简介

用一个具体的例子来演示策略梯度的应用。这里使用了OpenAI Gym中的CartPole-v0游戏来作为我们算法应用。这个游戏比较简单，基本要求就是控制下面的cart移动使连接在上面的pole保持垂直不倒。这个任务只有两个离散动作，要么向左用力，要么向右用力。而state状态就是这个cart的位置和速度， pole的角度和角速度，4维的特征。坚持到200分的奖励则为过关。
代码使用了一个三层的神经网络，输入层，一个隐藏层和一个输出层。下面我们看看关键部分的代码。代码参考了博客园的一个例子，并做了适当修改，适用于tensowflow2.0以上的版本。

#######################################################################
# Copyright (C)                                                       #
# 2016 - 2019 Pinard Liu(liujianping-ok@163.com)                      #
# https://www.cnblogs.com/pinard                                      #
# Permission given to modify the code as long as you keep this        #
# declaration at the top                                              #
#######################################################################
## https://www.cnblogs.com/pinard/p/10137696.html ##
## 强化学习(十三) 策略梯度(Policy Gradient) ##

import gym
import tensorflow as tf
import numpy as np
import random
from collections import deque

# Hyper Parameters
GAMMA = 0.95 # discount factor
LEARNING_RATE=0.01

class Policy_Gradient():
    def __init__(self, env):
        # init some parameters
        self.time_step = 0
        self.state_dim = env.observation_space.shape[0]
        self.action_dim = env.action_space.n
        self.ep_obs, self.ep_as, self.ep_rs = [], [], []
        self.create_softmax_network()

        # Init session
        self.session = tf.compat.v1.InteractiveSession()
        self.session.run(tf.compat.v1.global_variables_initializer())

    def create_softmax_network(self):
        # network weights
        tf.compat.v1.disable_eager_execution()

        W1 = self.weight_variable([self.state_dim, 20])
        b1 = self.bias_variable([20])
        W2 = self.weight_variable([20, self.action_dim])
        b2 = self.bias_variable([self.action_dim])
        # input layer
        self.state_input = tf.compat.v1.placeholder("float", [None, self.state_dim])
        self.tf_acts = tf.compat.v1.placeholder(tf.int32, [None, ], name="actions_num")
        self.tf_vt = tf.compat.v1.placeholder(tf.float32, [None, ], name="actions_value")
        # hidden layers
        h_layer = tf.nn.relu(tf.matmul(self.state_input, W1) + b1)
        # softmax layer
        self.softmax_input = tf.matmul(h_layer, W2) + b2
        #softmax output
        self.all_act_prob = tf.nn.softmax(self.softmax_input, name='act_prob')
        self.neg_log_prob = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=self.softmax_input,
                                                                      labels=self.tf_acts)
        self.loss = tf.reduce_mean(self.neg_log_prob * self.tf_vt)  # reward guided loss

        self.train_op = tf.compat.v1.train.AdamOptimizer(LEARNING_RATE).minimize(self.loss)

    def weight_variable(self, shape):
        initial = tf.compat.v1.truncated_normal(shape)
        return tf.Variable(initial)

    def bias_variable(self, shape):
        initial = tf.constant(0.01, shape=shape)
        return tf.Variable(initial)

    def choose_action(self, observation):
        prob_weights = self.session.run(self.all_act_prob, feed_dict={self.state_input: observation[np.newaxis, :]})
        action = np.random.choice(range(prob_weights.shape[1]), p=prob_weights.ravel())  # select action w.r.t the actions prob
        return action

    def store_transition(self, s, a, r):
        self.ep_obs.append(s)
        self.ep_as.append(a)
        self.ep_rs.append(r)

    def learn(self):

        discounted_ep_rs = np.zeros_like(self.ep_rs)
        running_add = 0
        for t in reversed(range(0, len(self.ep_rs))):
            running_add = running_add * GAMMA + self.ep_rs[t]
            discounted_ep_rs[t] = running_add

        discounted_ep_rs -= np.mean(discounted_ep_rs)
        discounted_ep_rs /= np.std(discounted_ep_rs)

        # train on episode
        self.session.run(self.train_op, feed_dict={
             self.state_input: np.vstack(self.ep_obs),
             self.tf_acts: np.array(self.ep_as),
             self.tf_vt: discounted_ep_rs,
        })

        self.ep_obs, self.ep_as, self.ep_rs = [], [], []    # empty episode data
# Hyper Parameters
ENV_NAME = 'CartPole-v0'
EPISODE = 3000 # Episode limitation
STEP = 3000 # Step limitation in an episode
TEST = 10 # The number of experiment test every 100 episode

def main():
  # initialize OpenAI Gym env and dqn agent
  env = gym.make(ENV_NAME)
  agent = Policy_Gradient(env)

  for episode in range(EPISODE):
    # initialize task
    state = env.reset()
    # Train
    for step in range(STEP):
      action = agent.choose_action(state) # e-greedy action for train
      next_state,reward,done,_ = env.step(action)
      agent.store_transition(state, action, reward)
      state = next_state
      if done:
        #print("stick for ",step, " steps")
        agent.learn()
        break

    # Test every 100 episodes
    if episode % 100 == 0:
      total_reward = 0
      for i in range(TEST):
        state = env.reset()
        for j in range(STEP):
          env.render()
          action = agent.choose_action(state) # direct action for test
          state,reward,done,_ = env.step(action)
          total_reward += reward
          if done:
            break
      ave_reward = total_reward/TEST
      print ('episode: ',episode,'Evaluation Average Reward:',ave_reward)

if __name__ == '__main__':
  main()

总结

在策略梯度中更新价值函数需要的是完整的序列，所以必须每个序列运行一次才能更新一次策略参数，学习效率来说相对较低，但相对于q学习，策略梯度更适用于连续动作的状态环境。