-
Notifications
You must be signed in to change notification settings - Fork 3
/
agents.py
356 lines (301 loc) · 12.4 KB
/
agents.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
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
from captureAgents import CaptureAgent
from capture import SIGHT_RANGE
import random, time, util
from game import Directions
import game
class ApproximateAdversarialAgent(CaptureAgent):
"""
Superclass for agents choosing actions via alpha-beta search, with
positions of unseen enemies approximated by Bayesian inference
"""
#####################
# AI algorithm code #
#####################
SEARCH_DEPTH = 5
def registerInitialState(self, gameState):
CaptureAgent.registerInitialState(self, gameState)
# Get all non-wall positions on the board
self.legalPositions = gameState.data.layout.walls.asList(False)
# Initialize position belief distributions for opponents
self.positionBeliefs = {}
for opponent in self.getOpponents(gameState):
self.initializeBeliefs(opponent)
def initializeBeliefs(self, agent):
"""
Uniformly initialize belief distributions for opponent positions.
"""
self.positionBeliefs[agent] = util.Counter()
for p in self.legalPositions:
self.positionBeliefs[agent][p] = 1.0
def chooseAction(self, gameState):
# Update belief distribution about opponent positions and place hidden
# opponents in their most likely positions
myPosition = gameState.getAgentState(self.index).getPosition()
noisyDistances = gameState.getAgentDistances()
probableState = gameState.deepCopy()
for opponent in self.getOpponents(gameState):
pos = gameState.getAgentPosition(opponent)
if pos:
self.fixPosition(opponent, pos)
else:
self.elapseTime(opponent, gameState)
self.observe(opponent, noisyDistances[opponent], gameState)
self.displayDistributionsOverPositions(self.positionBeliefs.values())
for opponent in self.getOpponents(gameState):
probablePosition = self.guessPosition(opponent)
conf = game.Configuration(probablePosition, Directions.STOP)
probableState.data.agentStates[opponent] = game.AgentState(
conf, probableState.isRed(probablePosition) != probableState.isOnRedTeam(opponent))
# Run negamax alpha-beta search to pick an optimal move
bestVal, bestAction = float("-inf"), None
for opponent in self.getOpponents(gameState):
value, action = self.expectinegamax(opponent,
probableState,
self.SEARCH_DEPTH,
1,
retAction=True)
if value > bestVal:
bestVal, bestAction = value, action
return action
def fixPosition(self, agent, position):
"""
Fix the position of an opponent in an agent's belief distributions.
"""
updatedBeliefs = util.Counter()
updatedBeliefs[position] = 1.0
self.positionBeliefs[agent] = updatedBeliefs
def elapseTime(self, agent, gameState):
"""
Elapse belief distributions for an agent's position by one time step.
Assume opponents move randomly, but also check for any food lost from
the previous turn.
"""
updatedBeliefs = util.Counter()
for (oldX, oldY), oldProbability in self.positionBeliefs[agent].items():
newDist = util.Counter()
for p in [(oldX - 1, oldY), (oldX + 1, oldY),
(oldX, oldY - 1), (oldX, oldY + 1)]:
if p in self.legalPositions:
newDist[p] = 1.0
newDist.normalize()
for newPosition, newProbability in newDist.items():
updatedBeliefs[newPosition] += newProbability * oldProbability
lastObserved = self.getPreviousObservation()
if lastObserved:
lostFood = [food for food in self.getFoodYouAreDefending(lastObserved).asList()
if food not in self.getFoodYouAreDefending(gameState).asList()]
for f in lostFood:
updatedBeliefs[f] = 1.0/len(self.getOpponents(gameState))
self.positionBeliefs[agent] = updatedBeliefs
def observe(self, agent, noisyDistance, gameState):
"""
Update belief distributions for an agent's position based upon
a noisy distance measurement for that agent.
"""
myPosition = self.getAgentPosition(self.index, gameState)
teammatePositions = [self.getAgentPosition(teammate, gameState)
for teammate in self.getTeam(gameState)]
updatedBeliefs = util.Counter()
for p in self.legalPositions:
if any([util.manhattanDistance(teammatePos, p) <= SIGHT_RANGE
for teammatePos in teammatePositions]):
updatedBeliefs[p] = 0.0
else:
trueDistance = util.manhattanDistance(myPosition, p)
positionProbability = gameState.getDistanceProb(trueDistance, noisyDistance)
updatedBeliefs[p] = positionProbability * self.positionBeliefs[agent][p]
if not updatedBeliefs.totalCount():
self.initializeBeliefs(agent)
else:
updatedBeliefs.normalize()
self.positionBeliefs[agent] = updatedBeliefs
def guessPosition(self, agent):
"""
Return the most likely position of the given agent in the game.
"""
return self.positionBeliefs[agent].argMax()
def expectinegamax(self, opponent, state, depth, sign, retAction=False):
"""
Negamax variation of expectimax.
"""
if sign == 1:
agent = self.index
else:
agent = opponent
bestAction = None
if self.stateIsTerminal(agent, state) or depth == 0:
bestVal = sign * self.evaluateState(state)
else:
actions = state.getLegalActions(agent)
actions.remove(Directions.STOP)
bestVal = float("-inf") if agent == self.index else 0
for action in actions:
successor = state.generateSuccessor(agent, action)
value = -self.expectinegamax(opponent, successor, depth - 1, -sign)
if agent == self.index and value > bestVal:
bestVal, bestAction = value, action
elif agent == opponent:
bestVal += value/len(actions)
if agent == self.index and retAction:
return bestVal, bestAction
else:
return bestVal
def stateIsTerminal(self, agent, gameState):
"""
Check if the search tree should stop expanding at the given game state
on the given agent's turn.
"""
return len(gameState.getLegalActions(agent)) == 0
def evaluateState(self, gameState):
"""
Evaluate the utility of a game state.
"""
util.raiseNotDefined()
#####################
# Utility functions #
#####################
def getAgentPosition(self, agent, gameState):
"""
Return the position of the given agent.
"""
pos = gameState.getAgentPosition(agent)
if pos:
return pos
else:
return self.guessPosition(agent)
def agentIsPacman(self, agent, gameState):
"""
Check if the given agent is operating as a Pacman in its current position.
"""
agentPos = self.getAgentPosition(agent, gameState)
return (gameState.isRed(agentPos) != gameState.isOnRedTeam(agent))
def getOpponentDistances(self, gameState):
"""
Return the IDs of and distances to opponents, relative to this agent.
"""
return [(o, self.distancer.getDistance(
self.getAgentPosition(self.index, gameState),
self.getAgentPosition(o, gameState)))
for o in self.getOpponents(gameState)]
class CautiousAttackAgent(ApproximateAdversarialAgent):
"""
An attack-oriented agent that will retreat back to its home zone
after consuming 5 pellets.
"""
def registerInitialState(self, gameState):
ApproximateAdversarialAgent.registerInitialState(self, gameState)
self.retreating = False
def chooseAction(self, gameState):
if (gameState.getAgentState(self.index).numCarrying < 5 and
len(self.getFood(gameState).asList())):
self.retreating = False
else:
self.retreating = True
return ApproximateAdversarialAgent.chooseAction(self, gameState)
def evaluateState(self, gameState):
myPosition = self.getAgentPosition(self.index, gameState)
targetFood = self.getFood(gameState).asList()
distanceFromStart = abs(myPosition[0] - gameState.getInitialAgentPosition(self.index)[0])
opponentDistances = self.getOpponentDistances(gameState)
opponentDistance = min([dist for id, dist in opponentDistances])
if self.retreating:
return - len(targetFood) \
- 2 * distanceFromStart \
+ opponentDistance
else:
foodDistances = [self.distancer.getDistance(myPosition, food)
for food in targetFood]
minDistance = min(foodDistances) if len(foodDistances) else 0
return 2 * self.getScore(gameState) \
- 100 * len(targetFood) \
- 3 * minDistance \
+ 2 * distanceFromStart \
+ opponentDistance
class OpportunisticAttackAgent(ApproximateAdversarialAgent):
def evaluateState(self, gameState):
myPosition = self.getAgentPosition(self.index, gameState)
food = self.getFood(gameState).asList()
targetFood = None
maxDist = 0
opponentDistances = self.getOpponentDistances(gameState)
opponentDistance = min([dist for id, dist in opponentDistances])
if not food or gameState.getAgentState(self.index).numCarrying > self.getScore(gameState) > 0:
return 20 * self.getScore(gameState) \
- self.distancer.getDistance(myPosition, gameState.getInitialAgentPosition(self.index)) \
+ opponentDistance
for f in food:
d = min([self.distancer.getDistance(self.getAgentPosition(o, gameState), f)
for o in self.getOpponents(gameState)])
if d > maxDist:
targetFood = f
maxDist = d
if targetFood:
foodDist = self.distancer.getDistance(myPosition, targetFood)
else:
foodDist = 0
distanceFromStart = abs(myPosition[0] - gameState.getInitialAgentPosition(self.index)[0])
if not len(food):
distanceFromStart *= -1
return 2 * self.getScore(gameState) \
- 100 * len(food) \
- 2 * foodDist \
+ opponentDistance \
+ distanceFromStart
class DefensiveAgent(ApproximateAdversarialAgent):
"""
A defense-oriented agent that should never cross into the opponent's territory.
"""
TERMINAL_STATE_VALUE = -1000000
def stateIsTerminal(self, agent, gameState):
return self.agentIsPacman(self.index, gameState) or \
ApproximateAdversarialAgent.stateIsTerminal(self, agent, gameState)
class GoalieAgent(DefensiveAgent):
"""
A defense-oriented agent that tries to place itself between its team's
food and the closest opponent.
"""
def evaluateState(self, gameState):
if self.agentIsPacman(self.index, gameState):
return DefensiveAgent.TERMINAL_STATE_VALUE
myPosition = self.getAgentPosition(self.index, gameState)
shieldedFood = self.getFoodYouAreDefending(gameState).asList()
opponentPositions = [self.getAgentPosition(opponent, gameState)
for opponent in self.getOpponents(gameState)]
if len(shieldedFood):
opponentDistances = util.Counter()
opponentTotalDistances = util.Counter()
for f in shieldedFood:
for o in opponentPositions:
distance = self.distancer.getDistance(f, o)
opponentDistances[(f, o)] = distance
opponentTotalDistances[o] -= distance
threateningOpponent = opponentTotalDistances.argMax()
atRiskFood, shortestDist = None, float("inf")
for (food, opponent), dist in opponentDistances.iteritems():
if opponent == threateningOpponent and dist < shortestDist:
atRiskFood, shortestDist = food, dist
return len(shieldedFood) \
- 2 * self.distancer.getDistance(myPosition, atRiskFood) \
- self.distancer.getDistance(myPosition, threateningOpponent)
else:
return -min(self.getOpponentDistances(gameState), key=lambda t: t[1])[1]
class HunterDefenseAgent(DefensiveAgent):
"""
A defense-oriented agent that actively seeks out an enemy agent in its territory
and tries to hunt it down
"""
def evaluateState(self, gameState):
myPosition = self.getAgentPosition(self.index, gameState)
if self.agentIsPacman(self.index, gameState):
return DefensiveAgent.TERMINAL_STATE_VALUE
score = 0
pacmanState = [self.agentIsPacman(opponent, gameState)
for opponent in self.getOpponents(gameState)]
opponentDistances = self.getOpponentDistances(gameState)
for isPacman, (id, distance) in zip(pacmanState, opponentDistances):
if isPacman:
score -= 100000
score -= 5 * distance
elif not any(pacmanState):
score -= distance
return score