一 Value-Based

Q-Learning

(資料圖片僅供參考)

Q-Learning是RL算法中Value-Based的算法，Q即為Q(s,a)就是在某一時(shí)刻的s狀態(tài)下(s∈S)，采取 動(dòng)作a (a∈A)能夠獲得收益的期望，環(huán)境會(huì)根據(jù)agent的動(dòng)作反饋相應(yīng)的回報(bào)reward。所以算法的主要思想就是將State與Action構(gòu)建成一張Q-table來存儲(chǔ)Q值，然后根據(jù)Q值來選取能夠獲得最大的收益的動(dòng)作。 下面是Q-Learning的TensorFlow實(shí)現(xiàn)

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

當(dāng)狀態(tài)動(dòng)作很多時(shí)，Q-Learning使用Table存儲(chǔ)Value的方式不再實(shí)用（甚至不可行）。

如何不使用Table而得到每個(gè)狀態(tài)下采取各個(gè)動(dòng)作的Value呢？DQN用神經(jīng)網(wǎng)絡(luò)將State映射到Value。 DQN是在Q-Learning的主框架上做了擴(kuò)展，包括：

記憶庫（用于重復(fù)學(xué)習(xí)，隨機(jī)抽取的經(jīng)歷也打亂的狀態(tài)之間的相關(guān)性，使神經(jīng)網(wǎng)絡(luò)的更新更有效率）MLP計(jì)算Q值暫時(shí)凍結(jié)Q_target參數(shù)（切斷相關(guān)性），target網(wǎng)絡(luò)用來計(jì)算Q現(xiàn)實(shí)

下面是DQN的TensorFlow實(shí)現(xiàn)

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

直接輸出動(dòng)作，可以在連續(xù)區(qū)間內(nèi)選擇動(dòng)作；而Value-Based要在連續(xù)區(qū)間中，對(duì)無數(shù)個(gè)動(dòng)作計(jì)算價(jià)值選擇行為是不可行的。

誤差如何反向傳遞呢？沒有誤差，它的目的是選的動(dòng)作在下次更有可能被選擇，但怎么知道動(dòng)作的好壞呢，用reward，reward小，動(dòng)作在下次被選擇的可能性增加的少。

Actor-Critic

Actor：Policy-Based，輸入State，預(yù)測(cè)輸出采取各種Action的概率。 Critic；Value-Based，輸入State，預(yù)測(cè)輸出當(dāng)前State的Value，并與下一狀態(tài)的next_stateValue求TD_error 在Actor-Critic中，Actor可以每一步都更新學(xué)習(xí)（而單純的Policy-Based方法要在回合結(jié)束后才能更新）

但也帶來了問題：由于兩個(gè)網(wǎng)絡(luò)在連續(xù)狀態(tài)中更新參數(shù)，每次跟新前后的參數(shù)具有相關(guān)性，導(dǎo)致網(wǎng)絡(luò)只能片面的看待問題，甚至學(xué)不到有效的參數(shù)，不能收斂。

TRPO

PPO

Deep Deterministic Policy Gradient（DDPG）