def __init__(self, real, simulator, expParams, searchParams):
self.Real = real
self.Simulator = simulator
self.ExpParams = expParams
self.SearchParams = searchParams
if (self.ExpParams.AutoExploration):
self.SearchParams.ExplorationConstant = 1
MCTS.InitFastUCB(self.SearchParams.ExplorationConstant)
def __init__(self,real,simulator,outputFile, expParams, searchParams):
self.Real = real
self.Simulator = simulator
self.OutputFile= outputFile
self.ExpParams = expParams
self.SearchParams = searchParams
self.Results = RESULTS.RESULTS()
if (self.ExpParams.AutoExploration):
if (self.SearchParams.UseRave):
self.SearchParams.ExplorationConstant = 0
else:
self.SearchParams.ExplorationConstant = simulator.GetRewardRange()
MCTS.InitFastUCB(self.SearchParams.ExplorationConstant)
def RunOne(self, tries):
mcts = MCTS.MCTS(self.Simulator, self.SearchParams)
for x in range(tries):
self.RunSim(mcts)
state = self.Real.CreateStartState()
ei = state.experimentInstance
steps = 0
while not state.terminal_or_allReach():
steps = steps + 1
action = mcts.GreedyUCB(state, False)
state, terminal, observation, reward = self.Real.RealStep(
state, action)
ei = state.experimentInstance
print "Terminated"
precision, recall = ei.calc_precision_recall()
print "end", repr(ei)
print precision, recall, steps
def RunMultiple(self):
#boost::timer timer
mcts = MCTS.MCTS(self.Simulator, self.SearchParams)
steps = 0
state = self.Real.CreateStartState() #STATE*
terminal = False
while not (state.terminal_or_allReach() or terminal):
action = mcts.SelectAction()
ei = state.experimentInstance
# print "ei-repr" , repr(ei), ei.calc_precision_recall()
# print "selected action" , action ,"HP" , ei.next_tests_by_hp()
state, terminal, observation, reward = self.Real.RealStep(
state, action)
mcts.Update(state)
steps = steps + 1
# print "Terminated" ,state.getMaxProb(), len(state.experimentInstance.initial_tests)
ei = state.experimentInstance
precision, recall = ei.calc_precision_recall()
return precision, recall, steps
def Run(self):
#boost::timer timer
mcts = MCTS.MCTS(self.Simulator, self.SearchParams)
undiscountedReturn = 0.0
discountedReturn = 0.0
discount = 1.0
terminal = False
outOfParticles = False
t = 0
state = self.Real.CreateStartState() #STATE*
for t in range(int(self.ExpParams.NumSteps)):
action = mcts.SelectAction()
# print "action", action, state.experimentInstance.initial_tests, state.getMaxProb(), state.experimentInstance.error
terminal, observation, reward = self.Real.Step(state, action)
# print "state.getMaxProb", state.getMaxProb()
self.Results.Reward.Add(reward)
undiscountedReturn += reward
discountedReturn += reward * discount
discount *= self.Real.GetDiscount()
if (terminal):
print "Terminated", state.getMaxProb(), len(
state.experimentInstance.initial_tests
) #, str(mcts.StatTreeDepth) , str(mcts.StatRolloutDepth) , str(mcts.StatTotalReward)
break
outOfParticles = not mcts.Update(action, observation, reward)
if (outOfParticles):
break
# if (timer.elapsed() > self.ExpParams.TimeOut):
# print "Timed out after ", t , " steps in ", self.results.Time.GetTotal() , "seconds"
# break
if (outOfParticles):
print "Out of particles, finishing episode with SelectRandom"
history = mcts.GetHistory() #HISTORY
while (++t < self.ExpParams.NumSteps):
observation = 0
reward = 0.0
# This passes real state into simulator!
# SelectRandom must only use fully observable state
# to avoid "cheating"
action = self.Simulator.SelectRandom(state, history,
mcts.GetStatus())
terminal, observation, reward = self.Real.Step(state, action)
self.Results.Reward.Add(reward)
undiscountedReturn += reward
discountedReturn += reward * discount
discount *= self.Real.GetDiscount()
if (terminal):
print "Terminated" #, str(mcts.StatTreeDepth) , str(mcts.StatRolloutDepth) , str(mcts.StatTotalReward)
break
history.Add(action, observation)
#self.results.Time.Add(timer.elapsed())
self.Results.Time.Add(-1)
self.Results.UndiscountedReturn.Add(undiscountedReturn)
self.Results.DiscountedReturn.Add(discountedReturn)