一 Value-Based
Q-Learning
(资料图片仅供参考)
Q-Learning是RL算法中Value-Based的算法，Q即为Q(s,a)就是在某一时刻的s状态下(s∈S)，采取 动作a (a∈A)能够获得收益的期望，环境会根据agent的动作反馈相应的回报reward。所以算法的主要思想就是将State与Action构建成一张Q-table来存储Q值，然后根据Q值来选取能够获得最大的收益的动作。 下面是Q-Learning的TensorFlow实现
import numpy as npimport pandas as pdclass QLearning: def __init__(self, actions, learning_rate=0.01, reward_decay=0.9, e_greedy=0.9): """ QLearning :param actions: :param learning_rate: :param reward_decay: :param e_greedy: """ self.actions = actions self.lr = learning_rate self.gamma = reward_decay self.epsilon = e_greedy self.q_table = pd.DataFrame(columns=self.actions) def chooseAction(self, observation): """Choose action with state and observation""" self.checkStateExist(observation) if np.random.uniform()<SELF.EPSILON: :]="" opt_actions="self.q_table.loc[observation," #="" ="="np.max(opt_actions)].index)"" return="" action="opt_actions.argmax()" updateparams(self,="" state,="" action,="" reward,="" self.checkstateexist(state_)="" q_pre="self.q_table.loc[state," state_="" !="terminal" self.gamma="" self.q_table.loc[state_,="" :].max()="" q_target="reward" self.q_table.loc[state,="" action]="" +="" *="" (q_target="" -="" q_pre)="" def="" checkstateexist(self,="" if="" state="" not="" in="" self.q_table="self.q_table.append(" pd.series([0]*len(self.actions),="" index="self.q_table.columns," name="state)" )
DQN
当状态动作很多时，Q-Learning使用Table存储Value的方式不再实用（甚至不可行）。
如何不使用Table而得到每个状态下采取各个动作的Value呢？DQN用神经网络将State映射到Value。 DQN是在Q-Learning的主框架上做了扩展，包括：
记忆库（用于重复学习，随机抽取的经历也打乱的状态之间的相关性，使神经网络的更新更有效率）MLP计算Q值暂时冻结Q_target参数（切断相关性），target网络用来计算Q现实
下面是DQN的TensorFlow实现
import tensorflow as tfimport numpy as npclass DeepQNet: def __init__(self, n_actions, n_features, learning_rate=0.01, reward_decay=0.9, e_greedy=0.9, update_target_iter=300, memory_size=500, batch_size=32, e_greedy_increment=None, output_graph=False, ): """ DQN :param n_actions: :param n_features: :param learning_rate: :param reward_decay: :param e_greedy: :param update_target_iter: :param memory_size: :param batch_size: :param e_greedy_increment: :param output_graph: """ self.n_actions = n_actions self.n_actions = n_actions self.n_features = n_features self.lr = learning_rate self.gamma = reward_decay self.epsilon_max = e_greedy self.update_target_iter = update_target_iter self.memory_size = memory_size self.batch_size = batch_size self.epsilon_increment = e_greedy_increment self.epsilon = 0 if e_greedy_increment is not None else self.epsilon_max # total learning step(Cooperate with update_target_iter in learn() to update the parameters of target net) self.learn_step_counter = 0 # memory: row = memory_size, col = observation + observation_ + action + reward self.memory = np.zeros((self.memory_size, self.n_features*2+2)) self._buildNet() self.sess = tf.Session() if output_graph: tf.summary.FileWriter("logs/", self.sess.graph) self.sess.run(tf.global_variables_initializer()) self.cost = [] def _buildNet(self): """"Build evaluate network and target network""" # build evaluate net self.state = tf.placeholder(tf.float32, [None, self.n_features], name="state") self.q_target = tf.placeholder(tf.float32, [None, self.n_actions], name="Q_target") with tf.variable_scope("evaluate_net"): c_names, n_l1 = ["evaluate_net_params", tf.GraphKeys.GLOBAL_VARIABLES], 10 w_initializer, b_initializer = tf.random_normal_initializer(0, 0.3), tf.constant_initializer(0.1) with tf.variable_scope("layer1"): w1 = tf.get_variable("w1", [self.n_features, n_l1], initializer=w_initializer, collections=c_names) b1 = tf.get_variable("b1", [1, n_l1], initializer=b_initializer, collections=c_names) l1 = tf.nn.relu(tf.matmul(self.state, w1) + b1) with tf.variable_scope("layer2"): w2 = tf.get_variable("w2", [n_l1, self.n_actions], initializer=w_initializer, collections=c_names) b2 = tf.get_variable("b2", [1, self.n_actions], initializer=b_initializer, collections=c_names) self.q_evaluate = tf.nn.relu(tf.matmul(l1, w2) + b2) with tf.variable_scope("loss"): self.loss = tf.reduce_mean(tf.squared_difference(self.q_target, self.q_evaluate)) with tf.variable_scope("train"): self.opt = tf.train.RMSPropOptimizer(self.lr).minimize(self.loss) # build target net self.state_ = tf.placeholder(tf.float32, [None, self.n_features], name="state_") with tf.variable_scope("target_net"): c_names = ["target_net_params", tf.GraphKeys.GLOBAL_VARIABLES] with tf.variable_scope("layer1"): w1 = tf.get_variable("w1", [self.n_features, n_l1], initializer=w_initializer, collections=c_names) b1 = tf.get_variable("b1", [1, n_l1], initializer=b_initializer, collections=c_names) l1 = tf.nn.relu(tf.matmul(self.state_, w1) + b1) with tf.variable_scope("layer2"): w2 = tf.get_variable("w2", [n_l1, self.n_actions], initializer=w_initializer, collections=c_names) b2 = tf.get_variable("b2", [1, self.n_actions], initializer=b_initializer, collections=c_names) self.q_next = tf.nn.relu(tf.matmul(l1, w2) + b2) def storeTransition(self, state, action, reward, state_): """Store the state, observation and reward experienced during the train process to enable batch training""" if not hasattr(self, "memory_counter"): self.memory_counter = 0 transition = np.hstack((state, [action, reward], state_)) index = self.memory_counter % self.memory_size self.memory[index, :] = transition self.memory_counter += 1 def chooseAction(self, observation): """Choose action with state and observation""" observation = observation[np.newaxis, :] if np.random.uniform() < self.epsilon: actions = self.sess.run(self.q_evaluate, feed_dict={self.state: observation}) action = np.argmax(actions) else: action = np.random.randint(0, self.n_actions) return action def updateTargetNet(self): """Update the target net with the latest evaluate net parameters""" evaluate_params = tf.get_collection("evaluate_net_params") target_params = tf.get_collection("target_net_params") self.sess.run([tf.assign(t, e) for t, e in zip(target_params, evaluate_params)]) def learn(self): # check to update target net if self.learn_step_counter % self.update_target_iter == 0: self.updateTargetNet() print("Update target net!") # Get batch training data from the memory if self.memory_counter > self.memory_size: sample_index = np.random.choice(self.memory_size, size=self.batch_size) else: sample_index = np.random.choice(self.memory_counter, size=self.batch_size) batch_memory = self.memory[sample_index, :] q_evaluate, q_next = self.sess.run([self.q_evaluate, self.q_next], feed_dict={self.state: batch_memory[:, 0:self.n_features], self.state_: batch_memory[:, -self.n_features:]}) q_target = q_evaluate.copy() batch_index = np.arange(self.batch_size, dtype=np.int32) eval_act_index = batch_memory[:, self.n_features].astype(int) reward = batch_memory[:, self.n_features + 1] # Related to memory format, here means [action, reward] q_target[batch_index, eval_act_index] = reward + self.gamma * np.max(q_next, axis=1) _, cost = self.sess.run([self.opt, self.loss], feed_dict={self.state: batch_memory[:, 0:self.n_features], self.q_target: q_target }) self.cost.append(cost) self.epsilon = self.epsilon + self.epsilon_increment if self.epsilon < self.epsilon_max else self.epsilon_max self.learn_step_counter += 1 def showCost(self): import matplotlib.pyplot as plt plt.plot(np.arange(len(self.cost)), self.cost) plt.ylabel("Cost") plt.xlabel("training steps") plt.show()
二 Policy-Based
直接输出动作，可以在连续区间内选择动作；而Value-Based要在连续区间中，对无数个动作计算价值选择行为是不可行的。
误差如何反向传递呢？没有误差，它的目的是选的动作在下次更有可能被选择，但怎么知道动作的好坏呢，用reward，reward小，动作在下次被选择的可能性增加的少。
Actor-Critic
Actor：Policy-Based，输入State，预测输出采取各种Action的概率。 Critic；Value-Based，输入State，预测输出当前State的Value，并与下一状态的next_stateValue求TD_error 在Actor-Critic中，Actor可以每一步都更新学习（而单纯的Policy-Based方法要在回合结束后才能更新）
但也带来了问题：由于两个网络在连续状态中更新参数，每次跟新前后的参数具有相关性，导致网络只能片面的看待问题，甚至学不到有效的参数，不能收敛。
TRPO
PPO
Deep Deterministic Policy Gradient（DDPG）
