-
Notifications
You must be signed in to change notification settings - Fork 0
/
RMax.py
286 lines (245 loc) · 9.82 KB
/
RMax.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
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
import random
import copy
#!!!observation format: (1, 0, 1, hp, (1, 1, 1), (x, y))
class RMax:
def __init__(self, epsilon, gamma, hordQ, probQ, punishment):
self.punishment = punishment #the punishment is Integer[0, inf)
self.epsilon = epsilon
self.gamma = gamma
self.hordq = hordQ
self.oriProbQ = probQ
self.probQ = {}
self.adjState = {}
self.adjRoom = [
[1, 2],
[0, 3],
[0, 3, 4],
[1, 2, 5],
[2, 5],
[3, 4]
]
self.Qmodel = {}
self.stepNum = 0
def touchAll(self, observation):
actionList = self.getActionList(self.getPlanVar(observation))
for action in actionList:
self.touch(observation, action)
def touch(self, observation, action):
ob = self.getPlanVar(observation)
key = (ob, action)
envVar = self.getEnvVar(observation)
if not envVar in self.Qmodel:
self.Qmodel[envVar] = {}
self.probQ[envVar] = copy.deepcopy(self.oriProbQ)
if not key in self.Qmodel[envVar]:
self.Qmodel[envVar][key] = 0 #assign 0 as the initial value
def getPlanVar(self, ob):
assert(len(ob) == 6)
return (ob[0], ob[1], ob[2], ob[5])
def getEnvVar(self, ob):
assert(len(ob) == 6)
return (ob[3], ob[4])
def mergeVar(self, planVar, envVar):
return (planVar[0], planVar[1], planVar[2], envVar[0], envVar[1], planVar[3])
def getLoc(self, ob):
assert(len(ob)==4) #observation format: (1, 1, 1, (x, y))
return ob[3]
#observation in reduced format
def getRoom(self, ob):
loc = self.getLoc(ob)
y = int(loc[1]/2)
x = int(loc[0]/3)
id = 2*y + x
return id
#observation in reduced format
def getActionList(self, observation):
room = self.getRoom(observation)
actionList = self.adjRoom[room]
return actionList
def selectAction(self, observation):
actionList = self.getActionList(self.getPlanVar(observation))
#use epsilon-greedy
if random.random() < self.epsilon:
#select randomly
action = actionList[int(random.random()*len(actionList))]
self.touch(observation, action)
return action
else:
#select the best action
v = []
for action in actionList:
self.touch(observation, action)
v.append(self.getQ(observation, action))
assert len(v) > 0
m = max(v)
select = int(random.random()*v.count(m))
i = 0
maxCount = 0
for value in v:
if value == m:
if maxCount == select:
action = actionList[i]
break
maxCount = maxCount + 1
i = i + 1
return action
def start(self, observation):
self.lastObservation = observation
curRoom = self.getRoom(self.getPlanVar(observation))
self.lastAction = nextRoom = self.selectAction(observation)
action = self.hordq.start(((curRoom, nextRoom), observation))
self.lastPrimitiveAction = action #debug only
return action
def getQ(self, inOb, action):
ob = self.getPlanVar(inOb)
envVar = self.getEnvVar(inOb)
key = (ob, action)
return self.Qmodel[envVar][key]
def getV(self, ob):
actionList = self.getActionList(self.getPlanVar(ob))
maxQ = self.getQ(ob, actionList[0])
for action in actionList:
Q = self.getQ(ob, action)
if Q > maxQ:
maxQ = Q
return maxQ
def dumpProb():
#status = [0, 0, 0]
print "1->2:", self.probQ.getQ(((0, 0, 0, (1, 1)), (0, 0, 0, (1, 2))), 2)
def updateProbModel(self, inOb, inLastOb, lastAction):
self.touch(inLastOb, lastAction)
#use the old value to update
envVar = self.getEnvVar(inLastOb)
lastOb = self.getPlanVar(inLastOb)
ob = self.getPlanVar(inOb)
#check if the environment variable exists or not
if not envVar in self.Qmodel:
self.Qmodel[envVar] = {}
self.probQ[envVar] = copy.deepcopy(self.oriProbQ)
#update probability model
probKey = (lastOb, ob)
self.probQ[envVar].touch(probKey, lastAction)
self.probQ[envVar].updateQ(probKey, lastAction, 1, 0) #gamma for probQ is not useful here
#add ob to lastOb's adjacency list
if not lastOb in self.adjState:
self.adjState[lastOb] = [ob]
else:
if not ob in self.adjState[lastOb]:
self.adjState[lastOb].append(ob)
for state in self.adjState[lastOb]:
if state != ob:
tmpProbKey = (lastOb, state)
self.probQ[envVar].touch(tmpProbKey, lastAction)
self.probQ[envVar].updateQ(tmpProbKey, lastAction, 0, 0) #gamma for probQ is not useful here
def updateQModel(self, inOb):
self.touchAll(inOb)
#use the old value to update
envVar = self.getEnvVar(inOb)
#use the new value to update
#update Q value
for i in range(0, 3):
for state in self.adjState:
actionList = self.getActionList(state)
room = self.getRoom(state)
fullState = self.mergeVar(state, envVar)
for action in actionList:
actionR = (room, action)
self.hordq.touchAll((actionR, fullState))
r = self.hordq.getVc((actionR, fullState))
c = 0
for adj in self.adjState[state]:
tmpKey = (state, adj)
self.probQ[envVar].touch(tmpKey, action)
prob = self.probQ[envVar].getQ(tmpKey, action)
fullAdjState = self.mergeVar(adj, envVar)
self.touchAll(fullAdjState)
v = self.getV(fullAdjState)
c = c + prob*v
#the Bellman's equation
self.Qmodel[envVar][(state, action)] = self.gamma*c + r
#observation is original format
def step(self, reward, observation):
lastRoom = self.getRoom(self.getPlanVar(self.lastObservation))
curRoom = self.getRoom(self.getPlanVar(observation))
#check for termination of subtask
if lastRoom == curRoom:
#continue execute last action
primitiveAction = self.hordq.step(reward, ((curRoom, self.lastAction), observation), reward)
else:
#update model
if self.epsilon != 1:
self.updateProbModel(observation, self.lastObservation, self.lastAction)
self.updateQModel(observation)
#choose the next action
action = self.selectAction(observation)
#room changed
if curRoom == self.lastAction:
#achieve the subgoal
internalReward = reward + self.punishment
#internalReward = reward
else:
internalReward = reward
#mission failed. punish the agent
#internalReward = reward - self.punishment
#debugging
#curLoc = self.getLoc(self.getPlanVar(observation))
#prevLoc = self.getLoc(self.getPlanVar(self.lastObservation))
#print "lastOb:", self.lastObservation
#print "ob:", observation
#print "move: ",prevLoc, "->", curLoc, " ", self.lastPrimitiveAction, " reward ", internalReward
#print "move: ",lastRoom, "->", curRoom, " ", self.lastAction, " next action:", action, " reward ", internalReward
primitiveAction = self.hordq.step(reward, ((curRoom, action), observation), internalReward)
self.lastObservation = observation
self.lastAction = action
self.lastPrimitiveAction = primitiveAction #debug only
self.stepNum = self.stepNum + 1
#if self.stepNum % 100000 == 0:
self.punishment = self.punishment - 5.0/100000.0
#if self.punishment < 0:
#self.punishment = 0
return primitiveAction
def end(self, reward):
#assume bus ends at (0, 0)
#update prob model
inLastOb = self.lastObservation
lastAction = self.lastAction
lastOb = self.getPlanVar(inLastOb)
envVar = self.getEnvVar(inLastOb)
if lastOb in self.adjState:
for state in self.adjState[lastOb]:
tmpProbKey = (lastOb, state)
self.probQ[envVar].touch(tmpProbKey, lastAction)
self.probQ[envVar].updateQ(tmpProbKey, lastAction, 0, 0) #gamma for probQ is not useful here
self.hordq.end(reward, reward)
#add comparison to random planner
import EmptySARSA
import HORDQ
import SARSA
if __name__ == "__main__":
alpha = 0.2
epsilon = 0.1
gamma = 0.9
#ob = (-1, -1, -1, (1, 2))
#ob2 = (-1, -1, -1, (5, 4))
#ob3 = (-1, -1, -1, (0, 0))
ob4 = (-1, -1, -1, (1, 1))
punishment = 10
isRORDQ = False
hordQ = HORDQ.HORDQ(alpha, epsilon, gamma, [1, -1], isRORDQ)
probQ = SARSA.SARSA(alpha, epsilon, gamma, [1, -1])
controller = RMax(epsilon, gamma, hordQ, probQ, punishment)
#unit test for get room
val = controller.getRoom(ob4)
print "value: ", val
assert( val == 0)
#controller.start(ob)
#for i in range(0, 1000):
#
#controller.step(1, ob)
#controller.step(1, ob)
#controller.step(1, ob)
#controller.step(1, ob2)
#controller.end(10)
#print controller.Q
#print hordQ.Qc
#print probQ.Q