def testGDelta(self):
# usually, Gdelta==G
GP = GaussianProcess(GaussianKernel_iso([0.05]))
X = lhcSample([[0., 1.]], 5, seed=10)
Y = [x**2 for x in X]
GP.train(X, Y)
G = (Y[0] - max(Y)) / (Y[0] - 1)
self.failUnlessEqual(G, Gdelta(GP, [[0., 1.]], Y[0], 1.0, 0.01))
# sometimes, though, Gdelta > G -- this GP has a very high confidence
# prediction of a very good point at x ~ .65
GP = GaussianProcess(GaussianKernel_iso([0.1]))
X = array([[.5], [.51], [.59], [.6]])
Y = array([1., 2., 2., 1.])
GP.train(X, Y)
# figure(1)
# A = arange(0, 1, 0.01)
# post = [GP.posterior(x) for x in A]
# plot(A, [p[0] for p in post], 'k-')
# plot(A, [p[0]+p[1] for p in post], 'k:')
# show()
G = (Y[0] - max(Y)) / (Y[0] - 4.0)
Gd = Gdelta(GP, [[0., 1.]], Y[0], 4.0, 0.01)
self.failUnless(G < Gd)
# however, if there is more variance, we will collapse back to G
GP = GaussianProcess(GaussianKernel_iso([.001]))
GP.train(X, Y)
G = (Y[0] - max(Y)) / (Y[0] - 4.0)
self.failUnlessEqual(G, Gdelta(GP, [[0., 1.]], Y[0], 4.0, 0.01))
def testGDelta(self):
# usually, Gdelta==G
GP = GaussianProcess(GaussianKernel_iso([0.05]))
X = lhcSample([[0., 1.]], 5, seed=10)
Y = [x**2 for x in X]
GP.train(X, Y)
G = (Y[0]-max(Y)) / (Y[0]-1)
self.failUnlessEqual(G, Gdelta(GP, [[0.,1.]], Y[0], 1.0, 0.01))
# sometimes, though, Gdelta > G -- this GP has a very high confidence
# prediction of a very good point at x ~ .65
GP = GaussianProcess(GaussianKernel_iso([0.1]))
X = array([[.5], [.51], [.59], [.6]])
Y = array([1., 2., 2., 1.])
GP.train(X, Y)
# figure(1)
# A = arange(0, 1, 0.01)
# post = [GP.posterior(x) for x in A]
# plot(A, [p[0] for p in post], 'k-')
# plot(A, [p[0]+p[1] for p in post], 'k:')
# show()
G = (Y[0]-max(Y)) / (Y[0]-4.0)
Gd = Gdelta(GP, [[0., 1.]], Y[0], 4.0, 0.01)
self.failUnless(G < Gd)
# however, if there is more variance, we will collapse back to G
GP = GaussianProcess(GaussianKernel_iso([.001]))
GP.train(X, Y)
G = (Y[0]-max(Y)) / (Y[0]-4.0)
self.failUnlessEqual(G, Gdelta(GP, [[0., 1.]], Y[0], 4.0, 0.01))
