def generateVPIHuntersBoard(seed=None):
width = 11
height = 11
foodHouseLeft = util.flipCoin(PROB_FOOD_LEFT)
layoutTextGrid = [[' ' for _ in xrange(width)] for _ in xrange(height)]
layoutTextGrid[0] = ['%' for _ in xrange(width)]
layoutTextGrid[-1] = layoutTextGrid[0][:]
for i in xrange(height):
layoutTextGrid[i][0] = layoutTextGrid[i][-1] = '%'
possibleLocations = pickPossibleLocations(width, height)
# (foodX, foodY), (ghostX, ghostY) = tuple(random.sample(possibleLocations, 2))
bottomLeft, topLeft, bottomRight, topRight = tuple(possibleLocations)
foodX, foodY = topLeft
ghostX, ghostY = topRight
if not util.flipCoin(PROB_FOOD_LEFT):
(foodX, foodY), (ghostX, ghostY) = (ghostX, ghostY), (foodX, foodY)
layoutTextGrid[-foodY-1][foodX] = '.'
layoutTextGrid[-ghostY-1][ghostX] = 'G'
for foodWallX, foodWallY in buildHouseAroundCenter(foodX, foodY):
if util.flipCoin(PROB_FOOD_RED):
layoutTextGrid[-foodWallY-1][foodWallX] = 'R'
else:
layoutTextGrid[-foodWallY-1][foodWallX] = 'B'
for ghostWallX, ghostWallY in buildHouseAroundCenter(ghostX, ghostY):
if util.flipCoin(PROB_GHOST_RED):
layoutTextGrid[-ghostWallY-1][ghostWallX] = 'R'
else:
layoutTextGrid[-ghostWallY-1][ghostWallX] = 'B'
layoutTextGrid[5][5] = 'P'
layoutTextRowList = [''.join(row) for row in layoutTextGrid]
return layoutTextRowList
def mutate(crossed,prob,maxLength,threshold = 1e6): for k,history in enumerate(crossed): for i,ele in enumerate(history): if util.flipCoin(prob): #if util.flipCoin(threshold * 1./fitness(ele,s_belief,t_belief,source_M, M_proj)) if util.flipCoin(0.8): if util.flipCoin(0.5): mutated = list(ele) mutated[0] = random.choice(Actions) crossed[k][i] = tuple(mutated) else: mutated = list(ele) mutated[1] = Obs[np.random.choice(range(len(Obs)),1,p=Obs_p)[0]] crossed[k][i] = tuple(mutated) else: mutated = list(ele) mutated[0] = random.choice(Actions) mutated[1] = Obs[np.random.choice(range(len(Obs)),1,p=Obs_p)[0]] crossed[k][i] = tuple(mutated) if util.flipCoin(prob): if util.flipCoin(0.5) and len(history) < maxLength: mutated = [0,0] #mutated[0] = random.choice(Actions) #mutated[1] = Obs[np.random.choice(range(len(Obs)),1,p=Obs_p)[0]] mutated = history[-1] crossed[k] = history + [tuple(mutated)] elif len(history)>=maxLength-1: crossed[k].pop() return crossed
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
"*** YOUR CODE HERE ***"
if len(legalActions) == 0:
return None
# pick random one of the legal actions if
# flipCoin returns true (exploration)
if util.flipCoin(self.epsilon):
return random.choice(legalActions)
# otherwise, pick the best move determined by the policy (exploitation)
return self.getPolicy(state)
def makeGrid(gridString):
walk=False
if gridString[0][0]=='W':
walk=True
if walk:
obstacleProb=0.0
else:
obstacleProb=0.2
width, height = 10, 3
grid = Grid(width, height)
for h in range(height):
if walk:
grid[width-1][h]=10
elif gridString[0][0]=='C':
grid[width-1][h]=35
else:
grid[width-1][h]=2
for x in range(0,width-1):
for y in range(0,height):
if util.flipCoin(obstacleProb):
grid[x][y]=-2
else:
#grid[x][y]='-1'
grid[x][y]=' '
grid[0][0]='S'
return grid
def chooseAction(self, state):
#return random.choice( state.getLegalActions( self.index ) )
if not self.firstTurnComplete:
self.registerInitialState(state)
self.firstTurnComplete = True
"""
Picks among the actions with the highest Q(s,a).
"""
actions = state.getLegalActions(self.index)
if util.flipCoin(self.explorationRate):
return random.choice(actions)
# You can profile your evaluation time by uncommenting these lines
# start = time.time()
values = [(a, self.evaluate(state, a)) for a in actions]
# print 'eval time for agent %d: %.4f' % (self.index, time.time() - start)
#print 'VALUES: ' + str(values)
maxValue = max(values, key=lambda val : val[1])
bestActions = [a for a, v in zip(actions, values) if v == maxValue]
action = random.choice(bestActions)
self.update(state, action, self.getSuccessor(state, action))
#print 'Features: ' + str(self.getFeatures)
#print 'Weights: ' + str(self.weights)
#print 'Action: ' + str(action) + ' - ' + str(self.getPosition(state)) + '--->' + str(self.getPosition(self.getSuccessor(state, action)))
return action
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
"""Description:
If the flip of the coin is favorable it will chose a random action, else it will get the best
"""
""" YOUR CODE HERE """
if not legalActions:
return None
if util.flipCoin(self.epsilon):
action = random.choice(legalActions)
else:
action = self.getPolicy(state)
""" END CODE """
return action
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
legal_actions = self.getLegalActions(state)
action = None
"*** YOUR CODE HERE ***"
# util.raiseNotDefined()
# return action
if not legal_actions:
return None
# using flip coin fuction for adding randomness when
# choosing action
if util.flipCoin(self.epsilon):
action = random.choice(legal_actions)
else:
action = self.computeActionFromQValues(state)
return action
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
"*** YOUR CODE HERE ***"
if len(legalActions) == 0:
return None
# epsilon greedy, explotation or exploration
if util.flipCoin(self.epsilon):
action = random.choice(legalActions)
else:
action = self.computeActionFromQValues(state)
return action
def playDefense(self,state, offensePlayerIndex):
if state.isWinner(offensePlayerIndex):
return True
if util.flipCoin(self.naiveFactor):
return True
else:
return False
def getAction(self, state):
legalActions = self.getLegalActions(state)
if (util.flipCoin(self.epsilon)): # Epsilon chance to choose random Hit or Stand, or follow policy. Epsilon 0 = Always policy
return random.choice(legalActions)
return self.getPolicy(state)
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
if state not in self.LegalActions:
self.LegalActions[state] = self.getLegalActions(state)
legalActions = self.LegalActions[state]
if util.flipCoin(self.epsilon):
return random.choice(legalActions)
return self.getPolicy(state)
"*** YOUR CODE HERE ***"
util.raiseNotDefined()
return action
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
"*** YOUR CODE HERE ***"
if not self.qTable.has_key(state):
self.qTable[state] = {}
for action in legalActions:
self.qTable[state][action] = 0
if len(legalActions) == 0:
return None
coin = util.flipCoin(self.epsilon)
if coin == True :
action = random.choice(legalActions)
else:
v = -9999
for act in legalActions:
if self.qTable[state][act] > v:
v = self.qTable[state][act]
action = act
return action
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
legal_actions = self.getLegalActions(state)
action = None
"*** YOUR CODE HERE ***"
if len(legal_actions) > 0:
if util.flipCoin(self.epsilon):
action = min([(self.state_uses[(state, act)], act) for act in legal_actions])[1]
print action
if not self.state_uses[(state, action)]:
self.state_uses[(state, action)] = 0
else:
self.state_uses[(state, action)] += 1
else:
action = self.getPolicy(state)
if not self.state_uses[(state, action)]:
self.state_uses[(state, action)] = 0
else:
self.state_uses[(state, action)] += 1
return action
def getAction(self, state):
# Acc to some probability, we take a random action
# Otherwise, we follow the best action available
if util.flipCoin(self.epsilon):
return random.choice(self.getLegalActions(state))
else:
return self.getPolicy(state)
def getAction(self, hitOrnot,position):
"""
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
targetlist=[]
legalDistance=[]
target=()
legalDistance =self.getAllpossibleDistance(position)
# print"GA position is", position
if legalDistance:
if util.flipCoin(self.epsilon):
# print "length",len(legalDistance)
random_Distance = random.randint(0, len(legalDistance)-1)
shoot_distance=legalDistance[random_Distance]
# print "GA shoot_distance:",shoot_distance
targetlist=self.findLocationWithShootDistance(position,shoot_distance)
# print"GA TARGET LIST",targetlist,"len is",len(targetlist)
randomTarget=random.randint(0, len(targetlist)-1)
target=targetlist[randomTarget]
print "shoot randomly at",target,self.q_shot_counter
else:
target = self.getPolicy(hitOrnot,position)
return target
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
"*** YOUR CODE HERE ***"
#Calculate probability for taking actionS
if len(legalActions) > 0:
#Using probability from self.epsilon
if util.flipCoin( self.epsilon):
#Get random action from list using random.choice
action = random.choice( legalActions)
else:
#Get action from Policy pie.
action = self.getPolicy( state)
return action
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
"*** YOUR CODE HERE ***"
if len(legalActions) == 0:
return None
else:
prob = util.flipCoin(self.epsilon)
if prob:
return random.choice(legalActions)
else:
q = util.Counter()
for a in legalActions:
"""if self.getQValue(state, a) > result or action == None :
action = a
result = self.getQValue(state, a)"""
q[state, a] = self.getQValue(state, a)
return q.argMax()[1]
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
"*** YOUR CODE HERE ***"
action = self.computeActionFromQValues(state)
legal_actions=self.getLegalActions(state)
if len(legal_actions)<=1:
return action
#suboptimal_actions.remove(action)
# if state not in self.visit:
# self.visit[state]=0
# self.visit[state]+=1
if util.flipCoin(self.epsilon):#/self.visit[state]
return random.choice(legal_actions)
return action
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
"*** YOUR CODE HERE ***"
# comprobamos si hay legal actions, sino, retornamos None
if not legalActions:
return action
# lanzamos la moneda con epsilon para decidir que accion retornamos, al azar o best policy
if util.flipCoin(self.epsilon):
# retornamos una accion al azar, si no devuelve nada retornamos None
return random.choice(legalActions) or None
# retornamos bestPolicy
return self.getPolicy(state)
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
"*** YOUR CODE HERE ***"
if len(legalActions) < 1:
return None
else:
randomAction = util.flipCoin(self.epsilon) #epsilon = prob of true; 1-epsilon = prob false
if randomAction:
action = random.choice(legalActions)
else:
action = self.getPolicy(state)
print "action", action
return action
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
"*** YOUR CODE HERE ***"
if (self.isTerminal(state)):
return action
prob = self.epsilon;
if util.flipCoin(prob):
action = random.choice(legalActions)
else:
action = self.getPolicy(state)
self.doAction(state,action)
return action
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)
qValLegalActions = util.Counter()
for legalAction in legalActions:
qValLegalAction = self.getQValue(state, legalAction)
qValLegalActions[legalAction] = qValLegalAction
action = None
if len(qValLegalActions) > 0:
action = qValLegalActions.argMax()
if util.flipCoin(self.epsilon):
action = random.choice(legalActions)
return action
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)
actionsToExplore = []
for action in legalActions:
if self.getQValue(state,action) == 0:
actionsToExplore.append(action)
action = None
chooseRandom=util.flipCoin(self.epsilon)
if (chooseRandom):
if (len(actionsToExplore) > 0):
action = random.choice(actionsToExplore)
else:
action = random.choice(legalActions)
else:
action=self.getPolicy(state)
return action
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
"*** YOUR CODE HERE ***"
#if terminal state return None
if len(legalActions)==0:
return None
#check random true or false
randomOrNot= util.flipCoin(self.epsilon)
if randomOrNot:
#Chose east, west, north, south? how do I get the list?
return random.choice(legalActions)
else:
#best policy action get policy or compute action from q values?
return self.computeActionFromQValues(state)
util.raiseNotDefined()
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
"*** YOUR CODE HERE ***"
#OUR CODE HERE
if legalActions==None or len(legalActions) is 0:
return None
#So do we take a random action or not?
if util.flipCoin(self.epsilon): #lyee says: no idea what epsilon is!
#We will take a random action
action= random.choice(legalActions)
else:
#We follow the policy
action = self.getPolicy(state) #lyee fix: kendall previously had just getPolicy.. I added the 'self' part. hope that's what kendall meant D:
return action
def getAction(self, state):
"""
What action to take in the current state. With
probability self.epsilon, we should take a random
action and take the best policy action otherwise.
After you choose an action make sure to
inform your parent self.doAction(state,action)
This is done for you, just don't clobber it
HINT: you might want to use util.flipCoin
here..... (see util.py)
"""
# Pick Action
action = None
epsilon = self.epsilon
take_random_action = util.flipCoin(epsilon)
list_of_actions = self.getLegalActions(state)
if take_random_action:
action = random.choice(list_of_actions)
else:
action = self.getPolicy(state)
#return action
# Need to inform parent of action for Pacman
self.doAction(state,action)
return action
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
"*** YOUR CODE HERE ***"
s = state
legalActions = self.getLegalActions(state)
#print 'LEGAL:'+str(legalActions)
if len(legalActions)==0:
#print 'NONE'
return None
action = None
if util.flipCoin(self.epsilon):
action = random.choice(legalActions)
else:
action = self.getPolicy(s)
#print action
#print 'return' +str(action)
return action
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
"*** YOUR CODE HERE ***"
if len(legalActions) == 0:
#print "No legal actions"
action = None
elif util.flipCoin(self.epsilon):
#print "Random Choice of Action"
action = random.choice(legalActions)
else:
#print "Choice of action based on Policy"
action = self.getPolicy(state)
#print "Action:", action
return action
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)
"""
actions = self.getLegalActions(state)
bestQValue = -99999999
bestActions = []
for action in actions:
q = self.getQValue(state, action)
if q == bestQValue:
bestActions.append(action)
elif q > bestQValue:
bestActions = [action]
bestQValue = q
if len(bestActions) == 0:
return None
# Pick Action
legalActions = self.getLegalActions(state)
action = None
if legalActions:
if util.flipCoin(self.epsilon):
action = random.choice(legalActions)
else:
action = random.choice(bestActions)
return action
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)
"*** YOUR CODE HERE ***"
# terminal state
if not legalActions:
return None
else:
# pick if we should explore by flipping a coin
goRandom = util.flipCoin(self.epsilon)
if goRandom:
# randomly choose an action
return random.choice(legalActions)
else:
# choose the best action
return self.getPolicy(state)
def chooseAction(self, gameState):
"""
@version 1.2.1
"""
self.observationHistory.append(gameState)
actions = gameState.getLegalActions(self.index)
if len(actions) == 0:
return None
if util.flipCoin(self.learningRate):
action = random.choice(actions)
print("now do exploitation")
else:
action = self.computeActionFromQValues(gameState)
self.stepLeft -= 1
if self.considerBack:
#double check actions or bestActions as argument
action = self.updateCarryFood(
gameState, self.backToSafetyPosition(gameState, actions))
self.lastState = gameState
# action = self.updateCarryFood(gameState, random.choice(bestActions))
self.lastAction = action
# self.updateWeights(self.lastState, self.lastAction, gameState)
print "above all, I choose ", action
print "============================="
print "=============ends============"
print "============================="
return self.updateCarryFood(gameState, action)
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
"*** YOUR CODE HERE ***"
################################################
# 0039026 #
###########
"""Instance variables you have access to
- self.epsilon (exploration prob)
- self.alpha (learning rate)
- self.discount (discount rate)"""
#util.flipCoin(p)
#print util.flipCoin(10)
#print self.epsilon
bool = util.flipCoin(self.epsilon)
if not bool:
#print "COINhere false ----------------"
action = self.computeActionFromQValues(state)
else:
#print "COINhere true ----------------"
action = random.choice(legalActions)
return action
def getAction(self, state):
legalActions = self.getLegalActions(state)
action = None
"end game"
if not self.getLegalActions(state):
return action
if self.new_episode:
"Analyze state"
self.state_discription = self.getStateDiscription(state)
self.new_episode = False
"Explore or Exploit"
if util.flipCoin(self.epsilon):
action = random.choice(legalActions)
else:
"predict based on current state"
values = self.first_model.predict(
np.array([self.state_discription]))
actions = copy.deepcopy(ACTIONS)
actions = [
action
for _, action in sorted(zip(values, actions), reverse=True)
]
for a in actions:
if a in legalActions:
actions = a
break
if action not in legalActions:
action = ACTIONS[4]
self.doAction(state, action)
return action
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
"*** YOUR CODE HERE ***"
if len(legalActions) == 0:
return action
if util.flipCoin(self.epsilon):
action = random.choice(legalActions)
else:
action = self.computeActionFromQValues(state)
return action
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)
isHeads = util.flipCoin(self.epsilon)
if len(legalActions) is 0:
return None
if isHeads:
#print "Taking the known policy"
return random.choice(legalActions)
else:
#print "Taking the random choice"
return self.getPolicy(state)
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
"*** YOUR CODE HERE ***"
#util.raiseNotDefined()
if len(legalActions
) != 0: #if there are no legal actions, returns action none
if util.flipCoin(self.epsilon
) == True: #getting probability of exploration.
action = random.choice(legalActions) #getting randomly actions
else:
action = self.getPolicy(state)
return action
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)
"""
"*** YOUR CODE HERE ***"
# Pick Action
random_choice = util.flipCoin(self.epsilon)
legalActions = self.getLegalActions(state)
action = self.getPolicy(state)
if random_choice:
if not legalActions:
return None
return util.random.choice(legalActions)
return action
def getAction(self, state, withEpsilon=True):
"""
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.
"""
# print "getAction QLearningAgentt"
# Pick Action
print(self.epsilon)
legalActions = self.getLegalActions(state)
action = None
"*** YOUR CODE HERE ***"
if len(legalActions) == 0:
return action
if withEpsilon and util.flipCoin(self.epsilon):
action = random.choice(legalActions)
else:
action = self.computeActionFromQValues(state)
return action
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
actionList = self.getLegalActions(state)
#returns true with self.epsilon probability
probability = util.flipCoin(self.epsilon)
#if true, explore
if probability:
return random.choice(actionList)
#if false, do basic value-iteration
else:
return self.computeActionFromQValues(state)
def chooseAction(self, state):
# Append game state to observation history...
self.observationHistory.append(state)
# Pick Action
legalActions = state.getLegalActions(self.index)
action = None
if (DEBUG):
print self.newline()
print "AGENT " + str(self.index) + " choosing action!"
if len(legalActions):
if util.flipCoin(self.epsilon) and self.isTraining():
action = random.choice(legalActions)
if (DEBUG):
print "ACTION CHOSE FROM RANDOM: " + action
else:
action = self.computeActionFromQValues(state)
if (DEBUG):
print "ACTION CHOSE FROM Q VALUES: " + action
self.lastAction = action
foodLeft = len(self.getFood(state).asList())
# Prioritize going back to start if we have <= 2 pellets left
if foodLeft <= 2:
bestDist = 9999
for a in legalActions:
successor = self.getSuccessor(state, a)
pos2 = successor.getAgentPosition(self.index)
dist = self.getMazeDistance(self.start, pos2)
if dist < bestDist:
action = a
bestDist = dist
if (DEBUG):
print "AGENT " + str(self.index) + " chose action " + action + "!"
return action
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
"*** YOUR CODE HERE ***"
possible = util.flipCoin(self.epsilon)
# print('This is the possibility', possible)
if possible:
action = random.choice(legalActions)
else:
action = self.getPolicy(state)
# util.raiseNotDefined()
return action
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)
"""
"*** YOUR CODE HERE ***"
#Local Declarations
action = None
actionList = self.getLegalActions(state)
#Error check, make sure list of actions is not empty
if not (actionList):
return None
#Check over probabiliy choice; choose action randomly or from computed QValues
if (util.flipCoin(self.epsilon)):
action = random.choice(actionList)
else:
action = self.computeActionFromQValues(state)
return action
def generateRandomHuntersBoard(seed=None, width=None, height=None):
"""Note that this is constructing a string, so indexing is [-y-1][x] rather than [x][y]"""
random.seed(seed)
leftHouseTop = util.flipCoin(PROB_LEFT_TOP)
if not width or not height:
width = random.randrange(11, 20, 4)
height = random.randrange(11, 16, 4)
layoutTextGrid = [[' ' for _ in range(width)] for _ in range(height)]
layoutTextGrid[0] = ['%' for _ in range(width)]
layoutTextGrid[-1] = layoutTextGrid[0][:]
for i in range(height):
layoutTextGrid[i][0] = layoutTextGrid[i][-1] = '%'
possibleLocations = pickPossibleLocations(width, height)
# (foodX, foodY), (ghostX, ghostY) = tuple(random.sample(possibleLocations, 2))
bottomLeft, topLeft, bottomRight, topRight = tuple(possibleLocations)
if leftHouseTop:
foodX, foodY = topLeft
ghostX, ghostY = bottomRight if util.flipCoin(PROB_OPPOSITE_CORNERS) else topRight
else:
foodX, foodY = bottomLeft
ghostX, ghostY = topRight if util.flipCoin(PROB_OPPOSITE_CORNERS) else bottomRight
if not util.flipCoin(PROB_FOOD_LEFT):
(foodX, foodY), (ghostX, ghostY) = (ghostX, ghostY), (foodX, foodY)
layoutTextGrid[-foodY-1][foodX] = '.'
layoutTextGrid[-ghostY-1][ghostX] = 'G'
for foodWallX, foodWallY in buildHouseAroundCenter(foodX, foodY):
if util.flipCoin(PROB_FOOD_RED):
layoutTextGrid[-foodWallY-1][foodWallX] = 'R'
else:
layoutTextGrid[-foodWallY-1][foodWallX] = 'B'
for ghostWallX, ghostWallY in buildHouseAroundCenter(ghostX, ghostY):
if util.flipCoin(PROB_GHOST_RED):
layoutTextGrid[-ghostWallY-1][ghostWallX] = 'R'
else:
layoutTextGrid[-ghostWallY-1][ghostWallX] = 'B'
layoutTextGrid[-2][1] = 'P'
layoutTextRowList = [''.join(row) for row in layoutTextGrid]
return layoutTextRowList
def mutate(self, p):
newString = ""
for i in xrange(self.n):
if util.flipCoin(p): newString += str(int(not int(self.string[i])))
else: newString += self.string[i]
self.string = newString
def getAction(self, state):
"""
Compute the action to take in the current state.
HINT: You might want to use util.flipCoin(prob)
HINT: To pick randomly from a list, use random.choice(list)
"""
action = None
legalActions = self.getLegalActions(state)
# List of actions, excluding those that are known to simply keep the agent in the same state
possibleActions = []
# List of actions and their respective Q values
possibleActionQValues = util.Counter()
# List of actions and the absolute values of their respective Q values,
# excluding actions that would cause the agent to backtrack to the previous state
possibleActionsNoBacktrack = util.Counter()
absPossibleActionsNoBacktrack = util.Counter()
# Assemble lists of actions that are permitted depending on the circumstances
for action in legalActions:
if (state, action) not in self.forbiddenActions:
possibleActionQValues[action] = self.getQValue(state, action)
possibleActions.append(action)
if not self.isBacktrackAction(action):
possibleActionsNoBacktrack[action] = self.getQValue(
state, action)
absPossibleActionsNoBacktrack[action] = abs(
self.getQValue(state, action))
if len(possibleActionQValues) > 0:
print "goalMode: ", self.goalMode
print "epsilon 1: ", self.epsilon1
print "epsilon 2: ", self.epsilon2
print "meanQValue: ", self.getValue(state)
print "possibleActions: ", possibleActions
print "possibleActionQValues: ", possibleActionQValues
# Training to populate Q table
if self.goalMode == 'maxQMode':
if util.flipCoin(0.5):
# action = possibleActionsNoBacktrack.argMax()
action = possibleActionQValues.argMax()
# action = random.choice(possibleActions)
else:
action = random.choice(possibleActions)
elif self.goalMode == 'minQMode':
if util.flipCoin(self.epsilon1):
# action = possibleActionsNoBacktrack.argMin()
action = possibleActionQValues.argMin()
# action = random.choice(possibleActions)
else:
action = random.choice(possibleActions)
print "Random: ", action
# Training to find deceptive path
else:
largestQValue = possibleActionQValues.argMax()
print "Equilibrium state: ", state, self.isEquilibriumState(
state)
# If agent has already found an equidistant state with the largest-Q-value action seen so far,
# then continue to the true goal
if self.maxQValuePolicy:
action = possibleActionsNoBacktrack.argMax()
# Otherwise, keep searching for the equidistant state that has the largest-Q-value action.
else:
# If the agent has arrived at (what was thought to be) the LDP, and found that this state
# no longer has at least one positively valued action and at least one negatively valued action,
# then forget about this state.
if self.lastDeceptivePoint is not None and state == self.lastDeceptivePoint[
0] and not self.isEquilibriumState(state):
self.lastDeceptivePoint = None
# If an equidistant state has been found...
if self.isEquilibriumState(state):
# If the agent has arrived at an equidistant state that has the largest-Q-value action
# seen so far (or if the agent has arrived at what is currently thought to be the LDP),
# then update the details of the likeliest LDP candidate...
if self.lastDeceptivePoint is None\
or possibleActionQValues.get(largestQValue) > self.lastDeceptivePoint[1]\
or state == self.lastDeceptivePoint[0]:
self.lastDeceptivePoint = (
state,
possibleActionQValues.get(largestQValue))
# Now head directly to the true goal, with probability epsilon2...
if util.flipCoin(1 - self.epsilon2):
self.maxQValuePolicy = True
action = possibleActionsNoBacktrack.argMax()
# Or continue searching for equidistant states that might have a larger Q value
else:
action = absPossibleActionsNoBacktrack.argMin()
if self.epsilon2 >= 1.0 / float(
self.phaseTwoEpisodes):
self.epsilon2 -= 1.0 / float(
self.phaseTwoEpisodes)
# If this equidistant state does NOT have the largest-Q-value action
# of all equidistant states seen so far, then keep searching for such an equidistant state
else:
action = absPossibleActionsNoBacktrack.argMin()
# Otherwise, keep searching for an equidistant state:
else:
if self.getValue(state) > 0:
action = possibleActionsNoBacktrack.argMin()
elif self.getValue(state) < 0:
action = possibleActionsNoBacktrack.argMax()
else:
action = absPossibleActionsNoBacktrack.argMin()
print "self.lastDeceptivePoint: ", self.lastDeceptivePoint
return action
def getAction(self, state):
if util.flipCoin(self.epsilon) is True:
return random.choice(self.legalActions)
return self.computeActionFromQValues(state)
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
"*** YOUR CODE HERE ***"
if not legalActions:
action = None
elif util.flipCoin(self.epsilon):
action = random.choice(legalActions)
else:
action = self.computeActionFromQValues(state)
# if action not in legalActions:
# print("WTFFFF")
action_copy = action
if self.shield:
safe = False
legal_qval = []
# print(legalActions)
# legalActions.remove(action)
for ac in legalActions:
if ac != action:
legal_qval.append((ac, self.getQValue(state, ac)))
sorted(legal_qval, key=lambda x: x[1], reverse=True)
i = 0
self.discarded = []
# print(legal_qval)
while not safe and len(legalActions) != 0:
# if action not in legalActions:
# print("WTFFFFF")
px, py = state.getPacmanPosition()
if action == 'East':
px += 1
elif action == 'West':
px -= 1
elif action == 'North':
py += 1
elif action == 'South':
py -= 1
num_ghosts = len(state.data.agentStates) - 1
safe = True
for j in range(num_ghosts):
ghostpos = state.getGhostPosition(j + 1)
dist = manhattanDistance((px, py), ghostpos)
# print(dist)
if dist < 2.0:
safe = False
self.discarded.append(action)
legalActions.remove(action)
break
# if action == 'Stop':
# safe = False
# print(action)
if not safe and i < len(legal_qval):
action = legal_qval[i][0]
i += 1
if len(legalActions) == 0:
action = action_copy
return action
def chooseOfAction(self, gameState):
# Pick Action
ghost=[]
ghostIndex = 0
opAgents = CaptureAgent.getOpponents(self,gameState)
currentPos = gameState.getAgentPosition(self.index)
# Get ghost locations and states if observable
if opAgents:
for opponent in opAgents:
opPos = gameState.getAgentPosition(opponent)
opIsPacman = gameState.getAgentState(opponent).isPacman
if opPos and not opIsPacman:
dis = abs(currentPos[0]-opPos[0])+abs(currentPos[1]-opPos[1])
if dis<=6:
ghost.append(opPos)
ghostIndex = opponent
if len(self.getFood(gameState).asList())>2:
if len(ghost) ==0 :
if gameState.getAgentState(self.index).numCarrying>1 and gameState.data.timeleft<200:
self.weights =self.weights4
print("444444444444444444444")
else:
self.weights = self.weights1
print("111111111111111111111")
else:
if min([self.getMazeDistance(gameState.getAgentPosition(self.index),a) for a in ghost])>6:
self.weights = self.weights1
print("111111111111111111111")
else:
if gameState.getAgentState(ghostIndex).scaredTimer<10:
if gameState.data.timeleft<200 :
if gameState.getAgentState(self.index).numCarrying>2:
self.weights = self.weights3
print("33333333333333333333")
else:
self.weights = self.weights2
print("2222222222222222222222")
else:
if gameState.getAgentState(self.index).numCarrying>10:
if self.red:
middle = int((gameState.data.layout.width - 2)/2 )
else:
middle = int((gameState.data.layout.width - 2)/2 + 1)
if abs(gameState.getAgentPosition(self.index)[0]-middle) < middle/2:
self.weights = self.weights3
print("33333333333333333333")
else :
self.weights = self.weights2
print("2222222222222222222222")
else:
self.weights = self.weights2
print("2222222222222222222222")
else :
self.weights = self.weights1
print("111111111111111111111")
else :
if len(ghost) ==0:
self.weights = self.weights4
print("44444444444444444444")
else:
if gameState.getAgentState(ghostIndex).scaredTimer<10:
self.weights = self.weights3
print("33333333333333333333")
else:
self.weights = self.weights4
print("44444444444444444444")
legalActions = gameState.getLegalActions(self.index)
legalActions.remove(Directions.STOP)
action = None
if len(legalActions) != 0:
prob = util.flipCoin(self.epsilon)
if prob:
action = random.choice(legalActions)
else:
if self.weights ==self.weights1:
action = self.getSafePolicy(gameState)
else:
action = self.getPolicy(gameState)
if self.weights == self.weights2:
food = self.getFeatures(gameState,action)["closest-food"]
ghost = self.getFeatures(gameState,action)["closest-ghosts"]
print(food*100,ghost*100)
if not gameState.getAgentState(self.index).isPacman:
if self.red:
if self.finish:
self.mode =2
else:
if self.finish:
self.mode = 2
return action
def getAction(self, state):
############## Build network map ##############
# This builds a network of positions pacman can move to based on a state.
# Required for search algorithms. This is specific to this maze.
if not self.map_built: # If map isn't built, build it
self.w_pos = self.wall_pos(state) # Get wall locations
self.map_graph = self.map_graph(self.w_pos) # Get network
self.map_built = True # Turn off map building
# Update weights
if self.played: # Have we played yet?
if self.learning:
############## Get Reward of state ##############
self.r = self.reward_signal(state.getScore(), self.old_score)
self.old_score = state.getScore(
) # Udate the old score for next time
############## Extract legal moves ##############
# Get legal actions & remove STOP
legal = state.getLegalPacmanActions()
if Directions.STOP in legal:
legal.remove(Directions.STOP)
# Convert available actions in form of int
available_actions = self.AvailableActions(legal)
############## Calculate Max Q(s', a') ##############
# Coordinates for grids around pacman
south = (state.getPacmanPosition()[0],
state.getPacmanPosition()[1] - 1)
east = (state.getPacmanPosition()[0] + 1,
state.getPacmanPosition()[1])
west = (state.getPacmanPosition()[0] - 1,
state.getPacmanPosition()[1])
north = (state.getPacmanPosition()[0],
state.getPacmanPosition()[1] + 1)
directions = [south, east, west, north]
possible_directions = [
] # List with +1 moves to examine in legal
direction_integer = [] # Mapping grid examined with direction
# Examine available actions and store grids to look at for pacman
if 0 in available_actions:
possible_directions.append(east)
direction_integer.append(0)
if 1 in available_actions:
possible_directions.append(south)
direction_integer.append(1)
if 2 in available_actions:
possible_directions.append(west)
direction_integer.append(2)
if 3 in available_actions:
possible_directions.append(north)
direction_integer.append(3)
# Stores Q(s', a') values, and their respective function scores
Q_values = []
f1_score = []
f2_score = []
# Looking at possible actions and compute Q values
for i in possible_directions:
f1 = self.Distance_to_food(state, state.getFood(), i,
self.map_graph,
state.getGhostPositions())
f2 = self.Distance_to_Ghost(state,
state.getGhostPositions(), i,
self.map_graph)
f1_score.append(f1)
f2_score.append(f2)
Q = self.weights[
0] + self.weights[1] * f1 + self.weights[2] * f2
Q_values.append(Q)
# Choose the best action
index = Q_values.index(max(Q_values))
action = self.IntToMove(direction_integer[index])
############## Weight Updates ##############
# UPDATE THE WEIGHTS
difference = self.r + self.gamma * max(Q_values) - self.Qsa
self.weights[0] = self.weights[0] + self.alpha * difference
self.weights[1] = self.weights[
1] + self.alpha * difference * f1_score[index]
self.weights[2] = self.weights[
2] + self.alpha * difference * f2_score[index]
# Save the chosen action's previous function scores
self.Qsa = max(
Q_values
) # This will be Q(s,a) after the move has been made
# Save f values in case of death
self.f1_death = f1_score[index]
self.f2_death = f2_score[index]
else:
############## Extract legal moves ##############
# Get legal actions & remove STOP
legal = state.getLegalPacmanActions()
if Directions.STOP in legal:
legal.remove(Directions.STOP)
# Convert available actions in form of int
available_actions = self.AvailableActions(legal)
############## Calculate Max Q(s', a') ##############
# Coordinates for grids around pacman
south = (state.getPacmanPosition()[0],
state.getPacmanPosition()[1] - 1)
east = (state.getPacmanPosition()[0] + 1,
state.getPacmanPosition()[1])
west = (state.getPacmanPosition()[0] - 1,
state.getPacmanPosition()[1])
north = (state.getPacmanPosition()[0],
state.getPacmanPosition()[1] + 1)
directions = [south, east, west, north]
possible_directions = [] # List with +1 moves to examine in legal
direction_integer = [] # Mapping grid examined with direction
# Examine available actions and store grids to look at for pacman
if 0 in available_actions:
possible_directions.append(east)
direction_integer.append(0)
if 1 in available_actions:
possible_directions.append(south)
direction_integer.append(1)
if 2 in available_actions:
possible_directions.append(west)
direction_integer.append(2)
if 3 in available_actions:
possible_directions.append(north)
direction_integer.append(3)
# Stores Q(s', a') values, and their respective function scores
Q_values = []
f1_score = []
f2_score = []
# Looking at possible actions and compute Q values
for i in possible_directions:
f1 = self.Distance_to_food(state, state.getFood(), i,
self.map_graph,
state.getGhostPositions())
f2 = self.Distance_to_Ghost(state, state.getGhostPositions(),
i, self.map_graph)
f1_score.append(f1)
f2_score.append(f2)
Q = self.weights[
0] + self.weights[1] * f1 + self.weights[2] * f2
Q_values.append(Q)
# Choose the best action
index = Q_values.index(max(Q_values))
action = self.IntToMove(direction_integer[index])
# Save the function values
self.Qsa = max(
Q_values) # This will be Q(s,a) after the move has been made
# Begin learning
if self.learning:
self.played = True
# Exploration function
if util.flipCoin(self.epsilon) and self.learning:
choices = range(0, len(Q_values))
index = random.choice(choices)
action = self.IntToMove(direction_integer[index])
self.Qsa = Q_values[
index] # This will be Q(s,a) after the move has been made
print self.weights
return action
for i in range(20):
nasa = open('i%d.txt' % i, 'r')
skipLine = 9
lineIndex = 0
result = {}
p = [0, 0, 0, 0]
for line in nasa:
if lineIndex > skipLine:
line = line.strip()
v = line.split(' ')
r = [int(v[0]), int(v[1]), int(v[2]), float(v[6])]
r[-1] *= random.choice(surface)
if not util.flipCoin(random.choice(health_prob)):
r[-1] = 0
r[-1] = int(r[-1])
if p[1] not in result:
result[p[1]] = {}
if p[2] not in result[p[1]]:
result[p[1]][p[2]] = {}
if p[3] not in result[p[1]][p[2]]:
result[p[1]][p[2]][p[3]] = {}
if r[3] not in result[p[1]][p[2]][p[3]]:
result[p[1]][p[2]][p[3]][r[3]] = 0
def getAction(self, state):
############## Build network map ##############
# This builds a network of positions pacman can move to based on a state.
# Required for search algorithms. This is specific to this maze.
if not self.map_built: # If map isn't built, build it
self.w_pos = self.wall_pos(state) # Get wall locations
self.map_graph = self.map_graph(self.w_pos) # Get network
self.map_built = True # Turn off map building
############## Extract legal moves ##############
# Get legal actions & remove STOP
legal = state.getLegalPacmanActions()
if Directions.STOP in legal:
legal.remove(Directions.STOP)
# Convert available actions in form of int
available_actions = self.AvailableActions(legal)
############## Extract State information S' ##############
# Get CURRENT state info S' (in form of key)
self.s_cur = self.key(state.getPacmanPosition(),
state.getGhostPositions(), state.getFood(),
self.map_graph, state)
############## Update Q-Table ##############
# If this is not the first action we make in a game. If it is, skip to "else"
if self.played:
# If this is the first time we have seen that state, initialize key-value pairs of all possible actions to 0. If not the dictionary will be empty and not function. Allows us to add states as we see them.
for i in available_actions:
if self.Q.get((self.s_cur, i)) == None:
self.Q[self.s_cur, i] = 0
# Get the current reward (R')
self.r_cur = self.reward_signal(state.getScore(), self.old_score)
# Update old score
self.old_score = state.getScore()
# Increment the state/action pair that we were in previously. (Nsa += 1)
self.Nsa[(self.s, self.a)] += 1
# Calculate alpha adjustment based on Nsa (if activated)
if self.alpha_adjustment:
self.alpha = self.adjusted_alpha(self.Nsa[(self.s, self.a)])
else: # Use regular alpha if not active
self.alpha = self.alpha
# Update the Q Table for previous state/action pair
self.Q[(self.s,
self.a)] = self.Q[(self.s, self.a)] + self.alpha * (
self.r + self.gamma * max(self.Q[(self.s_cur, i)]
for i in available_actions) -
self.Q[(self.s, self.a)])
else:
# This code is only run once at the beginning of each game.
# Initialize the current reward for starting
self.r_cur = state.getScore()
self.old_score = state.getScore(
) # "Old score" is the same as current score at t = 0.
# Initialize playing state. We will not come here again until the new game.
self.played = True
# Ensure dictionary is not empty for current starting position and available actions.
# They are initialized to 0.
for i in available_actions:
if self.Q.get((self.s_cur, i)) == None:
self.Q[self.s_cur, i] = 0
############## Update S, R ##############
# Adjust state, and reward. We have already updated the Q table so this will only be relevant at the next table update (given we survived an extra move)
self.s = self.s_cur
self.r = self.r_cur
############## Chosing argmax Q(s', a') and updating A ##############
self.scores = [
] # Will keep track of all rewards for each action in legal
############## If using function exploration
if self.function_exploration:
## Adjust action (need the arg max Q(s', a'))
# Obtaining the action which maximizes the rewards. Examine all possible actions
# and store their Q-values in a list
for i in available_actions:
# If the state I can go to hasn't been visited enough time, incentivise it properly using the large reward. Agent must also be in a state of learning
if (self.Nsa[(self.s_cur, i)] < self.Ne) and self.learning:
self.scores.append(self.L_Reward)
# If it has, get the true calculated utility of that state
else:
self.scores.append(self.Q[(self.s_cur, i)])
# Verify that the number of scores which are equal to max score. This will be used to make a random choice if we have several unseen state-action pairs.
counter = 0 # Serves as a counter and index for max score
max_score_index = []
for i in self.scores:
if i == max(self.scores):
max_score_index.append(counter)
counter += 1
# Extract the index for the highest score. Either randomly when there is more than one max score, or the first element in the list when there is only 1. This is needed to map the score back with the action which produced it.
if max_score_index > 1:
max_ = random.choice(max_score_index)
else:
max_ = max_score_index[0]
# Map the index corresponding to the highest score back to its respective action in available_actions
self.a = available_actions[max_]
# Convert int action to actual action and return move.
action = self.IntToMove(self.a)
############## If using epsilon exploration (not used)
if self.epsilon_exploration:
for i in available_actions:
self.scores.append(self.Q[(self.s_cur, i)])
# If less than epsilon, and we're learning, make a random choice
if util.flipCoin(self.epsilon) and self.learning:
self.a = random.choice(available_actions)
else:
max_ = self.scores.index(max(self.scores))
self.a = available_actions[max_]
action = self.IntToMove(self.a)
############## Return Action Arg Max Q(S', A') ##############
return action
def __le__(self, other):
return util.flipCoin(0.5)
def chooseAction(self, state):
start = time.time()
self.debugDraw([(0, 0)], [0, 0, 0], clear=True)
self_agent = state.getAgentState(self.index)
actions = state.getLegalActions(self.index)
food = state.getBlueFood().count(
True) if self.isOnRedTeam else state.getRedFood().count(True)
# Particle filtering
self.observeState()
particle_filtering_time = time.time() - start
# If we're carrying enough, just go home!
if self_agent.numCarrying >= 3:
return self.returnHome(state)
# Otherwise, run minimax
elif self.use_minimax:
max_score = -99999
max_action = None
alpha = -99999
beta = 99999
for action in actions:
# Update successor ghost positions to be the max pos in our particle distributions
successor = state.generateSuccessor(self.index, action)
ghosts = self.getBeliefDistribution().argMax()
successor = self.setGhostPositions(successor, ghosts,
self.getOpponents(state))
time_depth = 1 - particle_filtering_time - 0.2
result = self.minimax(successor, start, 1, alpha, beta, 1,
time_depth)
if result >= max_score:
max_score = result
max_action = action
if max_score > beta:
return max_action
alpha = max(alpha, max_score)
action = max_action
# Or compute action from q-values
else:
action = random.choice(actions) if util.flipCoin(
self.epsilon) else self.computeActionFromQValues(state)
# Q-learning
if self.learn:
reward = self.getReward(
state.generateSuccessor(self.index, action), state)
self.update(state, action,
state.generateSuccessor(self.index, action), reward)
# Draw particle distribution
# self.drawBeliefs()
# Update particles
self.elapseTime(state)
end = time.time()
if end - start > 1: print("Overtime --> total time was ", end - start)
return action
def actionSelector(self, gameState):
if util.flipCoin(self.epsilon):
return random.choice(self.getLegalActions(gameState))
return self.getPolicy(gameState)
def chooseAction(self,gameState):
start = time.time()
actions = gameState.getLegalActions(self.index)
action = None
foodLeft = len(self.getFood(gameState).asList())
myCurrentPos = gameState.getAgentState(self.index).getPosition()
InitialPosition = gameState.getInitialAgentPosition(self.index)
enemies = []
enemyGhost = []
enemyPacman = []
for opponent in self.getOpponents(gameState):
enemy = gameState.getAgentState(opponent)
enemies.append(enemy)
#print"enemies", enemies
enemyGhost = [a for a in enemies if not a.isPacman and a.getPosition() != None]
enemyPacman = [a for a in enemies if a.isPacman and a.getPosition() != None]
ghostPositions = []
disToG = 6666
ranges = []
enemyGhostPosition = [Ghost.getPosition() for Ghost in enemyGhost]
enemyPacmanPosition = [Pacman.getPosition() for Pacman in enemyPacman]
mid = gameState.data.layout.width / 2
if gameState.isOnRedTeam(self.index):
mid = mid - 1
else:
mid = mid + 1
legalPositions = [p for p in gameState.getWalls().asList(False) if p[1] > 1]
border = [p for p in legalPositions if p[0] == mid]
if len(enemyGhostPosition) >0 and not gameState.getAgentState(self.index).isPacman:
disToG = min([self.getMazeDistance(myCurrentPos, ghostPos) for ghostPos in enemyGhostPosition])
print'position in action',myCurrentPos
print'If Astar111',self.aSt
if disToG <5:
randomPoint = random.choice(border)
while randomPoint == myCurrentPos:
randomPoint = random.choice(border)
self.farPoint = randomPoint
self.aSt = True
print'If Astar',self.aSt
print'current pacman position',myCurrentPos
self.stopAction = True
actionList = []
if myCurrentPos!= self.farPoint:
if self.aSt:
bestDist = 9999
#for action2 in actions:
#successor = self.getSuccessor(gameState, action2,self.farPoint)
#pos2 = successor.getAgentPosition(self.index)
#action3 = None
print'self.actionList outside if-else',self.actionList
if len(self.actionList) == 0:
#self.actionList.remove('Stop')
self.actionList = self.aStar(gameState,self.farPoint,myCurrentPos,start)
#self.actionList = action3
print'self.actionList',self.actionList
if len(self.actionList) != 0:
if self.actionList[0] == 9999:
self.aSt = False
self.skipAstar = True
self.actionList.remove(9999)
print'skip astar first'
elif self.actionList[0] == 9999:
self.aSt = False
self.skipAstar = True
self.actionList.remove(9999)
print'skip a star'
else:
#actionList = self.actionList
#if len(self.actionList)>0:
bestAction = self.actionList[0]
print 'eval time in A STAR for agent %d: %.4f' % (self.index, time.time() - start)
print'bestAction',bestAction
self.actionList.remove(bestAction)
if myCurrentPos == InitialPosition or len(self.actionList) == 0:
self.aSt = False
else:
#for move in actions:
if bestAction in actions:
return bestAction
else:
self.aSt = False
#return bestAction
else:
self.aSt = False
#bestDist = dist
#print'bestAction',bestAction
#return bestAction
else:
self.aSt = False
if len(actions) !=0:
probability = util.flipCoin(self.epsilon)
if probability:
#print"random"
action = random.choice(actions)
else:
action = self.getPolicy(gameState)
action = self.getPolicy(gameState)
"""if self.stopAction and not self.skipAstar:
print'stop for a sec'
action = Directions.STOP
self.stopAction = False"""
"""elif self.aSt:
action = Directions.STOP
self.stopAction = False
print'stop second time'"""
print"final action",action
print 'eval time in Q-Learning for agent %d: %.4f' % (self.index, time.time() - start)
return action
def p():return n0tf0rn00bs.flipCoin(mutationProbability) q=np.array([z if p()else 1-z for z in chromosome])
def chooseAction(self, gameState):
"""
Picks the best of all legal actions based on their estimated Q values, which are computed
with a linear combination of the feature values and their weights.
This is the function that is called at every turn; any other functions that should be called
each turn should be called from here.
"""
# Only update weights if we are currently training
if TRAINING:
# First, update weights based on reward received from the move we just took
s_prime = self.getCurrentObservation()
a = self.lastAction
agentName = "Offensive Agent" if isinstance(self, OffensiveDummyAgent) else "Defensive Agent"
showOutput = (DEBUG_OFFENSE_ONLY and agentName is "Offensive Agent") or (DEBUG_DEFENSE_ONLY and agentName is "Defensive Agent")
if showOutput:
print(agentName + " action just taken: " + str(a))
s = self.getPreviousObservation() if len(
self.observationHistory) > 1 else s_prime
reward = self.getReward(s, a, s_prime) # sets self.justDied = 25
self.updateWeights(s, a, s_prime, reward)
# Choose our next action!
actions = gameState.getLegalActions(self.index)
qValuesOfNextActions = [self.evaluatePotentialNextState(
gameState, a) for a in actions]
maxValue = max(qValuesOfNextActions)
bestActions = [a for a, v in zip(
actions, qValuesOfNextActions) if v == maxValue]
# If there are 2 (or fewer) pellets left, the game is pretty much over, so the best action will be
# the one that moves us closer to where we initially started
foodLeft = len(self.getFood(gameState).asList())
if foodLeft <= 2:
actionChoice = self.getActionToGoBackHome(gameState, actions)
self.lastAction = actionChoice
agentName = "Offensive Agent" if isinstance(self, OffensiveDummyAgent) else "Defensive Agent"
showOutput = (DEBUG_OFFENSE_ONLY and agentName is "Offensive Agent") or (DEBUG_DEFENSE_ONLY and agentName is "Defensive Agent")
if showOutput:
print(agentName + " CHOOSING ACTION: " + str(actionChoice))
return actionChoice
# Do a coin flip with probability self.epsilon to choose randomly instead of optimally, only if TRAINING
coin_flip = util.flipCoin(self.epsilon)
if coin_flip and TRAINING:
actionChoice = random.choice(actions)
self.lastAction = actionChoice
agentName = "Offensive Agent" if isinstance(self, OffensiveDummyAgent) else "Defensive Agent"
showOutput = (DEBUG_OFFENSE_ONLY and agentName is "Offensive Agent") or (DEBUG_DEFENSE_ONLY and agentName is "Defensive Agent")
if showOutput:
print(agentName + "CHOOSING ACTION: " + str(actionChoice))
return actionChoice
# In all other cases, choose the best action based on computed Q values
# If multiple actions are tied, break ties randomly.
actionChoice = random.choice(bestActions)
self.lastAction = actionChoice
agentName = "Offensive Agent" if isinstance(self, OffensiveDummyAgent) else "Defensive Agent"
showOutput = (DEBUG_OFFENSE_ONLY and agentName is "Offensive Agent") or (DEBUG_DEFENSE_ONLY and agentName is "Defensive Agent")
if showOutput:
print(agentName + " CHOOSING ACTION: " + str(actionChoice))
return actionChoice
messageCallback = lambda x: None
else:
if opts.manual and opts.agent == None:
displayCallback = lambda state: display.displayNullValues(state)
else:
if opts.agent == 'random': displayCallback = lambda state: display.displayValues(a, state,
"CURRENT VALUES")
if opts.agent == 'value': displayCallback = lambda state: display.displayValues(a, state,
"CURRENT VALUES")
if opts.agent == 'q': displayCallback = lambda state: display.displayQValues(a, state,
"CURRENT Q-VALUES")
if episode < a.phaseOneEpisodes:
if util.flipCoin(0.5):
goalMode = 'maxQMode'
else:
goalMode = 'minQMode'
else:
goalMode = 'deceptiveMode'
# Run episode according to goal mode
returns += runEpisode(a, env, opts.discount, decisionCallback, displayCallback, messageCallback, pauseCallback,
episode, goalMode)
a.episodesSoFar += 1
a.maxQValuePolicy = False
a.stepCount = 0
if episodes > 0:
print
first = False
else:
# alpha = (alpha) / (alpha + count / 100) # decay learning rate
count += 1
#if count % 300 == 0:
# Q_values = util.Counter()
status = getStatus(input, output) # win or tie or lost
if status == 'win':
reward = 1
elif status == 'tie':
reward = 0
else:
reward = -1
change = getChange(input, last_input) # + or - or o
new_state = (status, change)
new_q = max([Q_values[(new_state, action)] for action in actions])
Q_values[(state, output)] = (1.0 - alpha) * Q_values[(state, output)] + \
alpha * (reward + gamma * new_q)
state = new_state
if count < 20: #count % 100 == 1:
f.write('{0} {1} {2} {3}\n'.format(input, output, reward, Q_values))
if util.flipCoin(epsilon): # epsilon greedy
output = rd.choice(actions)
else:
value = max([Q_values[(state, action)] for action in actions])
output = rd.choice([
action for action in actions if Q_values[(state, action)] == value
])
# TODO, exploration
f.close()
def MCTS(self, curState):
# First, determine the tactics and set the timer
startTime = time.time()
QValues = util.Counter()
Values = util.Counter()
tactic = self.getTactics(curState, self.index)
print(self.index, " ", tactic) # FIXME for debug purpose
# Do the main loop of MCTS
fringe = PriorityQueue()
fringe.push(curState, 0)
tempActions = Deque()
bestActions = Deque()
pathAndReward = util.Stack()
bestReward = float("-inf")
expandedStates = util.Counter()
nextStates = util.Counter()
curDepth = 0
while fringe.isEmpty() is False:
state = fringe.pop()
topPos = state.getAgentPosition(self.index)
if curDepth >= self.depth:
# Backpropagation
cumulativeReward = 0
while pathAndReward.isEmpty() is False:
state, reward = pathAndReward.pop()
cumulativeReward = reward + cumulativeReward * self.discount
Values[state] = cumulativeReward
if cumulativeReward > bestReward:
bestReward = cumulativeReward
bestActions = tempActions
# print(bestActions.list)
# util.pause()
(priority, _, _) = fringe.peekPriority()
depthDiff = curDepth + priority
curDepth = -priority
for _ in range(depthDiff):
tempActions.popBack()
else:
reward = self.getReward(state)
pathAndReward.push((state, reward))
if expandedStates[state] > 0:
# Not only calculate Q(s, a), should consider V(s) for some descendants
expandedStates[state] += 1
actionProb = util.Counter()
for action in nextStates[state]:
nextState = nextStates[state][action]
# If next state is expanded, use V(s)
if expandedStates[nextState] > 0:
actionProb[action] = Values[nextState]
# If next state is not expanded, use Q(s, a)
else:
actionProb[action] = QValues[topPos][action]
# Calculate probability according to Q(s, a) or V(s)
actionProb = softmax(actionProb)
else:
# If the state has not been expanded, expand the state
expandedStates[state] += 1
legalActions = state.getLegalActions(self.index)
actionProb = util.Counter()
for action in legalActions:
# print(self.getQValue(topState, action, tactic), QValues[topPos][action])
if QValues[topPos] == 0:
QValues[topPos] = util.Counter()
QValues[topPos][action] = self.getQValue(
state, action, tactic)[0]
actionProb[action] = QValues[topPos][action]
if nextStates[state] == 0:
nextStates[state] = util.Counter()
nextStates[state][action] = self.getNextState(
state, action)
actionProb = softmax(
actionProb
) # Calculate probability according to Q(s, a)
# Choose action according to action probability
flip = random.random()
cumulative = 0
chosenAction = "Error" # Marking error
if util.flipCoin(self.epsilon):
for prob in actionProb:
if cumulative <= flip <= cumulative + actionProb[prob]:
chosenAction = prob
break
else:
cumulative += actionProb[prob]
else:
chosenAction = actionProb.argMax()
tempActions.push((chosenAction, QValues[topPos][chosenAction]))
nextState = nextStates[state][chosenAction]
# Determine whether to do a back track
if util.flipCoin(1 / exp(.4 * (curDepth + self.bias))):
fringe.push(curState, -curDepth)
curDepth += 1
fringe.push(nextState, -curDepth)
endTime = time.time()
if endTime - startTime > self.timeInterval:
break
self.actionsChosen = bestActions