def runExp(gamma=0, epsilon=0.1, xor=False, lr = 0.02):
if xor:
print "Attempting the XOR task"
else:
print "Attempting the AND task"
task = XORTask()
task.and_task = not xor
l = Q_LinFA(task.nactions, task.nsenses)
l.rewardDiscount = gamma
l.learningRate = lr
agent = LinFA_QAgent(l)
agent.epsilon = epsilon
exp = Experiment(task, agent)
sofar = 0
for i in range(30):
exp.doInteractions(100)
print exp.task.cumreward - sofar,
if i%10 == 9:
print
sofar = exp.task.cumreward
l._decayLearningRate()
def __init__(self, task, agent):
if isinstance(agent, OptimizationAgent):
self.doOptimization = True
self.optimizer = agent.learner
self.optimizer.setEvaluator(task, agent.module)
self.optimizer.maxEvaluations = self.optimizer.numEvaluations
else:
Experiment.__init__(self, task, agent)
def __init__(self, task, agent):
if isinstance(agent, OptimizationAgent):
self.doOptimization = True
self.optimizer = agent.learner
self.optimizer.setEvaluator(task, agent.module)
self.optimizer.maxEvaluations = self.optimizer.numEvaluations
else:
Experiment.__init__(self, task, agent)
def _oneInteraction(self):
""" Do an interaction between the Task and the Agent. """
if self.doOptimization:
raise Exception(
'When using a black-box learning algorithm, only full episodes can be done.'
)
else:
return Experiment._oneInteraction(self)
def _oneInteraction(self):
""" Do an interaction between the Task and the Agent. """
if self.doOptimization:
raise Exception('When using a black-box learning algorithm, only full episodes can be done.')
else:
return Experiment._oneInteraction(self)
def __init__(self, task, multiAgent):
assert isinstance(
task,
EpisodicTaskSG), "task should be the subclass of EpisodicTaskSG."
assert isinstance(multiAgent, MultiAgent), "task should be MultAgent."
Experiment.__init__(self, task, multiAgent)
def __init__(self, task, multiAgent):
assert isinstance(task, EpisodicTaskSG), "task should be the subclass of EpisodicTaskSG."
assert isinstance(multiAgent, MultiAgent), "task should be MultAgent."
Experiment.__init__(self, task, multiAgent)
