1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
|
#!/usr/bin/env python
import sys,random,os
import numpy as np
# Import snake game
from snake import Snake
class QTable:
"""
# Boolean features:
# Snake go up?
# Snake go right?
# Snake go down?
# Snake go left?
# Apple at up?
# Apple at right?
# Apple at down?
# Apple at left?
# Obstacle at up?
# Obstacle at right?
# Obstacle at down?
# Obstacle at left?
# Tail in front?
##### Totally 13 boolean features so 2^13=8192 states
##### Totally 4 actions for the AI (up, right,down,left)
##### Totally 4*2^13 thus 32768 table entries
##### Reward +1 when eat an apple
##### Reward -10 when hit obstacle
"""
def __init__(self, file, save_every=10):
self.file=file
self.save_every=save_every
self.update_counter=0
if os.path.exists(file):
self.qtable=np.loadtxt(file)
else:
self.qtable=np.zeros((2**13, 4))
def isWall(self,h,game):
if h[0]<0 or h[1]<0 or h[0] >= game.grid_width or h[1] >= game.grid_height:
return(True)
return(False)
def get_state(self,game):
# First compute usefull values
h=game.snake[0]
left=(h[0]-1,h[1])
right=(h[0]+1,h[1])
up=(h[0],h[1]-1)
down=(h[0],h[1]+1)
a=game.apple
snake_go_up=(game.direction==12)
snake_go_right=(game.direction==3)
snake_go_down=(game.direction==6)
snake_go_left=(game.direction==9)
apple_up=(a[1]<h[1])
apple_right=(a[0]>h[0])
apple_down=(a[1]>h[1])
apple_left=(a[0]<h[0])
obstacle_up=(up in game.snake or self.isWall(up, game))
obstacle_right=(right in game.snake or self.isWall(right, game))
obstacle_down=(down in game.snake or self.isWall(down, game))
obstacle_left=(left in game.snake or self.isWall(left, game))
tail_in_front=0
if game.direction == 3:
for x in range(h[0],game.grid_width):
if (x,h[1]) in game.snake[1:]:
tail_in_front=1
break
elif game.direction == 9:
for x in range(0,h[0]):
if (x,h[1]) in game.snake[1:]:
tail_in_front=1
break
elif game.direction == 12:
for y in range(0,h[1]):
if (h[0],y) in game.snake[1:]:
tail_in_front=1
break
elif game.direction == 6:
for y in range(h[1],game.grid_height):
if (h[0],y) in game.snake[1:]:
tail_in_front=1
break
# This come from me I do not now if it is the best way to identify a state
state=2**12*tail_in_front+2**11*snake_go_up+2**10*snake_go_right+2**9*snake_go_down+2**8*snake_go_left+2**7*apple_up+2**6*apple_right+2**5*apple_down+2**4*apple_left+2**3*obstacle_up+2**2*obstacle_right+2**1*obstacle_down+obstacle_left
return(state)
def apply_bellman(self,state,action,new_state,reward):
alpha=0.5
gamma=0.9
self.qtable[state,action]=self.qtable[state,action]+alpha*(reward+gamma*np.max(self.qtable[new_state])-self.qtable[state,action])
self.update_counter+=1
if self.update_counter>=self.save_every:
np.savetxt(self.file,self.qtable)
self.update_counter=0
def get_action(self,state):
# Choose an action
action=random.choice((0,1,2,3))
if np.max(self.qtable[state]) > 0:
#qactions=qtable[state]
#options=np.flatnonzero(qactions == np.max(qactions)) # Since Q value might be equals for several actions
#action = random.choice(options)
action=np.argmax(self.qtable[state])
return(action)
# Perform learning
perf=0
last_state=None
last_action=None
game=Snake(length=4,fps=300,startat=(10,10))
qtable=QTable("qtable.txt")
for i in range(0,10000):
result=0
while result >= 0:
state=qtable.get_state(game)
action=qtable.get_action(state)
result=game.play3(action)
if last_state!=None:
reward=0
if result==-1:
reward=-10
elif result==1:
reward=1
qtable.apply_bellman(last_state,last_action,state,reward)
last_state=state
last_action=action
# Measurements
score=game.last_score
perf=max(perf,score)
print("Game ended with "+str(score)+" best so far is "+str(perf))
|
