完整代码：https://github.com/zle1992/Reinforcement_Learning_Game
Policy Gradient 可以直接预测出动作，也可以预测连续动作，但是无法单步更新。
QLearning 先预测出Q值，根据Q值选动作，无法预测连续动作、或者动作种类多的情况，但是可以单步更新。
一句话概括 Actor Critic 方法:
结合了 Policy Gradient (Actor) 和 Function Approximation (Critic) 的方法.
Actor 基于概率选行为 Critic 基于 Actor 的行为评判行为的得分,
Actor 根据 Critic 的评分修改选行为的概率.
Actor Critic 方法的优势: 可以进行单步更新, 比传统的 Policy Gradient 要快.
Actor Critic 方法的劣势: 取决于 Critic 的价值判断, 但是 Critic 难收敛, 再加上 Actor 的更新, 就更难收敛. 为了解决收敛问题, Google Deepmind 提出了 Actor Critic 升级版 Deep Deterministic Policy Gradient. 后者融合了 DQN 的优势, 解决了收敛难的问题.
Actor Critic 方法与Policy Gradinet的区别：
Policy Gradinet 中的梯度下降 ：
grad[logPi(s,a) * v_t]
其中v_t是真实的reward ,通过记录每个epoch的每一个state,action,reward得到。
而Actor中的v_t 是td_error 由Critic估计得到，不一定准确哦。
Actor Critic 方法与DQN的区别：
DQN 评价网络与动作网络其实是一个网络，只是采用了TD的方法，用滞后的网络去评价当前的动作。
Actor-Critic 就是在求解策略的同时用值函数进行辅助，用估计的值函数替代采样的reward，提高样本利用率。
Q-learning 是一种基于值函数估计的强化学习方法，Policy Gradient是一种策略搜索强化学习方法
Critic估计td_error跟DQN一样，用到了贝尔曼方程，
贝尔曼方程 :
Critic利用的是V函数的贝尔曼方程，来得到TD_error,
gradient = grad[r + gamma * V(s_) - V(s)]
Q-learning 利用的是Q函数的贝尔曼方程，来更新Q函数。
q_target = r + gamma * maxq(s_next)
q_eval = maxq(s)
Actor网络的输入（st,at,TDerror）
Actor 网络与policy gradient 差不多，多分类网络，在算loss时候，policy gradient需要乘一个权重Vt,而Vt是根据回报R 累计计算的。
在Actor中，在算loss时候，loss的权重是TDerror
TDerror是Critic网络计算出来的。
Critic网络的输入（st,vt+1,r）,输出TDerror
V_eval = network(st)
# TD_error = (r+gamma*V_next) - V_eval
学习的时候输入：（st, r, st+1)
vt+1 = network(st+1)
Critic网络(st,vt+1,r)
ACNetwork.py
1 import os 2 import numpy as np 3 import tensorflow as tf 4 from abc import ABCMeta, abstractmethod 5 np.random.seed(1) 6 tf.set_random_seed(1) 7 8 import logging # 寮曞叆logging妯″潡 9 logging.basicConfig(level=logging.DEBUG, 10 format='%(asctime)s - %(filename)s[line:%(lineno)d] - %(levelname)s: %(message)s') # logging.basicConfig鍑芥暟瀵规棩蹇楃殑杈撳嚭鏍煎紡鍙婃柟寮忓仛鐩稿叧閰嶇疆 11 # 鐢变簬鏃ュ織鍩烘湰閰嶇疆涓�骇鍒��缃�负DEBUG锛屾墍浠ヤ竴涓嬫墦鍗颁俊鎭�皢浼氬叏閮ㄦ樉绀哄湪鎺у埗鍙颁笂 12 13 tfconfig = tf.ConfigProto() 14 tfconfig.gpu_options.allow_growth = True 15 session = tf.Session(config=tfconfig) 16 17 18 class ACNetwork(object): 19 __metaclass__ = ABCMeta 20 """docstring for ACNetwork""" 21 def __init__(self, 22 n_actions, 23 n_features, 24 learning_rate, 25 memory_size, 26 reward_decay, 27 output_graph, 28 log_dir, 29 model_dir, 30 ): 31 super(ACNetwork, self).__init__() 32 33 self.n_actions = n_actions 34 self.n_features = n_features 35 self.learning_rate=learning_rate 36 self.gamma=reward_decay 37 self.memory_size =memory_size 38 self.output_graph=output_graph 39 self.lr =learning_rate 40 41 self.log_dir = log_dir 42 43 self.model_dir = model_dir 44 # total learning step 45 self.learn_step_counter = 0 46 47 48 self.s = tf.placeholder(tf.float32,[None]+self.n_features,name='s') 49 self.s_next = tf.placeholder(tf.float32,[None]+self.n_features,name='s_next') 50 51 self.r = tf.placeholder(tf.float32,[None,],name='r') 52 self.a = tf.placeholder(tf.int32,[None,],name='a') 53 54 55 56 57 58 with tf.variable_scope('Critic'): 59 60 self.v = self._build_c_net(self.s, scope='v', trainable=True) 61 self.v_ = self._build_c_net(self.s_next, scope='v_next', trainable=False) 62 63 self.td_error =self.r + self.gamma * self.v_ - self.v 64 self.loss_critic = tf.square(self.td_error) 65 with tf.variable_scope('train'): 66 self.train_op_critic = tf.train.AdamOptimizer(self.lr).minimize(self.loss_critic) 67 68 69 70 with tf.variable_scope('Actor'): 71 self.acts_prob = self._build_a_net(self.s, scope='actor_net', trainable=True) 72 # this is negative log of chosen action 73 log_prob = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=self.acts_prob, labels=self.a) 74 75 self.loss_actor = tf.reduce_mean(log_prob*self.td_error) 76 with tf.variable_scope('train'): 77 self.train_op_actor = tf.train.AdamOptimizer(self.lr).minimize(-self.loss_actor) 78 79 80 self.sess = tf.Session() 81 if self.output_graph: 82 tf.summary.FileWriter(self.log_dir,self.sess.graph) 83 84 self.sess.run(tf.global_variables_initializer()) 85 86 self.cost_his =[0] 87 88 89 self.saver = tf.train.Saver() 90 91 if not os.path.exists(self.model_dir): 92 os.mkdir(self.model_dir) 93 94 checkpoint = tf.train.get_checkpoint_state(self.model_dir) 95 if checkpoint and checkpoint.model_checkpoint_path: 96 self.saver.restore(self.sess, checkpoint.model_checkpoint_path) 97 print ("Loading Successfully") 98 self.learn_step_counter = int(checkpoint.model_checkpoint_path.split('-')[-1]) + 1 99 100 101 @abstractmethod 102 def _build_a_net(self,x,scope,trainable): 103 104 raise NotImplementedError 105 def _build_c_net(self,x,scope,trainable): 106 107 raise NotImplementedError 108 def learn(self,data): 109 110 111 112 113 batch_memory_s = data['s'] 114 batch_memory_a = data['a'] 115 batch_memory_r = data['r'] 116 batch_memory_s_ = data['s_'] 117 118 119 120 _, cost = self.sess.run( 121 [self.train_op_critic, self.loss_critic], 122 feed_dict={ 123 self.s: batch_memory_s, 124 self.a: batch_memory_a, 125 self.r: batch_memory_r, 126 self.s_next: batch_memory_s_, 127 128 }) 129 130 _, cost = self.sess.run( 131 [self.train_op_actor, self.loss_actor], 132 feed_dict={ 133 self.s: batch_memory_s, 134 self.a: batch_memory_a, 135 self.r: batch_memory_r, 136 self.s_next: batch_memory_s_, 137 138 }) 139 140 141 self.cost_his.append(cost) 142 143 self.learn_step_counter += 1 144 # save network every 100000 iteration 145 if self.learn_step_counter % 10000 == 0: 146 self.saver.save(self.sess,self.model_dir,global_step=self.learn_step_counter) 147 148 149 150 def choose_action(self,s): 151 s = s[np.newaxis,:] 152 153 probs = self.sess.run(self.acts_prob,feed_dict={self.s:s}) 154 return np.random.choice(np.arange(probs.shape[1]), p=probs.ravel())
game.py
1 import sys 2 import gym 3 import numpy as np 4 import tensorflow as tf 5 sys.path.append('./') 6 sys.path.append('model') 7 8 from util import Memory ,StateProcessor 9 from ACNetwork import ACNetwork 10 np.random.seed(1) 11 tf.set_random_seed(1) 12 13 import logging # 引入logging模块 14 logging.basicConfig(level=logging.DEBUG, 15 format='%(asctime)s - %(filename)s[line:%(lineno)d] - %(levelname)s: %(message)s') # logging.basicConfig函数对日志的输出格式及方式做相关配置 16 # 由于日志基本配置中级别设置为DEBUG，所以一下打印信息将会全部显示在控制台上 17 import os 18 os.environ["CUDA_VISIBLE_DEVICES"] = "1" 19 tfconfig = tf.ConfigProto() 20 tfconfig.gpu_options.allow_growth = True 21 session = tf.Session(config=tfconfig) 22 23 24 25 class ACNetwork4CartPole(ACNetwork): 26 """docstring for ClassName""" 27 def __init__(self, **kwargs): 28 super(ACNetwork4CartPole, self).__init__(**kwargs) 29 30 def _build_a_net(self,x,scope,trainable): 31 w_initializer, b_initializer = tf.random_normal_initializer(0., 0.3), tf.constant_initializer(0.1) 32 33 with tf.variable_scope(scope): 34 e1 = tf.layers.dense(inputs=x, 35 units=32, 36 bias_initializer = b_initializer, 37 kernel_initializer=w_initializer, 38 activation = tf.nn.relu, 39 trainable=trainable) 40 q = tf.layers.dense(inputs=e1, 41 units=self.n_actions, 42 bias_initializer = b_initializer, 43 kernel_initializer=w_initializer, 44 activation = tf.nn.softmax, 45 trainable=trainable) 46 47 return q 48 49 def _build_c_net(self,x,scope,trainable): 50 w_initializer, b_initializer = tf.random_normal_initializer(0., 0.3), tf.constant_initializer(0.1) 51 52 with tf.variable_scope(scope): 53 e1 = tf.layers.dense(inputs=x, 54 units=32, 55 bias_initializer = b_initializer, 56 kernel_initializer=w_initializer, 57 activation = tf.nn.relu, 58 trainable=trainable) 59 q = tf.layers.dense(inputs=e1, 60 units=1, 61 bias_initializer = b_initializer, 62 kernel_initializer=w_initializer, 63 activation =None, 64 trainable=trainable) 65 66 return q 67 68 69 70 batch_size = 32 71 72 memory_size =100 73 #env = gym.make('Breakout-v0') #离散 74 env = gym.make('CartPole-v0') #离散 75 76 77 n_features= list(env.observation_space.shape) 78 n_actions= env.action_space.n 79 env = env.unwrapped 80 81 def run(): 82 83 RL = ACNetwork4CartPole( 84 n_actions=n_actions, 85 n_features=n_features, 86 learning_rate=0.01, 87 reward_decay=0.9, 88 89 memory_size=memory_size, 90 91 output_graph=True, 92 log_dir = 'log/ACNetwork4CartPole/', 93 94 model_dir = 'model_dir/ACNetwork4CartPole/' 95 ) 96 97 memory = Memory(n_actions,n_features,memory_size=memory_size) 98 99 100 step = 0 101 ep_r = 0 102 for episode in range(2000): 103 # initial observation 104 observation = env.reset() 105 106 while True: 107 108 109 # RL choose action based on observation 110 action = RL.choose_action(observation) 111 # logging.debug('action') 112 # print(action) 113 # RL take action and get_collectiot next observation and reward 114 observation_, reward, done, info=env.step(action) # take a random action 115 116 # the smaller theta and closer to center the better 117 x, x_dot, theta, theta_dot = observation_ 118 r1 = (env.x_threshold - abs(x))/env.x_threshold - 0.8 119 r2 = (env.theta_threshold_radians - abs(theta))/env.theta_threshold_radians - 0.5 120 reward = r1 + r2 121 122 123 124 125 memory.store_transition(observation, action, reward, observation_) 126 127 128 if (step > 200) and (step % 1 == 0): 129 130 data = memory.sample(batch_size) 131 RL.learn(data) 132 #print('step:%d----reward:%f---action:%d'%(step,reward,action)) 133 # swap observation 134 observation = observation_ 135 ep_r += reward 136 # break while loop when end of this episode 137 if(episode>700): 138 env.render() # render on the screen 139 if done: 140 print('step: ',step, 141 'episode: ', episode, 142 'ep_r: ', round(ep_r, 2), 143 'loss: ',RL.cost_his[-1] 144 ) 145 ep_r = 0 146 147 break 148 step += 1 149 150 # end of game 151 print('game over') 152 env.destroy() 153 154 def main(): 155 156 run() 157 158 159 160 if __name__ == '__main__': 161 main() 162 #run2()