def _updateKernelParameters(self, S, A, random=True, normalize=True):
SA = self._getStateActionMatrix(S, A)
if random:
self.MuS = Helper.getRandomSubset(S, self.numFeatures)
self.MuSA = Helper.getRandomSubset(SA, self.numFeatures)
else:
self.MuS = Helper.getRepresentativeRows(S, self.numFeatures, normalize)
self.MuSA = Helper.getRepresentativeRows(SA, self.numFeatures, normalize)
NUM_SAMPLES_FOR_BW_ESTIMATE = 500
# bandwidth for PHI_S
bwNonKbS = Helper.getBandwidth(self.MuS[:, 0:self.NUM_NON_KB_DIM],
NUM_SAMPLES_FOR_BW_ESTIMATE, self.bwFactorNonKbS)
kbPosS = self._reshapeKbPositions(self.MuS[:, self.NUM_NON_KB_DIM:])
bwKbS = Helper.getBandwidth(kbPosS, NUM_SAMPLES_FOR_BW_ESTIMATE,
self.bwFactorKbS)
self.kernelS.setBandwidth(bwNonKbS, bwKbS)
self.kernelS.setWeighting(self.weightNonKbS)
# bandwidth for PHI_SA
bwNonKbSA = Helper.getBandwidth(self.MuSA[:, 0:(self.NUM_NON_KB_DIM + 2)],
NUM_SAMPLES_FOR_BW_ESTIMATE, self.bwFactorNonKbSA)
kbPosSA = self._reshapeKbPositions(self.MuSA[:, (self.NUM_NON_KB_DIM + 2):])
bwKbSA = Helper.getBandwidth(kbPosSA, NUM_SAMPLES_FOR_BW_ESTIMATE,
self.bwFactorKbSA)
self.kernelSA.setBandwidth(bwNonKbSA, bwKbSA)
self.kernelSA.setWeighting(self.weightNonKbSA)
def _getSubsetForGP(self, S, random=True, normalize=True):
Nsubset = min(self.numSamplesSubsetGP, S.shape[0])
if random:
return Helper.getRandomSubset(S, Nsubset)
else:
return Helper.getRepresentativeRows(S, Nsubset, normalize)
