乒乓球教程

强化学习实战（⼀）（tensorlayer乒乓球教程）

运⾏乒乓球例⼦

在本教程的第⼆部分，我们将运⾏⼀个深度强化学自我心理分析 习的例⼦，它在Karpathy的两篇博客DeepReinforcementLearning:Pongfrom

Pixels有介绍。

pythontutorial_atari_

在运⾏教程代码之前您需要安装OpenAIgymenvironment，焖鸡的做法 它提供了⼤量强化学习常⽤的游戏环境。如果⼀切运⾏正常，您将得到以

下的输出：

[2016-07-1209:31:59,760]Makingnewenv:Pong-v0

[TL]InputLayerinput_layer(?,6400)

[TL]DenLayerrelu1:200,relu

[TL]DenLayeroutput_layer:3,identity

param0:(6400,200)(mean:-0.000009median:-0.000018std:0.017393)

param1:(200,)(mean:0.000000median:0.000000std:0.000000)

param2:(200,3)(mean:0.002239median:0.003122std:0.096611)

param3:(3,)(mean:0.000000median:0.000000std:0.000000)

numofparams:1280803

layer0:Ten消亡的近义词 sor(“Relu:0”,shape=(?,200),dtype=float32)

layer1:Tensor(“add_1:0”,shape=(?,3),dtype防火顺口溜 =float32)

episode0:game0took0.17381s,reward:-1.000000

episode0:game1took0.12629s,reward:1.000000

episode0:game2took0.17082s,reward:-1.000000

episode0:game3took0.08944s,reward:-1.000000

episode0:game4took0.09446s,reward:-胖人穿搭 1.000000

episode0:game5took0.09440s,reward:-1.000000

episode0:game6took0.32798s,reward:-1.000000

episode0:game7took0.74437s,reward:-1.000000

episode0:game8took0.43013s,reward:-1.000000

episode0:game9took0.42496s,reward:-1.000000

episode0:game10took0.37128s,reward:-1.000000

episode0:game11took0.08979s,reward:-1.000000

episode0:game12took0.09138s,reward:-1.000000

episode0:game13took0.09142s,reward:-1.000000

episode0:game14took0.09639s,reward:-1.000000

episode0:game15took0.09852s,reward:-1.000000

episode0:game16took0.09984s,rewa木瓜炖牛奶 rd:-1.000000

episode0:game17took0.09575s,reward:-1.000000

episode0:game18took0.09416s,reward:-1.000000

episode0:game19took0.08674s,reward:-1.000000

episode0:game20took0.09628s,reward:-1.000000

gmean:-20.000000

episode1:game0took0.09910s,rewar1960年属什么 d:-1.000000

episode1:gam喜爱反义词 e1took0.17056s,reward:-1.000000

episode1:game2took0.09306s,reward:-1.000000

episode1:game3took0.09556s,reward:-1.000000

episode1:game4took0.12520s,reward:1.000000

episode1:game5t怎么做咖喱 ook0.17348s,reward:-1.000000

episode1:game6took0.09415s,reward:-1.000000

这个例⼦让神经⽹络通过游戏画⾯来学习如何像⼈类⼀样打乒乓球。神经⽹络将于伪AI电脑对战不断地对战，最后学会战胜它。在经过

15000个序列的训练之后，神经⽹络就可以赢得20%的⽐赛。在20000个序列的训练之后，神经⽹络可以赢得35%的⽐赛，我们可以看

到计算机学的越来越快，这是因为它有更多的胜利的数据来进⾏训练。训练了30000个序列后，神经⽹络再也不会输了。

render=Fallolfps

resume=Fal

如果您想显⽰游戏过程，那就设置render为True。当您再次运⾏该代码，您可以设置resume为True，那么代码将加载现有的模型并

且会基于它继续训练。

下⾯来介绍安装和运⾏demo教程。

pipinstallgym

OpenAIGym是开发和⽐较强化学习算法的⼯具包。

强化学习关注的是做出好决策，⽽监督式学习和⾮监督式学习主要关注的是做出预测。

强化学习有两个基本概念：环境（即外部世界）和智能体（即你正在编写的算法）。智能体向环境发送⾏为，环境回复观察和奖励（即分

数）。

OpenAIGym由两部分组成：

开源库：⼀个测试问题集合—环境（environment），可以⽤于⾃⼰的强化学习算法开发，这些环境有共享的接⼝，允许⽤户设计通

⽤的算法

Gym服务：⼀个站点和API，允许⽤户对他们训练的算法进⾏性能⽐较。

运⾏⼀个简单例⼦，移动平台使⽊棒不掉落。

importgym

rsimportMonitor

env=(‘CartPole-v0’)

env=Monitor(env,directory=’D:其他技术⽂献强化_深度学习gymcartpole-experiment-1’,video_callable=Fal,

write_upon_ret=True)

fori_episodeinrange(20):

obrvation=()

fortinrange(100):

()

print(obrvation)

action=_()

obrvation,reward,done,info=(action)

ifdone:

print(“Episodefinishedafter{}timesteps”.format(t+1))

break

接下来开始进⼊tensorlayer乒乓球教程：

pythontutorial_atari_

报错：Nomodulenamed‘atari_py’

再次运⾏pythontutorial_atari_

成功！