我和我的项目合作伙伴目前在我们最新的大学项目中遇到问题.
我们的任务是实现一个玩Pong游戏的神经网络.我们将球的速度和球拍的位置传递给网络,并提供三个输出：UP DOWN DO_NOTHING.玩家获得11分后,我们将训练所有状态的网络,做出的决策以及做出的决策的奖励(请参阅reward_cal()).我们面临的问题是,仅根据学习率,损失就一直保持在特定值上.因此,即使我们将其视为严重错误,网络也总是做出相同的决定.
请帮助我们找出我们做错了什么,我们感谢每一个建议！下面是我们的代码,请随时询问是否有任何问题.我们对这个话题还很新,所以如果有什么完全愚蠢的话,请不要粗鲁：D
这是我们的代码：
import sys,pygame,time
import numpy as np
import random
from os.path import isfile
import keras
from keras.optimizers import SGD
from keras.layers import Dense
from keras.layers.core import Flatten
pygame.init()
pygame.mixer.init()
#surface of the game
width = 400
height = 600
black = 0,0 #RGB value
screen = pygame.display.set_mode((width,height),32)
#(Resolution(x,y),flags,colour depth)
font = pygame.font.SysFont('arial',36,bold=True)
pygame.display.set_caption('PyPong') #title of window
#consts for the game
acceleration = 0.0025 # ball becomes faster during the game
mousematch = 1
delay_time = 0
paddleP = pygame.image.load("schlaeger.gif")
playerRect = paddleP.get_rect(center = (200,550))
paddleC = pygame.image.load("schlaeger.gif")
comRect = paddleC.get_rect(center=(200,50))
ball = pygame.image.load("ball.gif")
ballRect = ball.get_rect(center=(200,300))
#Variables for the game
pointsPlayer = [0]
pointsCom = [0]
playermove = [0,0]
speedbar = [0,0]
speed = [6,6]
hitX = 0
#neural const
learning_rate = 0.01
number_of_actions = 3
filehandler = open('logfile.log','a')
filename = sys.argv[1]
#neural variables
states,action_prob_grads,rewards,action_probs = [],[],[]
reward_sum = 0
episode_number = 0
reward_sums = []
pygame.display.flip()
def pointcontrol(): #having a look at the points in the game and restart()
if pointsPlayer[0] >= 11:
print('Player Won ',pointsPlayer[0],'/',pointsCom[0])
restart(1)
return 1
if pointsCom[0] >= 11:
print('Computer Won ',pointsCom[0])
restart(1)
return 1
elif pointsCom[0] < 11 and pointsPlayer[0] < 11:
restart(0)
return 0
def restart(finished): #resetting the positions and the ball speed and
(if point limit was reached) the points
ballRect.center = 200,300
comRect.center = 200,50
playerRect.center = 200,550
speed[0] = 6
speed[1] = 6
screen.blit(paddleC,comRect)
screen.blit(paddleP,playerRect)
pygame.display.flip()
if finished:
pointsPlayer[0] = 0
pointsCom[0] = 0
def reward_cal(r,gamma = 0.99): #rewarding every move
discounted_r = np.zeros_like(r) #making zero array with size of
reward array
running_add = 0
for t in range(r.size - 1,-1): #iterating beginning in the end
if r[t] != 0: #if reward -1 or 1 (point made or lost)
running_add = 0
running_add = running_add * gamma + r[t] #making every move
before the point the same reward but a little bit smaller
discounted_r[t] = running_add #putting the value in the new
reward array
#e.g r = 000001000-1 -> discounted_r = 0.5 0.6 0.7 0.8 0.9 1 -0.7
-0.8 -0.9 -1 values are not really correct just to make it clear
return discounted_r
#neural net
model = keras.models.Sequential()
model.add(Dense(16,input_dim = (8),kernel_initializer =
'glorot_normal',activation = 'relu'))
model.add(Dense(32,kernel_initializer = 'glorot_normal',activation =
'relu'))
model.add(Dense(number_of_actions,activation='softmax'))
model.compile(loss = 'categorical_crossentropy',optimizer = 'adam')
model.summary()
if isfile(filename):
model.load_weights(filename)
# one ball movement before the AI gets to make a decision
ballRect = ballRect.move(speed)
reward_temp = 0.0
if ballRect.left < 0 or ballRect.right > width:
speed[0] = -speed[0]
if ballRect.top < 0:
pointsPlayer[0] += 1
reward_temp = 1.0
done = pointcontrol()
if ballRect.bottom > height:
pointsCom[0] += 1
done = pointcontrol()
reward_temp = -1.0
if ballRect.colliderect(playerRect):
speed[1] = -speed[1]
if ballRect.colliderect(comRect):
speed[1] = -speed[1]
if speed[0] < 0:
speed[0] -= acceleration
if speed[0] > 0:
speed[0] += acceleration
if speed[1] < 0:
speed[1] -= acceleration
if speed[1] > 0 :
speed[1] += acceleration
while True: #game
for event in pygame.event.get():
if event.type == pygame.QUIT:
pygame.quit()
sys.exit()
state = np.array([ballRect.center[0],ballRect.center[1],speed[0],speed[1],playerRect.center[0],playerRect.center[1],comRect.center[0],comRect.center[1]])
states.append(state)
action_prob = model.predict_on_batch(state.reshape(1,8))[0,:]
action_probs.append(action_prob)
action = np.random.choice(number_of_actions,p=action_prob)
if(action == 0): playermove = [0,0]
elif(action == 1): playermove = [5,0]
elif(action == 2): playermove = [-5,0]
playerRect = playerRect.move(playermove)
y = np.array([-1,-1,-1])
y[action] = 1
action_prob_grads.append(y-action_prob)
#enemy move
comRect = comRect.move(speedbar)
ballY = ballRect.left+5
comRectY = comRect.left+30
if comRect.top <= (height/1.5):
if comRectY - ballY > 0:
speedbar[0] = -7
elif comRectY - ballY < 0:
speedbar[0] = 7
if comRect.top > (height/1.5):
speedbar[0] = 0
if(mousematch == 1):
done = 0
reward_temp = 0.0
ballRect = ballRect.move(speed)
if ballRect.left < 0 or ballRect.right > width:
speed[0] = -speed[0]
if ballRect.top < 0:
pointsPlayer[0] += 1
done = pointcontrol()
reward_temp = 1.0
if ballRect.bottom > height:
pointsCom[0] += 1
done = pointcontrol()
reward_temp = -1.0
if ballRect.colliderect(playerRect):
speed[1] = -speed[1]
if ballRect.colliderect(comRect):
speed[1] = -speed[1]
if speed[0] < 0:
speed[0] -= acceleration
if speed[0] > 0:
speed[0] += acceleration
if speed[1] < 0:
speed[1] -= acceleration
if speed[1] > 0 :
speed[1] += acceleration
rewards.append(reward_temp)
if (done):
episode_number += 1
reward_sums.append(np.sum(rewards))
if len(reward_sums) > 40:
reward_sums.pop(0)
s = 'Episode %d Total Episode Reward: %f,Mean %f' % (
episode_number,np.sum(rewards),np.mean(reward_sums))
print(s)
filehandler.write(s + '\n')
filehandler.flush()
# Propagate the rewards back to actions where no reward
was given.
# Rewards for earlier actions are attenuated
rewards = np.vstack(rewards)
action_prob_grads = np.vstack(action_prob_grads)
rewards = reward_cal(rewards)
X = np.vstack(states).reshape(-1,8)
Y = action_probs + learning_rate * rewards * y
print('loss: ',model.train_on_batch(X,Y))
model.save_weights(filename)
states,[]
reward_sum = 0
screen.fill(black)
screen.blit(paddleP,playerRect)
screen.blit(ball,ballRect)
screen.blit(paddleC,comRect)
pygame.display.flip()
pygame.time.delay(delay_time)
这是我们的输出：
pygame 1.9.4 Hello from the pygame community. https://www.pygame.org/contribute.html Using TensorFlow backend.
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
dense_1 (Dense) (None,16) 144
_________________________________________________________________
dense_2 (Dense) (None,32) 544
_________________________________________________________________
dense_3 (Dense) (None,3) 99
=================================================================
Total params: 787 Trainable params: 787 Non-trainable params: 0
_________________________________________________________________ 2019-02-14 11:18:10.543401: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your cpu supports instructions that this TensorFlow binary was not compiled to use: AVX2 AVX512F FMA 2019-02-14 11:18:10.666634: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1432] Found device 0 with properties: name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.6705 pciBusID: 0000:17:00.0 totalMemory:
10.92GiB freeMemory: 10.76GiB 2019-02-14 11:18:10.775144: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1432] Found device 1 with properties: name: GeForce GTX 1080 Ti major: 6 minor: 1 memoryClockRate(GHz): 1.6705 pciBusID: 0000:65:00.0 totalMemory:
10.91GiB freeMemory: 10.73GiB 2019-02-14 11:18:10.776037: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1511] Adding visible gpu devices: 0,1 2019-02-14 11:18:11.176560: I tensorflow/core/common_runtime/gpu/gpu_device.cc:982] Device interconnect StreamExecutor with strength 1 edge matrix: 2019-02-14 11:18:11.176590: I tensorflow/core/common_runtime/gpu/gpu_device.cc:988] 0 1 2019-02-14 11:18:11.176596: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1001] 0: N Y 2019-02-14 11:18:11.176600: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1001] 1: Y N 2019-02-14 11:18:11.176914: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1115] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 10403 MB memory) -> physical GPU (device: 0,name: GeForce GTX 1080 Ti,pci bus id: 0000:17:00.0,compute capability: 6.1) 2019-02-14 11:18:11.177216: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1115] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:1 with 10382 MB memory) -> physical GPU (device: 1,pci bus id: 0000:65:00.0,compute capability: 6.1)
Computer Won 0 / 11 Episode 1 Total Episode Reward: -11.000000,Mean -11.000000
loss: 0.254405
Computer Won 0 / 11 Episode 2 Total Episode Reward: -11.000000,Mean -11.000000
loss: 0.254304
Computer Won 0 / 11 Episode 3 Total Episode Reward: -11.000000,Mean -11.000000
loss: 0.254304
Computer Won 0 / 11 Episode 4 Total Episode Reward: -11.000000,Mean -11.000000
loss: 0.254304
Computer Won 0 / 11 Episode 5 Total Episode Reward: -11.000000,Mean -11.000000
loss: 0.254304
Computer Won 0 / 11 Episode 6 Total Episode Reward: -11.000000,Mean -11.000000
loss: 0.254304
最佳答案
那是邪恶的“露露”,显示出它的力量.
Relu具有一个没有渐变的“零”区域.当所有输出都为负时,Relu使所有输出均等于零并消除反向传播.
安全使用Relus的最简单解决方案是在它们之前添加BatchNormalization层：
model = keras.models.Sequential()
model.add(Dense(16,kernel_initializer = 'glorot_normal'))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(Dense(32,kernel_initializer = 'glorot_normal'))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(Dense(number_of_actions,activation='softmax'))
这将使该层的“输出”一半“为零”,而另一半为可训练的.