def VOI(x,a=aVec,grad=False):
bVec=bfunc(x)
a,b,keep=AffineBreakPointsPrep(a,bVec)
keep1,c=AffineBreakPoints(a,b)
keep1=keep1.astype(np.int64)
M=len(keep1)
keep2=keep[keep1]
return hvoi(b,c,keep1)+a_0(x)
def VOI(x,a2=aVec,grad=False):
lsum=np.sum(x)
newIndexes=[(i,j) for (i,j) in possiblePoints if np.sum(j)==totalOpenings-lsum]
bVec=bfunc(x,[j for (i,j) in newIndexes])
bVec=bVec+np.sqrt(sigma_0(x,x))
a2=np.array([a2[i] for (i,j) in newIndexes])
a2=a2+a_0(x)
a,b,keep=AffineBreakPointsPrep(a2,bVec)
keep1,c=AffineBreakPoints(a,b)
keep1=keep1.astype(np.int64)
M=len(keep1)
keep2=keep[keep1]
return hvoi(b,c,keep1)+a_0(x)
def VOIfunc(self, n, pointNew, L, data, kern, temp1, temp2, grad, a, onlyGrad=False):
n1 = self.n1
tempN = n + self._numberTraining
b, temp5, inner, tempB = self.aANDb(n, self._points, pointNew, L, data, kern, temp1, temp2)
a, b, keep = AffineBreakPointsPrep(a, b)
keep1, c = AffineBreakPoints(a, b)
keep1 = keep1.astype(np.int64)
M = len(keep1)
keep2 = keep[keep1]
if grad:
B2temp = np.zeros((M, tempN))
inv1temp = np.zeros((M, tempN))
for j in xrange(M):
inv1temp[j, :] = temp2[keep2[j], :]
if onlyGrad:
return self.evalVOI(
n, pointNew, a, b, c, keep, keep1, M, L, data.Xhist, kern, tempB, temp5, inner, inv1temp, grad, onlyGrad
)
if grad == False:
return self.evalVOI(n, pointNew, a, b, c, keep, keep1, M, L, data.Xhist, kern, tempB, grad=False)
return self.evalVOI(
n, pointNew, a, b, c, keep, keep1, M, L, data.Xhist, kern, tempB, temp5, inner, inv1temp, grad
)
def VOIfunc(self, n, pointNew, grad, L, temp2, a, scratch, kern, XW, B, onlyGradient=False):
"""
Output:
Evaluates the VOI and it can compute its derivative. It evaluates
the VOI, when grad and onlyGradient are False; it evaluates the
VOI and computes its derivative when grad is True and onlyGradient
is False, and computes only its gradient when gradient and
onlyGradient are both True.
Args:
-n: Iteration of the algorithm.
-pointNew: The VOI will be evaluated at this point.
-grad: True if we want to compute the gradient; False otherwise.
-L: Cholesky decomposition of the matrix A, where A is the covariance
matrix of the past obsevations (x,w).
-temp2: temp2=inv(L)*B.T, where B is a matrix such that B(i,j) is
\int\Sigma_{0}(x_{i},w,x_{j},w_{j})dp(w)
where points x_{p} is a point of the discretization of
the space of x; and (x_{j},w_{j}) is a past observation.
-a: Vector of the means of the GP on g(x)=E(f(x,w,z)).
The means are evaluated on the discretization of
the space of x.
-scratch: Matrix where scratch[i,:] is the solution of the linear system
Ly=B[j,:].transpose() (See above for the definition of B and L)
-kern: Kernel.
-XW: Past observations.
-B: Computes B(x,XW)=\int\Sigma_{0}(x,w,XW[0:n1],XW[n1:n1+n2])dp(w).
Its arguments are:
-x: Vector of points where B is evaluated
-XW: Point (x,w)
-n1: Dimension of x
-n2: Dimension of w
-onlyGradient: True if we only want to compute the gradient; False otherwise.
"""
n1 = self.n1
b, gamma, BN, temp1, aux4 = self.aANDb(
n,
self._points,
pointNew[0, 0:n1],
pointNew[0, n1 : n1 + self.n2],
L,
temp2=temp2,
past=XW,
kernel=kern,
B=B,
)
a, b, keep = AffineBreakPointsPrep(a, b)
keep1, c = AffineBreakPoints(a, b)
keep1 = keep1.astype(np.int64)
M = len(keep1)
nTraining = self._numberTraining
tempN = nTraining + n
keep2 = keep[keep1]
if grad:
scratch1 = np.zeros((M, tempN))
for j in xrange(M):
scratch1[j, :] = scratch[keep2[j], :]
if onlyGradient:
return self.evalVOI(
n,
pointNew,
a,
b,
c,
keep,
keep1,
M,
gamma,
BN,
L,
scratch=scratch1,
inv=temp1,
aux4=aux4,
grad=True,
onlyGradient=onlyGradient,
kern=kern,
XW=XW,
)
if grad == False:
return self.evalVOI(
n, pointNew, a, b, c, keep, keep1, M, gamma, BN, L, aux4=aux4, inv=temp1, kern=kern, XW=XW
)
return self.evalVOI(
n,
pointNew,
a,
b,
c,
keep,
keep1,
M,
gamma,
BN,
L,
aux4=aux4,
inv=temp1,
scratch=scratch1,
grad=True,
kern=kern,
XW=XW,
)
