【强化学习】DQN（Deep Q network）原理及实现

一、原理

DQN为融合了神经网络和Q-learning的方法。

面对复杂问题，state数量巨多，传统的表格学习已经不能满足此种情况。神经网络的的工作模式为通过对输入进行处理学习得到结果的过程。神经网络应用到强化学习中时，输入为状态和动作，价值作为其输出，或者输入为状态，输出为最大值的动作，省略了需要用表格记录动作及状态的过程，可更好的应用于复杂状态下的处理。

DQN中还有两种机理用于提升。一种为Experience replay（经验回放），随机对之前的经历进行学习，使其更新更有效率。Fixed Q-targets 也是一种打乱相关性的机理。

二、代码实现

1. RL_brain

建立一个数据库和一个暂时冻结的q_target参数，需要建立两个不通过的神经网络。两个神经网络一个有最新的参数，一个老的参数。

import numpy as np

import pandas as pd

import tensorflow as tf

np.random.seed(1)

tf.set_random_seed(1)

# Deep Q Network off-policy

class DeepQNetwork:

    def __init__(

            self,

            n_actions,

            n_features,

            learning_rate=0.01,

            reward_decay=0.9,

            e_greedy=0.9,

            replace_target_iter=300,

            memory_size=500,

            batch_size=32,

            e_greedy_increment=None,

            output_graph=True,

    ):

        self.n_actions = n_actions

        self.n_features = n_features

        self.lr = learning_rate

        self.gamma = reward_decay

        self.epsilon_max = e_greedy

        self.replace_target_iter = replace_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

        self.learn_step_counter = 0

        # initialize zero memory [s, a, r, s_]

        self.memory = np.zeros((self.memory_size, n_features * 2 + 2))

        # consist of [target_net, evaluate_net]

        self._build_net()

        t_params = tf.get_collection('target_net_params')

        e_params = tf.get_collection('eval_net_params')

        self.replace_target_op = [tf.assign(t, e) for t, e in zip(t_params, e_params)]

        self.sess = tf.Session()

        if output_graph:

            # $ tensorboard --logdir=logs

            # tf.train.SummaryWriter soon be deprecated, use following

            tf.summary.FileWriter("logs/", self.sess.graph)

        self.sess.run(tf.global_variables_initializer())

        self.cost_his = []

    def _build_net(self):

        # ------------------ build evaluate_net ------------------

        self.s = tf.placeholder(tf.float32, [None, self.n_features], name='s')  # input

        self.q_target = tf.placeholder(tf.float32, [None, self.n_actions], name='Q_target')  # for calculating loss

        with tf.variable_scope('eval_net'):

            # c_names(collections_names) are the collections to store variables

            c_names, n_l1, w_initializer, b_initializer = \

                ['eval_net_params', tf.GraphKeys.GLOBAL_VARIABLES], 10, \

                tf.random_normal_initializer(0., 0.3), tf.constant_initializer(0.1)  # config of layers

            # first layer. collections is used later when assign to target net

            #建立了两层神经网络l1和l2

            with tf.variable_scope('l1'):

                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.s, w1) + b1)

            # second layer. collections is used later when assign to target net

            with tf.variable_scope('l2'):

                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_eval = tf.matmul(l1, w2) + b2

        with tf.variable_scope('loss'):

            self.loss = tf.reduce_mean(tf.squared_difference(self.q_target, self.q_eval))

        with tf.variable_scope('train'):

            self._train_op = tf.train.RMSPropOptimizer(self.lr).minimize(self.loss)

        # ------------------ build target_net ------------------

        self.s_ = tf.placeholder(tf.float32, [None, self.n_features], name='s_')    # input

        with tf.variable_scope('target_net'):

            # c_names(collections_names) are the collections to store variables

            c_names = ['target_net_params', tf.GraphKeys.GLOBAL_VARIABLES]

            # first layer. collections is used later when assign to target net

            with tf.variable_scope('l1'):

                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.s_, w1) + b1)

            # second layer. collections is used later when assign to target net

            with tf.variable_scope('l2'):

                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.matmul(l1, w2) + b2

    def store_transition(self, s, a, r, s_):

        if not hasattr(self, 'memory_counter'):

            self.memory_counter = 0

        transition = np.hstack((s, [a, r], s_))

        # replace the old memory with new memory

        index = self.memory_counter % self.memory_size

        self.memory[index, :] = transition

        self.memory_counter += 1

    def choose_action(self, observation):

        # to have batch dimension when feed into tf placeholder

        observation = observation[np.newaxis, :]

        if np.random.uniform() < self.epsilon:

            # forward feed the observation and get q value for every actions

            actions_value = self.sess.run(self.q_eval, feed_dict={self.s: observation})

            action = np.argmax(actions_value)

        else:

            action = np.random.randint(0, self.n_actions)

        return action

    def learn(self):

        # check to replace target parameters

        if self.learn_step_counter % self.replace_target_iter == 0:

            self.sess.run(self.replace_target_op)

            print('\ntarget_params_replaced\n')

        # sample batch memory from all 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_next, q_eval = self.sess.run(

            [self.q_next, self.q_eval],

            feed_dict={

                self.s_: batch_memory[:, -self.n_features:],  # fixed params

                self.s: batch_memory[:, :self.n_features],  # newest params

            })

        # change q_target w.r.t q_eval's action

        q_target = q_eval.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]

        q_target[batch_index, eval_act_index] = reward + self.gamma * np.max(q_next, axis=1)

        """

        For example in this batch I have 2 samples and 3 actions:

        q_eval =

        [[1, 2, 3],

         [4, 5, 6]]

        q_target = q_eval =

        [[1, 2, 3],

         [4, 5, 6]]

        Then change q_target with the real q_target value w.r.t the q_eval's action.

        For example in:

            sample 0, I took action 0, and the max q_target value is -1;

            sample 1, I took action 2, and the max q_target value is -2:

        q_target =

        [[-1, 2, 3],

         [4, 5, -2]]

        So the (q_target - q_eval) becomes:

        [[(-1)-(1), 0, 0],

         [0, 0, (-2)-(6)]]

        We then backpropagate this error w.r.t the corresponding action to network,

        leave other action as error=0 cause we didn't choose it.

        """

        # train eval network

        _, self.cost = self.sess.run([self._train_op, self.loss],

                                     feed_dict={self.s: batch_memory[:, :self.n_features],

                                                self.q_target: q_target})

        self.cost_his.append(self.cost)

        # increasing epsilon

        self.epsilon = self.epsilon + self.epsilon_increment if self.epsilon < self.epsilon_max else self.epsilon_max

        self.learn_step_counter += 1

    def plot_cost(self):

        import matplotlib.pyplot as plt

        plt.plot(np.arange(len(self.cost_his)), self.cost_his)

        plt.ylabel('Cost')

        plt.xlabel('training steps')

        plt.show()

2.算法更新

from maze_env import Maze

from RL_brain import DeepQNetwork

def run_maze():

    step = 0

    for episode in range(300):

        # initial observation

        observation = env.reset()

        while True:

            # fresh env

            env.render()

            # RL choose action based on observation

            action = RL.choose_action(observation)

            # RL take action and get next observation and reward

            observation_, reward, done = env.step(action)

            RL.store_transition(observation, action, reward, observation_)

            if (step > 200) and (step % 5 == 0):#当记忆库有了一定的数量再进行学习，过了200没5步学习一次

                RL.learn()

            # swap observation

            observation = observation_

            # break while loop when end of this episode

            if done:

                break

            step += 1

    # end of game

    print('game over')

    env.destroy()

if __name__ == "__main__":

    # maze game

    env = Maze()

    RL = DeepQNetwork(env.n_actions, env.n_features,

                      learning_rate=0.01,

                      reward_decay=0.9,

                      e_greedy=0.9,

                      replace_target_iter=200,

                      memory_size=2000,

                      # output_graph=True

                      )

    env.after(100, run_maze)

    env.mainloop()

    RL.plot_cost()

转自morvan Zhou