DeepNetwork---tensorflow实现
https://github.com/zle1992/Reinforcement_Learning_Game
DeepQNetwork.py
1 import numpy as np 2 import tensorflow as tf 3 from abc import ABCMeta, abstractmethod 4 np.random.seed(1) 5 tf.set_random_seed(1) 6 7 import logging # 引入logging模块 8 logging.basicConfig(level=logging.DEBUG, 9 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 10 # 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 11 12 tfconfig = tf.ConfigProto() 13 tfconfig.gpu_options.allow_growth = True 14 session = tf.Session(config=tfconfig) 15 16 17 class DeepQNetwork(object): 18 __metaclass__ = ABCMeta 19 """docstring for DeepQNetwork""" 20 def __init__(self, 21 n_actions, 22 n_features, 23 learning_rate, 24 reward_decay, 25 e_greedy, 26 replace_target_iter, 27 memory_size, 28 e_greedy_increment, 29 output_graph, 30 log_dir, 31 ): 32 super(DeepQNetwork, self).__init__() 33 34 self.n_actions = n_actions 35 self.n_features = n_features 36 self.learning_rate=learning_rate 37 self.gamma=reward_decay 38 self.epsilon_max=e_greedy 39 self.replace_target_iter=replace_target_iter 40 self.memory_size=memory_size 41 self.epsilon_increment=e_greedy_increment 42 self.output_graph=output_graph 43 self.lr =learning_rate 44 # total learning step 45 self.learn_step_counter = 0 46 self.log_dir = log_dir 47 48 49 50 self.s = tf.placeholder(tf.float32,[None]+self.n_features,name='s') 51 self.s_next = tf.placeholder(tf.float32,[None]+self.n_features,name='s_next') 52 53 self.r = tf.placeholder(tf.float32,[None,],name='r') 54 self.a = tf.placeholder(tf.int32,[None,],name='a') 55 56 57 self.q_eval = self._build_q_net(self.s, scope='eval_net', trainable=True) 58 self.q_next = self._build_q_net(self.s_next, scope='target_net', trainable=False) 59 60 61 62 with tf.variable_scope('q_target'): 63 self.q_target = self.r + self.gamma * tf.reduce_max(self.q_next, axis=1, name='Qmax_s_') # shape=(None, ) 64 with tf.variable_scope('q_eval'): 65 a_indices = tf.stack([tf.range(tf.shape(self.a)[0], dtype=tf.int32), self.a], axis=1) 66 self.q_eval_wrt_a = tf.gather_nd(params=self.q_eval, indices=a_indices) # shape=(None, ) 67 with tf.variable_scope('loss'): 68 self.loss = tf.reduce_mean(tf.squared_difference(self.q_target, self.q_eval_wrt_a, name='TD_error')) 69 with tf.variable_scope('train'): 70 self._train_op = tf.train.RMSPropOptimizer(self.lr).minimize(self.loss) 71 72 73 74 t_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='target_net') 75 e_params = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='eval_net') 76 77 with tf.variable_scope("hard_replacement"): 78 self.target_replace_op=[tf.assign(t,e) for t,e in zip(t_params,e_params)] 79 80 81 82 self.sess = tf.Session() 83 if self.output_graph: 84 tf.summary.FileWriter(self.log_dir,self.sess.graph) 85 86 self.sess.run(tf.global_variables_initializer()) 87 88 self.cost_his =[] 89 90 @abstractmethod 91 def _build_q_net(self,x,scope,trainable): 92 raise NotImplementedError 93 94 def learn(self,data): 95 96 97 # check to replace target parameters 98 if self.learn_step_counter % self.replace_target_iter == 0: 99 self.sess.run(self.target_replace_op) 100 print('\ntarget_params_replaced\n') 101 102 batch_memory_s = data['s'], 103 batch_memory_a = data['a'], 104 batch_memory_r = data['r'], 105 batch_memory_s_ = data['s_'], 106 _, cost = self.sess.run( 107 [self._train_op, self.loss], 108 feed_dict={ 109 self.s: batch_memory_s, 110 self.a: batch_memory_a, 111 self.r: batch_memory_r, 112 self.s_next: batch_memory_s_, 113 }) 114 self.cost_his.append(cost) 115 116 # increasing epsilon 117 self.epsilon_max = self.epsilon_max + self.epsilon_increment if self.epsilon_max < self.epsilon_max else self.epsilon_max 118 self.learn_step_counter += 1 119 120 121 122 123 def choose_action(self,s): 124 s = s[np.newaxis,:] 125 aa = np.random.uniform() 126 #print("epsilon_max",self.epsilon_max) 127 if aa < self.epsilon_max: 128 action_value = self.sess.run(self.q_eval,feed_dict={self.s:s}) 129 action = np.argmax(action_value) 130 else: 131 action = np.random.randint(0,self.n_actions) 132 return action
Memory.py
1 import numpy as np 2 np.random.seed(1) 3 class Memory(object): 4 """docstring for Memory""" 5 def __init__(self, 6 n_actions, 7 n_features, 8 memory_size): 9 super(Memory, self).__init__() 10 self.memory_size = memory_size 11 self.cnt =0 12 13 self.s = np.zeros([memory_size]+n_features) 14 self.a = np.zeros([memory_size,]) 15 self.r = np.zeros([memory_size,]) 16 self.s_ = np.zeros([memory_size]+n_features) 17 18 def store_transition(self,s, a, r, s_): 19 #logging.info('store_transition') 20 index = self.cnt % self.memory_size 21 self.s[index] = s 22 self.a[index] = a 23 self.r[index] = r 24 self.s_[index] =s_ 25 self.cnt+=1 26 27 def sample(self,n): 28 #logging.info('sample') 29 #assert self.cnt>=self.memory_size,'Memory has not been fulfilled' 30 N = min(self.memory_size,self.cnt) 31 indices = np.random.choice(N,size=n) 32 d ={} 33 d['s'] = self.s[indices][0] 34 d['s_'] = self.s_[indices][0] 35 d['r'] = self.r[indices][0] 36 d['a'] = self.a[indices][0] 37 return d
主函数
1 import gym 2 import numpy as np 3 import tensorflow as tf 4 5 from Memory import Memory 6 from DeepQNetwork import DeepQNetwork 7 8 np.random.seed(1) 9 tf.set_random_seed(1) 10 11 import logging # 引入logging模块 12 logging.basicConfig(level=logging.DEBUG, 13 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 14 # 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 15 16 tfconfig = tf.ConfigProto() 17 tfconfig.gpu_options.allow_growth = True 18 session = tf.Session(config=tfconfig) 19 20 class DeepQNetwork4CartPole(DeepQNetwork): 21 """docstring for ClassName""" 22 def __init__(self, **kwargs): 23 super(DeepQNetwork4CartPole, self).__init__(**kwargs) 24 25 def _build_q_net(self,x,scope,trainable): 26 w_initializer, b_initializer = tf.random_normal_initializer(0., 0.3), tf.constant_initializer(0.1) 27 28 with tf.variable_scope(scope): 29 e1 = tf.layers.dense(inputs=x, 30 units=32, 31 bias_initializer = b_initializer, 32 kernel_initializer=w_initializer, 33 activation = tf.nn.relu, 34 trainable=trainable) 35 q = tf.layers.dense(inputs=e1, 36 units=self.n_actions, 37 bias_initializer = b_initializer, 38 kernel_initializer=w_initializer, 39 activation = tf.nn.sigmoid, 40 trainable=trainable) 41 42 return q 43 44 45 46 47 batch_size = 64 48 49 memory_size =2000 50 #env = gym.make('Breakout-v0') #离散 51 env = gym.make('CartPole-v0') #离散 52 53 54 n_features= list(env.observation_space.shape) 55 n_actions= env.action_space.n 56 57 env = env.unwrapped 58 59 def run(): 60 61 RL = DeepQNetwork4CartPole( 62 n_actions=n_actions, 63 n_features=n_features, 64 learning_rate=0.01, 65 reward_decay=0.9, 66 e_greedy=0.9, 67 replace_target_iter=200, 68 memory_size=memory_size, 69 e_greedy_increment=None, 70 output_graph=True, 71 log_dir = 'log/DeepQNetwork4CartPole/', 72 ) 73 74 memory = Memory(n_actions,n_features,memory_size=memory_size) 75 76 77 step = 0 78 ep_r = 0 79 for episode in range(2000): 80 # initial observation 81 observation = env.reset() 82 83 while True: 84 85 86 # RL choose action based on observation 87 action = RL.choose_action(observation) 88 # logging.debug('action') 89 # print(action) 90 # RL take action and get_collectiot next observation and reward 91 observation_, reward, done, info=env.step(action) # take a random action 92 93 # the smaller theta and closer to center the better 94 x, x_dot, theta, theta_dot = observation_ 95 r1 = (env.x_threshold - abs(x))/env.x_threshold - 0.8 96 r2 = (env.theta_threshold_radians - abs(theta))/env.theta_threshold_radians - 0.5 97 reward = r1 + r2 98 99 100 101 102 memory.store_transition(observation, action, reward, observation_) 103 104 105 if (step > 200) and (step % 5 == 0): 106 107 data = memory.sample(batch_size) 108 RL.learn(data) 109 #print('step:%d----reward:%f---action:%d'%(step,reward,action)) 110 # swap observation 111 observation = observation_ 112 ep_r += reward 113 # break while loop when end of this episode 114 if(episode>700): 115 env.render() # render on the screen 116 if done: 117 print('episode: ', episode, 118 'ep_r: ', round(ep_r, 2), 119 ' epsilon: ', round(RL.epsilon_max, 2)) 120 ep_r = 0 121 122 break 123 step += 1 124 125 # end of game 126 print('game over') 127 env.destroy() 128 129 def main(): 130 131 run() 132 133 134 135 if __name__ == '__main__': 136 main() 137 #run2()
