def partiallyObservableAction(self,gameState):
global possibleGhostStates
global frontierStates
pacmanLoc = gameState.getPacmanPosition()
ghostLoc = gameState.getGhostPosition(1)
room = gameState.getLoctoRoom()
if room[int(ghostLoc[0])][int(ghostLoc[1])] in gameState.data.roomsOn:
possibleGhostStates.clear()
possibleGhostStates[ghostLoc] = 1.0
frontierStates = [ghostLoc]
if (pacmanLoc, ghostLoc) in MDPUtil.AstarPolicy.keys():
action = self.convertTuple(MDPUtil.AstarPolicy[(pacmanLoc,ghostLoc)])
else:
action = self.A_star(pacmanLoc,ghostLoc,self.heuristic,gameState)
MDPUtil.AstarPolicy[(pacmanLoc,ghostLoc)] = self.convertAction(action)
else:
predictLoc = (-1,-1)
while True:
predictLoc = util.chooseFromDistribution(possibleGhostStates)
if len(possibleGhostStates.keys())==1:
predictLoc = possibleGhostStates.keys()[0]
break
while (room[int(predictLoc[0])][int(predictLoc[1])] in gameState.data.roomsOn) and len(possibleGhostStates.keys())!=1:
del possibleGhostStates[predictLoc]
predictLoc = util.chooseFromDistribution(possibleGhostStates)
if len(possibleGhostStates.keys())==1:
predictLoc = possibleGhostStates.keys()[0]
break
if predictLoc != pacmanLoc: break
tempFrontierStates = Set()
for frontierLoc in frontierStates:
tempNode = self.node(int(frontierLoc[0]),int(frontierLoc[1]))
possibleActions = self.getActions(gameState,tempNode)
for frontierAction in possibleActions:
actionTuple = self.convertAction(frontierAction)
newFrontierState = (frontierLoc[0]+actionTuple[0],frontierLoc[1]+actionTuple[1])
if newFrontierState in possibleGhostStates.keys(): continue
if room[int(newFrontierState[0])][int(newFrontierState[1])] in gameState.data.roomsOn: continue
tempFrontierStates.add(newFrontierState)
oldProb = possibleGhostStates[frontierLoc]
possibleGhostStates[newFrontierState] += oldProb/float(len(possibleActions))
possibleGhostStates.normalize()
frontierStates = tempFrontierStates
if (pacmanLoc, predictLoc) in MDPUtil.AstarPolicy.keys():
action = self.convertTuple(MDPUtil.AstarPolicy[(pacmanLoc,predictLoc)])
else:
action = self.A_star(pacmanLoc,predictLoc,self.heuristic,gameState)
MDPUtil.AstarPolicy[(pacmanLoc,predictLoc)] = self.convertAction(action)
return action
def get_generation(self):
#if self.id == 'i0':
timePeriod = self.environment.get_time()
if timePeriod != self.last_time:
tran = self.transitions[self.last_time]
disto = {}
for s in tran:
if s['from'] == self.last_generation:
for ns in s['to']:
disto[ns['to']] = ns['prob']
break
try:
ss = self.last_generation
self.last_generation = util.chooseFromDistribution(disto)
except Exception as e:
#if self.id == 'i0':
print 'vvvvvvvvvvvvvvvvvvvvvvvvvvvv'
print self.last_time, ss, self.last_generation, timePeriod, disto
print '****************************'
raise e
self.last_time = timePeriod
res = self.last_generation - 18
if res < 0:
res = 0
return res
def getDirectionalExpectimaxValue(self, gameState, agentIndex, depth):
if (agentIndex == 0
and depth == 1) or gameState.isWin() or gameState.isLose():
return self.evaluationFunction(gameState)
legalMoves = gameState.getLegalActions(agentIndex)
if agentIndex == 0:
return max(
self.getDirectionalExpectimaxValue(
state, getNextIndexAgent(agentIndex, gameState), depth - 1)
for state in [
gameState.generatePacmanSuccessor(action)
for action in legalMoves
])
else:
ghost = DirectionalGhost(index=agentIndex)
act_prob_dict = ghost.getDistribution(gameState)
val_prob_dict = util.Counter()
for action in legalMoves:
state = gameState.generateSuccessor(agentIndex, action)
val = self.getDirectionalExpectimaxValue(
state, getNextIndexAgent(agentIndex, gameState), depth)
val_prob_dict[val] = act_prob_dict[action]
val_prob_dict.normalize()
return util.chooseFromDistribution(val_prob_dict)
def getAction(self, state):
"""
Get the action the ghost will do based on a distribution
of possible actions.
"""
dist = self.getDistribution(state)
return chooseFromDistribution(dist)
def getAction( self, state ):
dist = self.getDistribution(state)
# return state.getLegalActions(self.index)[0]
if len(dist) == 0:
return Directions.STOP
else:
return util.chooseFromDistribution( dist )
def getAction( self, state ):
dist = self.getDistribution(state)
if 'Stop' in dist.keys() and len(dist) > 1:
del dist['Stop']
if len(dist) == 0:
return Directions.STOP
else:
return util.chooseFromDistribution( dist )
def getThreeAction( self, state, list_of_ghost_actions):
print evaluateHeuristic(state)
dist = self.getDistribution(state, list_of_ghost_actions)
if len(dist) == 0:
return Directions.STOP
else:
act = util.chooseFromDistribution( dist )
#print self.index, act
return act
def getAction(self, state):
if (state.isWin() or state.isLose()): #whoami
return Directions.STOP
dist = self.getDistribution(state)
if len(dist) == 0:
return Directions.STOP
else:
return util.chooseFromDistribution(dist)
def sample_data(self, trainingData, trainingLabels, sample_weights):
td = util.Counter()
tl = util.Counter()
for i in range(len(trainingLabels)):
k = util.chooseFromDistribution(sample_weights)
td[i] = trainingData[k]
tl[i] = trainingLabels[k]
return td, tl
"*** YOUR CODE HERE ***"
def getPartiallyObservableLoc(self,state):
global possiblePacmanStates
global frontierStates
pacmanLoc = state.getPacmanPosition()
ghostLoc = state.getGhostPosition(self.index)
room = state.getLoctoRoom()
if room[int(pacmanLoc[0])][int(pacmanLoc[1])] in state.data.roomsOn:
possiblePacmanStates.clear()
possiblePacmanStates[pacmanLoc] = 1.0
frontierStates = [pacmanLoc]
return pacmanLoc
else:
predictLoc = util.chooseFromDistribution(possiblePacmanStates)
while (room[int(predictLoc[0])][int(predictLoc[1])] in state.data.roomsOn) and len(possiblePacmanStates.keys())!=1:
del possiblePacmanStates[predictLoc]
predictLoc = util.chooseFromDistribution(possiblePacmanStates)
if len(possiblePacmanStates.keys())==1:
predictLoc = possiblePacmanStates.keys()[0]
break
tempFrontierStates = Set()
for frontierLoc in frontierStates:
tempNode = self.node(int(frontierLoc[0]),int(frontierLoc[1]))
possibleActions = self.getActions(state,tempNode)
for frontierAction in possibleActions:
actionTuple = self.convertAction(frontierAction)
newFrontierState = (frontierLoc[0]+actionTuple[0],frontierLoc[1]+actionTuple[1])
if room[int(newFrontierState[0])][int(newFrontierState[1])] in state.data.roomsOn: continue
tempFrontierStates.add(newFrontierState)
oldProb = possiblePacmanStates[frontierLoc]
possiblePacmanStates[newFrontierState] += oldProb/float(len(possibleActions))
possiblePacmanStates.normalize()
frontierStates = tempFrontierStates
pacmanLoc = predictLoc
return pacmanLoc
def getAction(self, gameState):
if not self.xl:
return max(gameState.getLegalActions(self.index), key = lambda action: (1 - 2 * self.index) * self.rawStateLib[gameState.generateSuccessor(self.index, action)])
# 增加随机性
sumprob = 0
dist = util.Counter()
for action in gameState.getLegalActions(self.index):
newState = gameState.generateSuccessor(self.index, action)
#sumprob += self.k ** self.rawStateLib[newState][self.index]
dist[action] = (self.k ** ((1 - 2 * self.index) * self.rawStateLib[newState]))
dist.normalize()
return util.chooseFromDistribution( dist )
def make_decision(self):
if self.dcopPhase == 0:
if not self.decision_made:
if self.last_state is not None:
probs = {}
for ns in self.next_states(self.last_state):
probs[ns] = self.probabilities[(self.last_state, ns)]
if len(probs) > 0:
self.last_state = chooseFromDistribution(probs)
self.commit_generators(
self.generatorsValues[self.last_state])
self.commit_children_powerLines(
self.powerLineValues[self.last_state])
self.decision_made = True
for n in self.relayNode.neighbours:
self.send(n, {'type': 'action-taken'})
else:
# Asking children for solving DCOP
self.printer('%d %s miss Unknown' %
(self.environment.get_time(), self.name))
self.test_log.append(2)
for c in self.relayNode.neighbours:
self.send(c, {'type': 'request-for-dcop'})
self.dcopPhase = 1
else:
raise 'last state is None'
elif self.all_neighbours_took:
# Checking for good decision
powerLines = self.relayNode.get_powerLine_values()
powerLineValues = tuple([powerLines[pl] for pl in powerLines])
# check for goodness
if self.last_state[1] == powerLineValues:
self.done = True
else:
# Asking children for solving DCOP
self.printer('%d %s miss Prediction' %
(self.environment.get_time(), self.name))
self.test_log.append(1)
for c in self.relayNode.neighbours:
self.send(c, {'type': 'request-for-dcop'})
self.dcopPhase = 1
elif self.dcopPhase == 3:
self.dcopPhase = 0
self.decision_made = True
self.done = True
def predictGhostMove(myPos, ghostPos): #TODO: NOTE: This actually doesn't work since it doesn't check validity of a move for the ghost
'''
Returns the expected position of the ghost
:param myPos:
:param ghostPos:
:return:
'''
moves = util.Counter()
#r = random.random() #where r is the degree of noise (randomness)
#if optimal:
if myPos[1] < ghostPos[1]:
moves[Directions.SOUTH] = 1
if myPos[0] < ghostPos[0]:
moves[Directions.SOUTH] = .8
moves[Directions.WEST] = .2
elif myPos[0] > ghostPos[0]:
moves[Directions.SOUTH] = .8
moves[Directions.EAST] = .2
elif myPos[1] > ghostPos[1]:
moves[Directions.NORTH] = 1
if myPos[0] < ghostPos[0]:
moves[Directions.NORTH] = .8
moves[Directions.WEST] = .2
elif myPos[0] > ghostPos[0]:
moves[Directions.NORTH] = .8
moves[Directions.EAST] = .2
elif myPos[1] == ghostPos[1]:
if myPos[0] < ghostPos[0]:
moves[Directions.WEST] = .9
moves[Directions.STOP] = .1
elif myPos[0] > ghostPos[0]:
moves[Directions.EAST] = 1
moves[Directions.STOP] = .1
else:
moves[Directions.STOP] = 1
#currently there is no check if ghost will move off the map or into a wall
## should just pick optimally for closer towards pac
## ghost might be using euc or man instead of maze distance though!!!
move = util.chooseFromDistribution(moves)
vector = Actions.directionToVector(move)
newPos = (ghostPos[0] + vector[0], ghostPos[1] + vector[1])
return newPos
def getAction(self, state):
"""
Compute the action to take in the current state. With
probability self.epsilon, we should take a random action and
take the best policy action otherwise. Note that if there are
no legal actions, which is the case at the terminal state, you
should choose None as the action.
HINT: You might want to use util.flipCoin(prob)
HINT: To pick randomly from a list, use random.choice(list)
"""
# Pick Action
legalActions = self.getLegalActions(state)
action = None
totalDeno = 0
probability_list = []
"*** YOUR CODE HERE ***"
if (len(legalActions) == 0):
return None
if (util.flipCoin(self.epsilon)):
for action in legalActions:
currQValue = self.getQValue(state, action)
currProb = math.exp(self.epsilon * currQValue)
totalDeno = totalDeno + currProb
for action in legalActions:
currQValue = self.getQValue(state, action)
currProb = math.exp(self.epsilon * currQValue)
prob = (currProb / totalDeno)
probability_list.append((prob, action))
action = util.chooseFromDistribution(probability_list)
else:
action = self.computeActionFromQValues(state)
return action
def assignJointActions(self, state, depth=4):
startTime = time.clock()
pacmanPosition = state.getPacmanPosition()
#pos = state.getGhostPosition( self.index )
allGhostPositions = state.getGhostPositions()
numGhosts = len(allGhostPositions)
jointActions = self.get_all_joint_actions(numGhosts, state)
#print jointActions
bestJointAction = None
bestJointActionValue = float("-inf")
for jointAction in jointActions:
value = evaluate_joint_action(jointAction, state, depth)
if value > bestJointActionValue:
bestJointActionValue = value
bestJointActions = []
bestJointActions.append(jointAction)
elif value == bestJointActionValue:
bestJointActions.append(jointAction)
# if value(jointAction) > bestJointActionValue:
# bestJointActionValue = value(jointAction)
# bestJointAction = jointAction
# elif value(jointAction) == bestJointActionValue:
# compare
bestProb = 0.95
distribution = util.Counter()
for a in bestJointActions: distribution[tuple(a)] = bestProb / len(bestJointActions)
for a in jointActions: distribution[tuple(a)] += ( 1-bestProb ) / len(jointActions)
# jointAction = {}
# for i in xrange(1, numGhosts+1):
# jointAction[i] = 'Stop'
#print "bestJointAction is : ", bestJointAction
print evaluateHeuristic(state)
return list(util.chooseFromDistribution( distribution ))
def getRandomSuccessor(self, gameState, agentIndex, currentDepth):
# Randomly uniform selection of an action
dist = self.getDistribution(agentIndex, gameState)
selectedAction = util.chooseFromDistribution(dist)
legal = gameState.getLegalActions(agentIndex)
if Directions.STOP in legal: legal.remove(Directions.STOP)
# NB : we browse this loop to expand all the gameState but it's not optimal
# In practise we can just expand call expectiMax on the action selected previously
for action in legal:
successor = gameState.generateSuccessor(agentIndex, action)
nextDepth = (currentDepth +
1) if (agentIndex == self.numGhosts) else currentDepth
nextAgent = (agentIndex + 1) % (self.numGhosts + 1)
(score, oldAction) = self.expectiMax(successor, nextAgent,
nextDepth)
# Check if the action is the one who was randomly selected
if action == selectedAction:
selectedScore = score
return (selectedScore, selectedAction)
def getAction(self, gameState):
leftnum = len(gameState.getLeftPiles(self.index))
legalaction = gameState.getLegalActions(self.index)
sumprob = 0
max = -99999
dist = util.Counter()
bestaction = legalaction[0]
for action in legalaction:
newgameState = gameState.generateSuccessor(self.index, action)
#print str(newgameState.getBoard()) in self.gameStateValue[leftnum - 1].keys()
if str(newgameState.getBoard()) in self.gameStateValue[self.index][leftnum - 1].keys():
'''
if self.gameStateValue[self.index][leftnum - 1][str(newgameState.getBoard())] > max:
bestaction = action
max = self.gameStateValue[self.index][leftnum - 1][str(newgameState.getBoard())]
'''
sumprob += (self.k ** self.gameStateValue[self.index][leftnum - 1][str(newgameState.getBoard())])
else:
'''
if self.evalFunc(newgameState, self.index) * 1.9 > max:
bestaction = action
max = self.evalFunc(newgameState, self.index) * 1.9
'''
sumprob += (self.k ** evalFunc(newgameState, self.index))
for action in legalaction:
newgameState = gameState.generateSuccessor(self.index, action)
if str(newgameState.getBoard()) in self.gameStateValue[self.index][leftnum - 1].keys():
dist[action] = float(self.k ** self.gameStateValue[self.index][leftnum - 1][str(newgameState.getBoard())]) / float(sumprob)
else:
dist[action] = float(self.k ** evalFunc(newgameState, self.index)) / float(sumprob)
dist.normalize()
return util.chooseFromDistribution( dist )
return bestaction
def getAction(self, state):
dist = self.getDistribution(state)
if len(dist) == 0:
return Directions.STOP
else:
return util.chooseFromDistribution(dist)
def getAction( self, state ):
dist = self.getDistribution(state)
if len(dist) == 0:
return Directions.STOP
else:
return util.chooseFromDistribution( dist )
def getAction( self, state ): dist = self.getDistribution( state ) if len(dist) == 0: return Actions.STOP return chooseFromDistribution( dist )
def pickAction(self, state): "Returns the action according to probability distribution." return util.chooseFromDistribution()
def getAction(self, state):
dist = self.getDistribution(state)
return chooseFromDistribution(dist)
def getAction(self, state):
dist = self.getDistribution(state)
return chooseFromDistribution(dist)
def getAction(self, state, total_pacmen, agentIndex):
dist = self.getDistribution(state, total_pacmen)
if len(dist) == 0:
return Directions.STOP
else:
return util.chooseFromDistribution(dist)
for i in range(21, 52): smallDice.append(i)
mediumDice = range(51, 256)
largeDice = range(256, 10000)
roundTenDice = range(30, 10000, 10)
roundHundredDice = range(100, 10000, 100)
score = 0
results = []
for attempt in range(numGames):
diceDist = util.Counter()
diceDist[1] = 5
for number in range(1, 11): diceDist[number] = 10*(1/float(len(range(2,10))))
for number in range(11, 101): diceDist[number] = 3*(1/float(len(range(11,100))))
for number in range(101, 1001): diceDist[number] = 1*(1/float(len(range(101,1000))))
diceDist.normalize()
numDice = util.chooseFromDistribution( diceDist )
sizeDist = util.Counter()
for number in smallDice: sizeDist[number] = 150*(1/float(len(smallDice)))
for number in mediumDice: sizeDist[number] = 40*(1/float(len(mediumDice)))
for number in largeDice: sizeDist[number] = 10*(1/float(len(largeDice)))
for number in dNDDice: sizeDist[number] += 350*(1/float(len(dNDDice)))
for number in roundTenDice: sizeDist[number] += 40*(1/float(len(roundTenDice)))
for number in roundHundredDice: sizeDist[number] += 60*(1/float(len(roundTenDice)))
sizeDist.normalize()
diceSize = util.chooseFromDistribution( sizeDist )
modDist = util.Counter()
modDist[0] = 150
def direcional_expectimax(self, gameState, agent, depth):
if gameState.isLose() or gameState.isWin() or len(
gameState.getLegalActions()) == 0:
return (gameState.getScore(), 0)
if depth == self.depth:
return (self.evaluationFunction(gameState), 0)
actions = gameState.getLegalActions(agent)
nextAgent = agent + 1
if agent == gameState.getNumAgents() - 1:
nextAgent = 0
depth = depth + 1
if agent == 0:
curMax = (-float('inf'), 0)
currentActions = [curMax]
for action in actions:
temp = self.direcional_expectimax(
gameState.generateSuccessor(agent, action), nextAgent,
depth)
if temp[0] == curMax[0]:
currentActions.append((temp[0], action))
elif temp[0] > curMax[0]:
curMax = (temp[0], action)
currentActions = [curMax]
return random.choice(currentActions)
else:
ghostState = gameState.getGhostState(agent)
isScared = ghostState.scaredTimer > 0
pacmanPosition = gameState.getPacmanPosition()
pos = gameState.getGhostPosition(agent)
speed = 1
if isScared: speed = 0.5
actionVectors = [
Actions.directionToVector(a, speed) for a in actions
]
newPositions = [(pos[0] + a[0], pos[1] + a[1])
for a in actionVectors]
distancesToPacman = [
util.manhattanDistance(pos, pacmanPosition)
for pos in newPositions
]
if isScared:
bestScore = max(distancesToPacman)
bestProb = 0.8
else:
bestScore = min(distancesToPacman)
bestProb = 0.8
succ_actions = []
for action in actions:
succ_actions.append(
self.direcional_expectimax(
gameState.generateSuccessor(agent, action), nextAgent,
depth))
bestActions = [
action
for action, distance in zip(succ_actions, distancesToPacman)
if distance == bestScore
]
# Construct distribution
dist = util.Counter()
for a in bestActions:
dist[a] = bestProb / len(bestActions)
for a in succ_actions:
dist[a] += (1 - bestProb) / len(succ_actions)
dist.normalize()
return util.chooseFromDistribution(dist)