-
Notifications
You must be signed in to change notification settings - Fork 11
/
q_table.py
155 lines (134 loc) · 6.46 KB
/
q_table.py
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
143
144
145
146
147
148
149
150
151
152
153
154
155
import numpy as np
import random
from collections import defaultdict
# 문제 구성 : {
# state : {x1, x2, x3, x4, x5, x6}
# action : {1, 0}
# reward : {1, -1}
# if prev_state_fp == 0,1,0,1,0 and currentAction == 0 {
# reward : 1000}
# }
# 개선 여지 : {
# epsilon을 1로 시작해서 점점 줄이면 더 성능이 좋아질 수 있을 것 같다.
# defalutdict 자료형을 사용하면 더 간단하게 만들 수 있을듯
# }
class QLearningAgent:
# 상태 변환 확률은 1이므로 생략
def __init__(self) :
self.actions = [0, 1]
self.learningLate = 0.01
self.discountFactor = 0.9
self.epsilon = 0.1
self.q_table = {"1": 0, "0": 0,
"00": 0, "01": 0, "10": 0, "11": 0,
"000": 0, "001": 0, "011": 0, "010": 0, "100": 0, "101": 0, "111": 0, "110": 0,
"0000": 0, "0001": 0, "0010": 0, "0011": 0, '0100': 0, '0101': 0, '0110': 0, "0111": 0,
'1000': 0, '1001': 0, '1010': 0, '1011': 0, '1100': 0, '1101': 0, '1110': 0, "1111": 0,
'00000': 0, '00001': 0, '00010': 0, '00011': 0, '00100': 0, '00101': 0, '00110': 0, '00111': 0,
'01000': 0, '01001': 0, '01010': 0, '01011': 0, '01100': 0, '01101': 0, '01110': 0, '01111': 0,
'10000': 0, '10001': 0, '10010': 0, '10011': 0, '10100': 0, '10101': 0, '10110': 0, '10111': 0,
'11000': 0, '11001': 0, '11010': 0, '11011': 0, '11100': 0, '11101': 0, '11110': 0, '11111': 0,
"000000": 0, "000001": 0, "000010": 0, "000011": 0, '000100': 0, '000101': 0, '000110': 0, '000111': 0,
"001000": 0, "001001": 0, "001010": 0, "001011": 0, '001100': 0, '001101': 0, '001110': 0, '001111': 0,
"010000": 0, "010001": 0, "010010": 0, "010011": 0, '010100': 0, '010101': 0, '010110': 0, '010111': 0,
"011000": 0, "011001": 0, "011010": 0, "011011": 0, '011100': 0, '011101': 0, '011110': 0, '011111': 0,
"100000": 0, "100001": 0, "100010": 0, "100011": 0, '100100': 0, '100101': 0, '100110': 0, '100111': 0,
"101000": 0, "101001": 0, "101010": 0, "101011": 0, '101100': 0, '101101': 0, '101110': 0, '101111': 0,
"110000": 0, "110001": 0, "110010": 0, "110011": 0, '110100': 0, '110101': 0, '110110': 0, '110111': 0,
'111000': 0, "111001": 0, "111010": 0, "111011": 0, "111100": 0, '111101': 0, '111110': 0, '111111': 0
}
# 최적경로를 찾기까지 epsilon을 1로 설정. 탐색력 최대화
def warmup(self):
wc = max(self.q_table.values())
if wc != 9999 :
self.epsilon = 1
else :
self.epsilon = 0.1
# s, a, r, s`를 이용해서 q-table 업데이트
def learn(self, state, action, reward, nextState) :
q_1 = self.q_table[state+str(action)]
q_2 = reward + self.discountFactor * self.q_table[nextState+self.argMax(nextState)]
self.q_table[state+str(action)] += self.learningLate * (q_2 - q_1)
# 마지막 state(6번째 갈림 길)일 때의 q-table 업데이트
def learnFinal(self, state, action, reward):
q_1 = reward
self.q_table[state+str(action)] = q_1
# e-greedy 정책에 따른 q-table내 해당 state의 action 반환
def get_action(self, state) :
if np.random.rand() < self.epsilon :
action = np.random.choice(self.actions)
# print("i'm greedy!")
else :
action = self.argMax(state)
return str(action)
# 최적의 action 반환
def argMax(self, state):
zeroValues = self.q_table[str(state)+'0']
oneValues = self.q_table[str(state)+'1']
if zeroValues > oneValues:
action = '0'
elif zeroValues < oneValues:
action = '1'
else:
action = str(np.random.choice(self.actions))
return str(action)
class CoProblem :
def __init__(self) :
self.fieldSize = 6 # 최대 행동 수
self.currentState = "" # 현재 상태
# 현재 state 반환
def getCurrentState(self):
self.currentState
return self.currentState
# 다음 state를 반환
def getNextState(self, action):
self.currentState += str(action)
return self.currentState
# 다음 state로 이동
def toNextState(self, aciton):
self.currentState += str(action)
# 현재 action에 대한 reward 반환
def getReward(self, action):
reward = 0
if self.currentState == "01010" and action == '0' :
reward += 9999
else :
if action == '0':
reward += -1
else:
reward += 1
return reward
# episode가 끝날 때마다 다시 environment 세팅
def setInit(self):
self.currentState = ""
if __name__ == "__main__" :
# Max Episode 설정
MAX_EPISODE = 5000
# environment와 agent 초기화
cop = CoProblem()
agent = QLearningAgent()
# 웜업을 통한 초기탐색
agent.warmup()
for episode in range(MAX_EPISODE) :
# episode가 시작할 때마다 environment 초기화
cop.setInit()
# 100 epi마다 q-table을 검사해서 warmup을 종료할지 결정
if episode % 100 == 0:
agent.warmup()
for stage in range(1,7):
# state 관측
state = cop.getCurrentState()
# state에 따른 agent의 action 선택
action = agent.get_action(state)
# action에 대한 reward 획득
reward = cop.getReward(action)
if len(cop.currentState) != cop.fieldSize-1:
# state와 action을 이용해서 nextState 관측
nextState = cop.getNextState(action)
# s, a, r, s`를 이용한 Q-table 학습
agent.learn(state, action, reward, nextState)
elif len(cop.currentState) == cop.fieldSize-1:
cop.toNextState(action)
# 마지막 단계일 때, s, a, r을 이용한 Q-table 학습
agent.learnFinal(state, action, reward)
print(episode, "episode's totoal state :", cop.currentState, "total rewards :", agent.q_table[cop.currentState])