    def train(self, features_array, labels_array):
        for i in range(np.shape(features_array)[0]):
            print("train")
            print(i)
            sess1 = tf.InteractiveSession()
            tf.global_variables_initializer().run()
            self.cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels = labels_array[i], logits = self.logits))
            self.train_step = tf.train.GradientDescentOptimizer(0.5).minimize(self.cross_entropy)
            y = tf.reshape(features_array[i], [4,1]).eval()
            sess1.run(self.train_step, feed_dict={self.input_layer_fake:y})

agent = Policy()
train_array = []
features_array = []
labels_array = []
main_sess = tf.InteractiveSession()
tf.global_variables_initializer().run()

for i_episode in range(100):
    observation = env.reset()

    for t in range(200):
        prevObservation = observation
        env.render()

        if np.random.uniform(0,1) < 0.2:
            action = env.action_space.sample()
        else:
            action = np.argmax(agent.predict((prevObservation)))

        observation, reward, done, info = env.step(action)
        add_in = np.random.uniform(0,1)
        if add_in < 0.5:
            features_array.append(prevObservation)
            sarPreprocessed = agent.predict(prevObservation)
            sarPreprocessed[0][action] = reward
            labels_array.append(sarPreprocessed)
        if done:
            break

    agent.train(features_array, labels_array)
    features_array = []
    labels_array = []

非常感谢任何帮助。

Answer 1

自从我尝试实现策略梯度以来，已经有一段时间了，但据我所知，问题是我在训练函数中使用了循环。

当我循环遍历 features_array 中的每个元素时，数组本身的长度不断增长(features_array 永远不会设置回 [] )，程序变慢。相反，我应该以“批量”方式进行训练，同时定期清除 features_array。

我在这里实现了一个更简洁的普通策略梯度算法版本: https://github.com/Ashboy64/rl-reimplementations/blob/master/Reimplementations/Vanilla-Policy-Gradient/vanilla_pg.py

可以在此处找到名为 PPO(近端策略优化)的性能更好的修改算法(仍然基于策略梯度)的实现: https://github.com/Ashboy64/rl-reimplementations/tree/master/Reimplementations/PPO