前往小程序,Get更优阅读体验!
立即前往
首页
学习
活动
专区
工具
TVP
发布
社区首页 >专栏 >强化学习之Q-learning简介

强化学习之Q-learning简介

作者头像
用户1147754
修改2019-08-07 16:20:39
3K0
修改2019-08-07 16:20:39
举报
文章被收录于专栏:YoungGy

强化学习在alphago中大放异彩,本文将简要介绍强化学习的一种q-learning。先从最简单的q-table下手,然后针对state过多的问题引入q-network,最后通过两个例子加深对q-learning的理解。

强化学习

强化学习通常包括两个实体agentenvironment。两个实体的交互如下,在environmentstatest下,agent采取actionat进而得到rewardrt 并进入statest+1。

强化学习的问题,通常有如下特点:

  • 不同的action产生不同的reward
  • reward有延迟性
  • 对某个action的reward是基于当前的state的

Q-learning

Q-Table

Q-learning的核心是Q-table。Q-table的行和列分别表示stateaction的值,Q-table的值Q(s,a)Q(s,a)衡量当前states采取actiona到底有多好。

Bellman Equation

在训练的过程中,我们使用Bellman Equation去更新Q-table。

Bellman Equation解释如下:Q(s,a)表示成当前s采取a后的即时r,加上折价γ后的最大reward max(Q(s′,a′)。

算法

根据Bellman Equation,学习的最终目的是得到Q-table,算法如下:

  1. 外循环模拟次数num_episodes
  2. 内循环每次模拟最大步数num_steps
  3. 根据当前的state和q-table选择action(可加入随机性)
  4. 根据当前的state和action获得下一步的state和reward
  5. 更新q-table: Q[s,a] = Q[s,a] + lr*(r + y*np.max(Q[s1,:]) - Q[s,a])

实例

以FrozenLake为例,代码如下:

代码语言:javascript
复制
# import lib
import gym
import numpy as np

# Load the environment
env = gym.make('FrozenLake-v0')

# Implement Q-Table learning algorithm
#Initialize table with all zeros
Q = np.zeros([env.observation_space.n,env.action_space.n])
# Set learning parameters
lr = .8
y = .95
num_episodes = 2000
#create lists to contain total rewards and steps per episode
#jList = []
rList = []
for i in range(num_episodes):
    #Reset environment and get first new observation
    s = env.reset()
    rAll = 0
    d = False
    j = 0
    #The Q-Table learning algorithm
    while j < 99:
        j+=1
        #Choose an action by greedily (with noise) picking from Q table
        a = np.argmax(Q[s,:] + np.random.randn(1,env.action_space.n)*(1./(i+1)))
        #Get new state and reward from environment
        s1,r,d,_ = env.step(a)
        #Update Q-Table with new knowledge
        Q[s,a] = Q[s,a] + lr*(r + y*np.max(Q[s1,:]) - Q[s,a])
        rAll += r
        s = s1
        if d == True:
            break
    #jList.append(j)
    rList.append(rAll)

print "Score over time: " +  str(sum(rList)/num_episodes)
print "Final Q-Table Values"
print Q

Deep-Q-learning

q-table存在一个问题,真实情况的state可能无穷多,这样q-table就会无限大,解决这个问题的办法是通过神经网络实现q-table。输入state,输出不同action的q-value。

Experience replay

强化学习由于state之间的相关性存在稳定性的问题,解决的办法是在训练的时候存储当前训练的状态到记忆体M,更新参数的时候随机从M中抽样mini-batch进行更新。

具体地,M中存储的数据类型为 <s,a,r,s′><s, a, r, s'>,M有最大长度的限制,以保证更新采用的数据都是最近的数据。

Exploration - Exploitation

  • Exploration:在刚开始训练的时候,为了能够看到更多可能的情况,需要对action加入一定的随机性。
  • Exploitation:随着训练的加深,逐渐降低随机性,也就是降低随机action出现的概率。

算法

实例

CartPole

代码语言:javascript
复制
# import lib
import gym
import tensorflow as tf
import numpy as np

# Create the Cart-Pole game environment
env = gym.make('CartPole-v0')

# Q-network
class QNetwork:
    def __init__(self, learning_rate=0.01, state_size=4, 
                 action_size=2, hidden_size=10, 
                 name='QNetwork'):
        # state inputs to the Q-network
        with tf.variable_scope(name):
            self.inputs_ = tf.placeholder(tf.float32, [None, state_size], name='inputs')

            # One hot encode the actions to later choose the Q-value for the action
            self.actions_ = tf.placeholder(tf.int32, [None], name='actions')
            one_hot_actions = tf.one_hot(self.actions_, action_size)

            # Target Q values for training
            self.targetQs_ = tf.placeholder(tf.float32, [None], name='target')

            # ReLU hidden layers
            self.fc1 = tf.contrib.layers.fully_connected(self.inputs_, hidden_size)
            self.fc2 = tf.contrib.layers.fully_connected(self.fc1, hidden_size)

            # Linear output layer
            self.output = tf.contrib.layers.fully_connected(self.fc2, action_size, 
                                                            activation_fn=None)

            ### Train with loss (targetQ - Q)^2
            # output has length 2, for two actions. This next line chooses
            # one value from output (per row) according to the one-hot encoded actions.
            self.Q = tf.reduce_sum(tf.multiply(self.output, one_hot_actions), axis=1)

            self.loss = tf.reduce_mean(tf.square(self.targetQs_ - self.Q))
            self.opt = tf.train.AdamOptimizer(learning_rate).minimize(self.loss)

# Experience replay
from collections import deque
class Memory():
    def __init__(self, max_size = 1000):
        self.buffer = deque(maxlen=max_size)

    def add(self, experience):
        self.buffer.append(experience)

    def sample(self, batch_size):
        idx = np.random.choice(np.arange(len(self.buffer)), 
                               size=batch_size, 
                               replace=False)
        return [self.buffer[ii] for ii in idx]

# hyperparameters
train_episodes = 1000          # max number of episodes to learn from
max_steps = 200                # max steps in an episode
gamma = 0.99                   # future reward discount

# Exploration parameters
explore_start = 1.0            # exploration probability at start
explore_stop = 0.01            # minimum exploration probability 
decay_rate = 0.0001            # exponential decay rate for exploration prob

# Network parameters
hidden_size = 64               # number of units in each Q-network hidden layer
learning_rate = 0.0001         # Q-network learning rate

# Memory parameters
memory_size = 10000            # memory capacity
batch_size = 20                # experience mini-batch size
pretrain_length = batch_size   # number experiences to pretrain the memory

tf.reset_default_graph()
mainQN = QNetwork(name='main', hidden_size=hidden_size, learning_rate=learning_rate)

# Populate the experience memory
# Initialize the simulation
env.reset()
# Take one random step to get the pole and cart moving
state, reward, done, _ = env.step(env.action_space.sample())

memory = Memory(max_size=memory_size)

# Make a bunch of random actions and store the experiences
for ii in range(pretrain_length):
    # Uncomment the line below to watch the simulation
    # env.render()

    # Make a random action
    action = env.action_space.sample()
    next_state, reward, done, _ = env.step(action)

    if done:
        # The simulation fails so no next state
        next_state = np.zeros(state.shape)
        # Add experience to memory
        memory.add((state, action, reward, next_state))

        # Start new episode
        env.reset()
        # Take one random step to get the pole and cart moving
        state, reward, done, _ = env.step(env.action_space.sample())
    else:
        # Add experience to memory
        memory.add((state, action, reward, next_state))
        state = next_state

# Training
# Now train with experiences
saver = tf.train.Saver()
rewards_list = []
with tf.Session() as sess:
    # Initialize variables
    sess.run(tf.global_variables_initializer())

    step = 0
    for ep in range(1, train_episodes):
        total_reward = 0
        t = 0
        while t < max_steps:
            step += 1
            # Uncomment this next line to watch the training
            env.render() 

            # Explore or Exploit
            explore_p = explore_stop + (explore_start - explore_stop)*np.exp(-decay_rate*step) 
            if explore_p > np.random.rand():
                # Make a random action
                action = env.action_space.sample()
            else:
                # Get action from Q-network
                feed = {mainQN.inputs_: state.reshape((1, *state.shape))}
                Qs = sess.run(mainQN.output, feed_dict=feed)
                action = np.argmax(Qs)

            # Take action, get new state and reward
            next_state, reward, done, _ = env.step(action)

            total_reward += reward

            if done:
                # the episode ends so no next state
                next_state = np.zeros(state.shape)
                t = max_steps

                print('Episode: {}'.format(ep),
                      'Total reward: {}'.format(total_reward),
                      'Training loss: {:.4f}'.format(loss),
                      'Explore P: {:.4f}'.format(explore_p))
                rewards_list.append((ep, total_reward))

                # Add experience to memory
                memory.add((state, action, reward, next_state))

                # Start new episode
                env.reset()
                # Take one random step to get the pole and cart moving
                state, reward, done, _ = env.step(env.action_space.sample())

            else:
                # Add experience to memory
                memory.add((state, action, reward, next_state))
                state = next_state
                t += 1

            # Sample mini-batch from memory
            batch = memory.sample(batch_size)
            states = np.array([each[0] for each in batch])
            actions = np.array([each[1] for each in batch])
            rewards = np.array([each[2] for each in batch])
            next_states = np.array([each[3] for each in batch])

            # Train network
            target_Qs = sess.run(mainQN.output, feed_dict={mainQN.inputs_: next_states})

            # Set target_Qs to 0 for states where episode ends
            episode_ends = (next_states == np.zeros(states[0].shape)).all(axis=1)
            target_Qs[episode_ends] = (0, 0)

            targets = rewards + gamma * np.max(target_Qs, axis=1)

            loss, _ = sess.run([mainQN.loss, mainQN.opt],
                                feed_dict={mainQN.inputs_: states,
                                           mainQN.targetQs_: targets,
                                           mainQN.actions_: actions})

    saver.save(sess, "checkpoints/cartpole.ckpt")

# Testing
test_episodes = 10
test_max_steps = 400
env.reset()
with tf.Session() as sess:
    saver.restore(sess, tf.train.latest_checkpoint('checkpoints'))

    for ep in range(1, test_episodes):
        t = 0
        while t < test_max_steps:
            env.render() 

            # Get action from Q-network
            feed = {mainQN.inputs_: state.reshape((1, *state.shape))}
            Qs = sess.run(mainQN.output, feed_dict=feed)
            action = np.argmax(Qs)

            # Take action, get new state and reward
            next_state, reward, done, _ = env.step(action)

            if done:
                t = test_max_steps
                env.reset()
                # Take one random step to get the pole and cart moving
                state, reward, done, _ = env.step(env.action_space.sample())

            else:
                state = next_state
                t += 1

env.close()

FrozenLake

代码语言:javascript
复制
# import lib
import gym
import numpy as np
import random
import tensorflow as tf
import matplotlib.pyplot as plt
%matplotlib inline

# laod env
env = gym.make('FrozenLake-v0')

# The Q-Network Approach
tf.reset_default_graph()

#These lines establish the feed-forward part of the network used to choose actions
inputs1 = tf.placeholder(shape=[1,16],dtype=tf.float32)
W = tf.Variable(tf.random_uniform([16,4],0,0.01))
Qout = tf.matmul(inputs1,W)
predict = tf.argmax(Qout,1)

#Below we obtain the loss by taking the sum of squares difference between the target and prediction Q values.
nextQ = tf.placeholder(shape=[1,4],dtype=tf.float32)
loss = tf.reduce_sum(tf.square(nextQ - Qout))
trainer = tf.train.GradientDescentOptimizer(learning_rate=0.1)
updateModel = trainer.minimize(loss)

# Training

init = tf.initialize_all_variables()

# Set learning parameters
y = .99
e = 0.1
num_episodes = 2000
#create lists to contain total rewards and steps per episode
jList = []
rList = []
with tf.Session() as sess:
    sess.run(init)
    for i in range(num_episodes):
        #Reset environment and get first new observation
        s = env.reset()
        rAll = 0
        d = False
        j = 0
        #The Q-Network
        while j < 99:
            j+=1
            #Choose an action by greedily (with e chance of random action) from the Q-network
            a,allQ = sess.run([predict,Qout],feed_dict={inputs1:np.identity(16)[s:s+1]})
            if np.random.rand(1) < e:
                a[0] = env.action_space.sample()
            #Get new state and reward from environment
            s1,r,d,_ = env.step(a[0])
            #Obtain the Q' values by feeding the new state through our network
            Q1 = sess.run(Qout,feed_dict={inputs1:np.identity(16)[s1:s1+1]})
            #Obtain maxQ' and set our target value for chosen action.
            maxQ1 = np.max(Q1)
            targetQ = allQ
            targetQ[0,a[0]] = r + y*maxQ1
            #Train our network using target and predicted Q values
            _,W1 = sess.run([updateModel,W],feed_dict={inputs1:np.identity(16)[s:s+1],nextQ:targetQ})
            rAll += r
            s = s1
            if d == True:
                #Reduce chance of random action as we train the model.
                e = 1./((i/50) + 10)
                break
        jList.append(j)
        rList.append(rAll)
print "Percent of succesful episodes: " + str(sum(rList)/num_episodes) + "%"

参考资料

  1. Simple Reinforcement Learning with Tensorflow Part 0: Q-Learning with Tables and Neural Networks
  2. Udacity Deep Learning Nano Degree
本文参与 腾讯云自媒体同步曝光计划,分享自作者个人站点/博客。
原始发表:2017年06月20日,如有侵权请联系 cloudcommunity@tencent.com 删除

本文分享自 作者个人站点/博客 前往查看

如有侵权,请联系 cloudcommunity@tencent.com 删除。

本文参与 腾讯云自媒体同步曝光计划  ,欢迎热爱写作的你一起参与!

评论
登录后参与评论
0 条评论
热度
最新
推荐阅读
目录
  • 强化学习
  • Q-learning
    • Q-Table
      • Bellman Equation
        • 算法
          • 实例
          • Deep-Q-learning
            • Experience replay
              • Exploration - Exploitation
                • 算法
                  • 实例
                    • CartPole
                    • FrozenLake
                • 参考资料
                领券
                问题归档专栏文章快讯文章归档关键词归档开发者手册归档开发者手册 Section 归档