class KMeansAgent(Agent):
"""
K-Means agent is built on the same idea that we can extract features
from K-means in the image recognition.
Unfortunately, this agent cannot get better as the euclidean distance
between states is meaningless !
"""
def __init__(self):
self.learn = learn
self.load = load
self.learner = Learner()
if self.load and os.path.isfile(learnerFile): self.learner.loadData(learnerFile)
self.learner.newGame()
def getAction(self, gameState):
"""
The selected action is simply the best action computed by the
(not that) smart evaluationFunction.
"""
# Before trying to find an action, we save the gameState
#print flattenGameState(gameState)
self.learner.addState(flattenGameState(gameState))
# Collect legal moves and successor states
legalMoves = gameState.getLegalActions()
legalMoves.remove('Stop')
# Choose one of the best actions
scores = [self.notThatSmartEvaluationFunction(gameState, action) for action in legalMoves]
#print scores
bestScore = max(scores)
bestIndices = [index for index in range(len(scores)) if scores[index] == bestScore]
chosenIndex = random.choice(bestIndices) # Pick randomly among the best
return legalMoves[chosenIndex]
def final(self,finalState):
self.learner.labelizeData(finalState)
if self.learner.gamesInMemory % processingFrequency == 0:
self.learner.processLearning()
if saveFile:
print 'Saving learner file...'
self.learner.saveData(learnerFile)
#self.learner.saveData(learnerFile)
self.learner.newGame()
#flattenGameState(finalState)
def notThatSmartEvaluationFunction(self, currGameState, pacManAction):
nextGameState = currGameState.generatePacmanSuccessor(pacManAction)
if self.learner.readyToPredict == False:
return nextGameState.getScore()
else:
nextGameState = currGameState.generatePacmanSuccessor(pacManAction)
stateFeatures = self.learner.extractFeatures(flattenGameState(nextGameState))
predictedScore = self.learner.predictScore(stateFeatures)
#print predictedScore
return float(predictedScore)
class KMeansAgent(Agent):
"""
K-Means agent is built on the same idea that we can extract features
from K-means in the image recognition.
Unfortunately, this agent cannot get better as the euclidean distance
between states is meaningless !
"""
def __init__(self):
self.learn = learn
self.load = load
self.learner = Learner()
if self.load and os.path.isfile(learnerFile):
self.learner.loadData(learnerFile)
self.learner.newGame()
def getAction(self, gameState):
"""
The selected action is simply the best action computed by the
(not that) smart evaluationFunction.
"""
# Before trying to find an action, we save the gameState
#print flattenGameState(gameState)
self.learner.addState(flattenGameState(gameState))
# Collect legal moves and successor states
legalMoves = gameState.getLegalActions()
legalMoves.remove('Stop')
# Choose one of the best actions
scores = [
self.notThatSmartEvaluationFunction(gameState, action)
for action in legalMoves
]
#print scores
bestScore = max(scores)
bestIndices = [
index for index in range(len(scores)) if scores[index] == bestScore
]
chosenIndex = random.choice(
bestIndices) # Pick randomly among the best
return legalMoves[chosenIndex]
def final(self, finalState):
self.learner.labelizeData(finalState)
if self.learner.gamesInMemory % processingFrequency == 0:
self.learner.processLearning()
if saveFile:
print 'Saving learner file...'
self.learner.saveData(learnerFile)
#self.learner.saveData(learnerFile)
self.learner.newGame()
#flattenGameState(finalState)
def notThatSmartEvaluationFunction(self, currGameState, pacManAction):
nextGameState = currGameState.generatePacmanSuccessor(pacManAction)
if self.learner.readyToPredict == False:
return nextGameState.getScore()
else:
nextGameState = currGameState.generatePacmanSuccessor(pacManAction)
stateFeatures = self.learner.extractFeatures(
flattenGameState(nextGameState))
predictedScore = self.learner.predictScore(stateFeatures)
#print predictedScore
return float(predictedScore)
