用Tensorflow基于Deep Q Learning DQN 玩Flappy Bird

前言

2013年DeepMind 在NIPS上发表Playing Atari with Deep Reinforcement Learning 一文，提出了DQN（Deep Q Network）算法，实现端到端学习玩Atari游戏，即只有像素输入，看着屏幕玩游戏。Deep Mind就凭借这个应用以6亿美元被Google收购。由于DQN的开源，在github上涌现了大量各种版本的DQN程序。但大多是复现Atari的游戏，代码量很大，也不好理解。

Flappy Bird是个极其简单又困难的游戏，风靡一时。在很早之前，就有人使用Q-Learning 算法来实现完Flappy Bird。http://sarvagyavaish.github.io/FlappyBirdRL/
但是这个的实现是通过获取小鸟的具体位置信息来实现的。

能否使用DQN来实现通过屏幕学习玩Flappy Bird是一个有意思的挑战。（话说本人和朋友在去年年底也考虑了这个idea，但当时由于不知道如何截取游戏屏幕只能使用具体位置来学习，不过其实也成功了）

最近，github上有人放出使用DQN玩Flappy Bird的代码，https://github.com/yenchenlin1994/DeepLearningFlappyBird【1】
该repo通过结合之前的repo成功实现了这个想法。这个repo对整个实现过程进行了较详细的分析，但是由于其DQN算法的代码基本采用别人的repo，代码较为混乱，不易理解。

为此，本人改写了一个版本https://github.com/songrotek/DRL-FlappyBird

对DQN代码进行了重新改写。本质上对其做了类的封装，从而使代码更具通用性。可以方便移植到其他应用。

当然，本文的目的是借Flappy Bird DQN这个代码来详细分析一下DQN算法极其使用。

DQN 伪代码

这个是NIPS13版本的伪代码：

Initialize replay memory D to size N
Initialize action-value function Q with random weights
for episode = 1, M doInitialize state s_1for t = 1, T doWith probability ϵ select random action a_totherwise select a_t=max_a  Q($s_t$,a; $θ_i$)Execute action a_t in emulator and observe r_t and s_(t+1)Store transition (s_t,a_t,r_t,s_(t+1)) in DSample a minibatch of transitions (s_j,a_j,r_j,s_(j+1)) from DSet y_j:=r_j for terminal s_(j+1)r_j+γ*max_(a^' )  Q(s_(j+1),a'; θ_i) for non-terminal s_(j+1)Perform a gradient step on (y_j-Q(s_j,a_j; θ_i))^2 with respect to θend for
end for

基本的分析详见Paper Reading 1 - Playing Atari with Deep Reinforcement Learning
基础知识详见Deep Reinforcement Learning 基础知识（DQN方面）

本文主要从代码实现的角度来分析如何编写Flappy Bird DQN的代码

编写FlappyBirdDQN.py

首先，FlappyBird的游戏已经编写好，是现成的。提供了很简单的接口：

nextObservation,reward,terminal = game.frame_step(action)

即输入动作，输出执行完动作的屏幕截图，得到的反馈reward，以及游戏是否结束。

那么，现在先把DQN想象为一个大脑，这里我们也用BrainDQN类来表示，这个类只需获取感知信息也就是上面说的观察（截图），反馈以及是否结束，然后输出动作即可。

完美的代码封装应该是这样。具体DQN里面如何存储。如何训练是外部不关心的。
因此，我们的FlappyBirdDQN代码只有如下这么短：

# -------------------------
# Project: Deep Q-Learning on Flappy Bird
# Author: Flood Sung
# Date: 2016.3.21
# -------------------------import cv2
import sys
sys.path.append("game/")
import wrapped_flappy_bird as game
from BrainDQN import BrainDQN
import numpy as np# preprocess raw image to 80*80 gray image
def preprocess(observation):observation = cv2.cvtColor(cv2.resize(observation, (80, 80)), cv2.COLOR_BGR2GRAY)ret, observation = cv2.threshold(observation,1,255,cv2.THRESH_BINARY)return np.reshape(observation,(80,80,1))def playFlappyBird():# Step 1: init BrainDQNbrain = BrainDQN()# Step 2: init Flappy Bird GameflappyBird = game.GameState()# Step 3: play game# Step 3.1: obtain init stateaction0 = np.array([1,0])  # do nothingobservation0, reward0, terminal = flappyBird.frame_step(action0)observation0 = cv2.cvtColor(cv2.resize(observation0, (80, 80)), cv2.COLOR_BGR2GRAY)ret, observation0 = cv2.threshold(observation0,1,255,cv2.THRESH_BINARY)brain.setInitState(observation0)# Step 3.2: run the gamewhile 1!= 0:action = brain.getAction()nextObservation,reward,terminal = flappyBird.frame_step(action)nextObservation = preprocess(nextObservation)brain.setPerception(nextObservation,action,reward,terminal)def main():playFlappyBird()if __name__ == '__main__':main()

核心部分就在while循环里面，由于要讲图像转换为80x80的灰度图，因此，加了一个preprocess预处理函数。

这里，显然只有有游戏引擎，换一个游戏是一样的写法，非常方便。

接下来就是编写BrainDQN.py 我们的游戏大脑

编写BrainDQN

基本架构：

class BrainDQN:def __init__(self):# init replay memoryself.replayMemory = deque()# init Q networkself.createQNetwork()def createQNetwork(self):def trainQNetwork(self):def setPerception(self,nextObservation,action,reward,terminal):def getAction(self):def setInitState(self,observation):

基本的架构也就只需要上面这几个函数，其他的都是多余了，接下来就是编写每一部分的代码。

CNN代码

也就是createQNetwork部分，这里采用如下图的结构（转自【1】）：

这里就不讲解整个流程了。主要是针对具体的输入类型和输出设计卷积和全连接层。

代码如下：

    def createQNetwork(self):# network weightsW_conv1 = self.weight_variable([8,8,4,32])b_conv1 = self.bias_variable([32])W_conv2 = self.weight_variable([4,4,32,64])b_conv2 = self.bias_variable([64])W_conv3 = self.weight_variable([3,3,64,64])b_conv3 = self.bias_variable([64])W_fc1 = self.weight_variable([1600,512])b_fc1 = self.bias_variable([512])W_fc2 = self.weight_variable([512,self.ACTION])b_fc2 = self.bias_variable([self.ACTION])# input layerself.stateInput = tf.placeholder("float",[None,80,80,4])# hidden layersh_conv1 = tf.nn.relu(self.conv2d(self.stateInput,W_conv1,4) + b_conv1)h_pool1 = self.max_pool_2x2(h_conv1)h_conv2 = tf.nn.relu(self.conv2d(h_pool1,W_conv2,2) + b_conv2)h_conv3 = tf.nn.relu(self.conv2d(h_conv2,W_conv3,1) + b_conv3)h_conv3_flat = tf.reshape(h_conv3,[-1,1600])h_fc1 = tf.nn.relu(tf.matmul(h_conv3_flat,W_fc1) + b_fc1)# Q Value layerself.QValue = tf.matmul(h_fc1,W_fc2) + b_fc2self.actionInput = tf.placeholder("float",[None,self.ACTION])self.yInput = tf.placeholder("float", [None]) Q_action = tf.reduce_sum(tf.mul(self.QValue, self.actionInput), reduction_indices = 1)self.cost = tf.reduce_mean(tf.square(self.yInput - Q_action))self.trainStep = tf.train.AdamOptimizer(1e-6).minimize(self.cost)

记住输出是Q值，关键要计算出cost，里面关键是计算Q_action的值，即该state和action下的Q值。由于actionInput是one hot vector的形式，因此tf.mul(self.QValue, self.actionInput)正好就是该action下的Q值。

training 部分。

这部分是代码的关键部分，主要是要计算y值，也就是target Q值。

    def trainQNetwork(self):# Step 1: obtain random minibatch from replay memoryminibatch = random.sample(self.replayMemory,self.BATCH_SIZE)state_batch = [data[0] for data in minibatch]action_batch = [data[1] for data in minibatch]reward_batch = [data[2] for data in minibatch]nextState_batch = [data[3] for data in minibatch]# Step 2: calculate y y_batch = []QValue_batch = self.QValue.eval(feed_dict={self.stateInput:nextState_batch})for i in range(0,self.BATCH_SIZE):terminal = minibatch[i][4]if terminal:y_batch.append(reward_batch[i])else:y_batch.append(reward_batch[i] + GAMMA * np.max(QValue_batch[i]))self.trainStep.run(feed_dict={self.yInput : y_batch,self.actionInput : action_batch,self.stateInput : state_batch})

其他部分

其他部分就比较容易了，这里直接贴出完整的代码：

# -----------------------------
# File: Deep Q-Learning Algorithm
# Author: Flood Sung
# Date: 2016.3.21
# -----------------------------import tensorflow as tf
import numpy as np
import random
from collections import deque class BrainDQN:# Hyper Parameters:ACTION = 2FRAME_PER_ACTION = 1GAMMA = 0.99 # decay rate of past observationsOBSERVE = 100000. # timesteps to observe before trainingEXPLORE = 150000. # frames over which to anneal epsilonFINAL_EPSILON = 0.0 # final value of epsilonINITIAL_EPSILON = 0.0 # starting value of epsilonREPLAY_MEMORY = 50000 # number of previous transitions to rememberBATCH_SIZE = 32 # size of minibatchdef __init__(self):# init replay memoryself.replayMemory = deque()# init Q networkself.createQNetwork()# init some parametersself.timeStep = 0self.epsilon = self.INITIAL_EPSILONdef createQNetwork(self):# network weightsW_conv1 = self.weight_variable([8,8,4,32])b_conv1 = self.bias_variable([32])W_conv2 = self.weight_variable([4,4,32,64])b_conv2 = self.bias_variable([64])W_conv3 = self.weight_variable([3,3,64,64])b_conv3 = self.bias_variable([64])W_fc1 = self.weight_variable([1600,512])b_fc1 = self.bias_variable([512])W_fc2 = self.weight_variable([512,self.ACTION])b_fc2 = self.bias_variable([self.ACTION])# input layerself.stateInput = tf.placeholder("float",[None,80,80,4])# hidden layersh_conv1 = tf.nn.relu(self.conv2d(self.stateInput,W_conv1,4) + b_conv1)h_pool1 = self.max_pool_2x2(h_conv1)h_conv2 = tf.nn.relu(self.conv2d(h_pool1,W_conv2,2) + b_conv2)h_conv3 = tf.nn.relu(self.conv2d(h_conv2,W_conv3,1) + b_conv3)h_conv3_flat = tf.reshape(h_conv3,[-1,1600])h_fc1 = tf.nn.relu(tf.matmul(h_conv3_flat,W_fc1) + b_fc1)# Q Value layerself.QValue = tf.matmul(h_fc1,W_fc2) + b_fc2self.actionInput = tf.placeholder("float",[None,self.ACTION])self.yInput = tf.placeholder("float", [None]) Q_action = tf.reduce_sum(tf.mul(self.QValue, self.actionInput), reduction_indices = 1)self.cost = tf.reduce_mean(tf.square(self.yInput - Q_action))self.trainStep = tf.train.AdamOptimizer(1e-6).minimize(self.cost)# saving and loading networkssaver = tf.train.Saver()self.session = tf.InteractiveSession()self.session.run(tf.initialize_all_variables())checkpoint = tf.train.get_checkpoint_state("saved_networks")if checkpoint and checkpoint.model_checkpoint_path:saver.restore(self.session, checkpoint.model_checkpoint_path)print "Successfully loaded:", checkpoint.model_checkpoint_pathelse:print "Could not find old network weights"def trainQNetwork(self):# Step 1: obtain random minibatch from replay memoryminibatch = random.sample(self.replayMemory,self.BATCH_SIZE)state_batch = [data[0] for data in minibatch]action_batch = [data[1] for data in minibatch]reward_batch = [data[2] for data in minibatch]nextState_batch = [data[3] for data in minibatch]# Step 2: calculate y y_batch = []QValue_batch = self.QValue.eval(feed_dict={self.stateInput:nextState_batch})for i in range(0,self.BATCH_SIZE):terminal = minibatch[i][4]if terminal:y_batch.append(reward_batch[i])else:y_batch.append(reward_batch[i] + GAMMA * np.max(QValue_batch[i]))self.trainStep.run(feed_dict={self.yInput : y_batch,self.actionInput : action_batch,self.stateInput : state_batch})# save network every 100000 iterationif self.timeStep % 10000 == 0:saver.save(self.session, 'saved_networks/' + 'network' + '-dqn', global_step = self.timeStep)def setPerception(self,nextObservation,action,reward,terminal):newState = np.append(nextObservation,self.currentState[:,:,1:],axis = 2)self.replayMemory.append((self.currentState,action,reward,newState,terminal))if len(self.replayMemory) > self.REPLAY_MEMORY:self.replayMemory.popleft()if self.timeStep > self.OBSERVE:# Train the networkself.trainQNetwork()self.currentState = newStateself.timeStep += 1def getAction(self):QValue = self.QValue.eval(feed_dict= {self.stateInput:[self.currentState]})[0]action = np.zeros(self.ACTION)action_index = 0if self.timeStep % self.FRAME_PER_ACTION == 0:if random.random() <= self.epsilon:action_index = random.randrange(self.ACTION)action[action_index] = 1else:action_index = np.argmax(QValue)action[action_index] = 1else:action[0] = 1 # do nothing# change episilonif self.epsilon > self.FINAL_EPSILON and self.timeStep > self.OBSERVE:self.epsilon -= (self.INITIAL_EPSILON - self.FINAL_EPSILON)/self.EXPLOREreturn actiondef setInitState(self,observation):self.currentState = np.stack((observation, observation, observation, observation), axis = 2)def weight_variable(self,shape):initial = tf.truncated_normal(shape, stddev = 0.01)return tf.Variable(initial)def bias_variable(self,shape):initial = tf.constant(0.01, shape = shape)return tf.Variable(initial)def conv2d(self,x, W, stride):return tf.nn.conv2d(x, W, strides = [1, stride, stride, 1], padding = "SAME")def max_pool_2x2(self,x):return tf.nn.max_pool(x, ksize = [1, 2, 2, 1], strides = [1, 2, 2, 1], padding = "SAME")

一共也只有160代码。
如果这个任务不使用深度学习，而是人工的从图像中找到小鸟，然后计算小鸟的轨迹，然后计算出应该怎么按键，那么代码没有好几千行是不可能的。深度学习大大减少了代码工作。

小结

本文从代码角度对于DQN做了一定的分析，对于DQN的应用，大家可以在此基础上做各种尝试。