def __init__(self, module, learner = None):
StateDependentAgent.__init__(self, module, learner)
# gaussian process
self.gp = GaussianProcess(self.explorationlayer.module.paramdim, -2, 2, 1)
self.gp.mean = -1.5
self.gp.hyper = (2.0, 2.0, 0.1)
def __init__(self, task, agent):
EpisodicExperiment.__init__(self, task, agent)
# create model and training set (action dimension + 1 for time)
self.modelds = SequentialDataSet(self.task.indim + 1, 1)
self.model = [
GaussianProcess(indim=self.modelds.getDimension('input'),
start=(-10, -10, 0),
stop=(10, 10, 300),
step=(5, 5, 100)) for _ in range(self.task.outdim)
]
# change hyper parameters for all gps
for m in self.model:
m.hyper = (20, 2.0, 0.01)
#!/usr/bin/env python """ A simple example on how to use the GaussianProcess class in pybrain, for one and two dimensions. """ __author__ = "Thomas Rueckstiess, [email protected]" from pybrain.auxiliary import GaussianProcess from pybrain.datasets import SupervisedDataSet from scipy import mgrid, sin, cos, array, ravel from pylab import show, figure ds = SupervisedDataSet(1, 1) gp = GaussianProcess(indim=1, start=-3, stop=3, step=0.05) figure() x = mgrid[-3:3:0.2] y = 0.1*x**2 + x + 1 z = sin(x) + 0.5*cos(y) ds.addSample(-2.5, -1) ds.addSample(-1.0, 3) gp.mean = 0 # new feature "autonoise" adds uncertainty to data depending on # it's distance to other points in the dataset. not tested much yet. # gp.autonoise = True gp.trainOnDataset(ds) gp.plotCurves(showSamples=True)
