def distmat(X, Y):
n = len(X)
m = len(Y)
xx = ml.sum(X*X, axis=1)
yy = ml.sum(Y*Y, axis=1)
xy = ml.dot(X, Y.T)
return npy.tile(xx, (m, 1)).T+npy.tile(yy, (n, 1)) - 2*xy
def zl_distvec(X, Y):
n = len(X)
m = len(Y)
xx = ml.sum(X*X, axis=1)
yy = ml.sum(Y*Y, axis=1)
xy = ml.dot(X, Y.T)
return tile(xx, (m, 1)).T+tile(yy, (n, 1)) - 2*xy
def nk_bhatta(X1, X2, eta):
# Returns the non-kernelized Bhattacharrya
#I.e. fits normal distributions in input space and calculates Bhattacharrya overlap between them
(n1, d1) = X1.shape
(n2, d) = X2.shape
assert d1 == d
mu1 = mat.sum(X1, 0) / n1
mu2 = mat.sum(X2, 0) / n2
X1c = X1 - mat.tile(mu1, (n1, 1))
X2c = X2 - mat.tile(mu2, (n2, 1))
Eta = mat.eye(d) * eta
S1 = X1c.T * X1c / n1 + Eta
S2 = X2c.T * X2c / n2 + Eta
mu3 = .5 * (S1.I * mu1.T + S2.I * mu2.T).T
S3 = 2 * (S1.I + S2.I).I
d1 = la.det(S1)**-.25
d2 = la.det(S2)**-.25
d3 = la.det(S3)**.5
dterm = d1 * d2 * d3
e1 = -.25 * mu1 * S1.I * mu1.T
e2 = -.25 * mu2 * S2.I * mu2.T
e3 = .5 * mu3 * S3 * mu3.T
eterm = math.exp(e1 + e2 + e3)
return float(dterm * eterm)
def nk_bhatta(X1, X2, eta):
# Returns the non-kernelized Bhattacharrya
#I.e. fits normal distributions in input space and calculates Bhattacharrya overlap between them
(n1, d1) = X1.shape
(n2, d ) = X2.shape
assert d1 == d
mu1 = mat.sum(X1,0) / n1
mu2 = mat.sum(X2,0) / n2
X1c = X1 - mat.tile(mu1, (n1,1))
X2c = X2 - mat.tile(mu2, (n2,1))
Eta = mat.eye(d) * eta
S1 = X1c.T * X1c / n1 + Eta
S2 = X2c.T * X2c / n2 + Eta
mu3 = .5 * (S1.I * mu1.T + S2.I * mu2.T).T
S3 = 2 * (S1.I + S2.I).I
d1 = la.det(S1) ** -.25
d2 = la.det(S2) ** -.25
d3 = la.det(S3) ** .5
dterm = d1 * d2 * d3
e1 = -.25 * mu1 * S1.I * mu1.T
e2 = -.25 * mu2 * S2.I * mu2.T
e3 = .5 * mu3 * S3 * mu3.T
eterm = math.exp(e1 + e2 + e3)
return float(dterm * eterm)
def zl_distvec(X, Y):
n = len(X)
m = len(Y)
xx = ml.sum(X * X, axis=1)
yy = ml.sum(Y * Y, axis=1)
xy = ml.dot(X, Y.T)
return tile(xx, (m, 1)).T + tile(yy, (n, 1)) - 2 * xy
def kernel_matrix(X, kernel, n1, n2):
(n, d) = X.shape
assert n == n1 + n2
K = mat.zeros((n,n))
for i in xrange(n):
for j in xrange(i+1):
K[i,j] = kernel(X[i,:], X[j,:])
K[j,i] = K[i,j]
U1 = mat.sum(K[0:n1,:],0) / n1
U2 = mat.sum(K[n1:n,:],0) / n2
U1m = mat.tile(U1, (n1,1))
U2m = mat.tile(U2, (n2,1))
U = mat.bmat('U1m; U2m')
m1m1 = mat.sum(K[0:n1, 0:n1]) / (n1*n1)
m1m2 = mat.sum(K[0:n1, n1:n]) / (n1*n2)
m2m2 = mat.sum(K[n1:n, n1:n]) / (n2*n2)
mumu = mat.zeros((n,n))
mumu[0:n1, 0:n1] = m1m1
mumu[0:n1, n1:n] = m1m2
mumu[n1:n, 0:n1] = m1m2
mumu[n1:n, n1:n] = m2m2
Kcu = K - U
Kuc = Kcu.T
N = mat.ones((n,n))/n
Kc = K - U - U.T + mumu
return (K, Kuc, Kc)
def distance(X, Y):
n = len(X)
m = len(Y)
xx = matlib.sum(X*X, axis=1)
yy = matlib.sum(Y*Y, axis=1)
xy = matlib.dot(X, Y.T)
return tile(xx, (m, 1)).T+tile(yy, (n, 1)) - 2*xy
def kernel_matrix(X, kernel, n1, n2):
(n, d) = X.shape
assert n == n1 + n2
K = mat.zeros((n, n))
for i in xrange(n):
for j in xrange(i + 1):
K[i, j] = kernel(X[i, :], X[j, :])
K[j, i] = K[i, j]
U1 = mat.sum(K[0:n1, :], 0) / n1
U2 = mat.sum(K[n1:n, :], 0) / n2
U1m = mat.tile(U1, (n1, 1))
U2m = mat.tile(U2, (n2, 1))
U = mat.bmat('U1m; U2m')
m1m1 = mat.sum(K[0:n1, 0:n1]) / (n1 * n1)
m1m2 = mat.sum(K[0:n1, n1:n]) / (n1 * n2)
m2m2 = mat.sum(K[n1:n, n1:n]) / (n2 * n2)
mumu = mat.zeros((n, n))
mumu[0:n1, 0:n1] = m1m1
mumu[0:n1, n1:n] = m1m2
mumu[n1:n, 0:n1] = m1m2
mumu[n1:n, n1:n] = m2m2
Kcu = K - U
Kuc = Kcu.T
N = mat.ones((n, n)) / n
Kc = K - U - U.T + mumu
return (K, Kuc, Kc)
def distMat(X, Y):
n = len(X)
m = len(Y)
xx = ml.sum(X * X, axis=1)
yy = ml.sum(Y * Y, axis=1)
xy = ml.dot(X, Y.T)
return np.tile(xx, (m, 1)).T + np.tile(yy, (n, 1)) - 2 * xy
def distmat(X,Y):
n = len(X)
m = len(Y)
xx = ml.sum(X*X,axis=1) # #axis=1 是按行求和
print "xx:{}".format(xx)
yy = ml.sum(Y*Y,axis=1)
print "yy:{}".format(yy)
xy = ml.dot(X,Y.T) #dot矩阵相乘
return tile(xx,(m,1)).T + tile(yy,(n,1)) - 2*xy #tile矩阵复制
def prepare_bhatta(X1, X2, kernel, eta, verbose=False):
(n1, d1) = X1.shape
(n2, d) = X2.shape
assert d1 == d
n = n1 + n2
X = mat.bmat('X1;X2')
(K, Kuc, Kc) = kernel_matrix(X, kernel, n1, n2)
G = GS_basis(Kc, verbose)
(null, g_dim) = G.shape
mu1 = mat.sum(Kuc[0:n1, :] * G, 0) / n1
mu2 = mat.sum(Kuc[n1:n, :] * G, 0) / n2
return (Kc, G, mu1, mu2)
def prepare_bhatta(X1, X2, kernel, eta, verbose=False):
(n1, d1) = X1.shape
(n2, d ) = X2.shape
assert d1 == d
n = n1 + n2
X = mat.bmat('X1;X2')
(K, Kuc, Kc) = kernel_matrix(X, kernel, n1, n2)
G = GS_basis(Kc, verbose)
(null, g_dim) = G.shape
mu1 = mat.sum(Kuc[0:n1,:] * G,0) / n1
mu2 = mat.sum(Kuc[n1:n,:] * G,0) / n2
return (Kc, G, mu1, mu2)
def kernel_submatrix(self):
# Cache kernel evaluations between vectors in this dataset so we don't repeat this work every
# time we call Bhattacharrya
X = self.X
(n,d) = X.shape
K = mat.zeros((n,n))
for i in xrange(n):
for j in xrange(i+1):
K[i,j] = self.kernel(X[i,:], X[j,:])
K[j,i] = K[i,j]
Ki = mat.sum(K,1) / n
k = mat.sum(K) / (n*n)
Kc = K - Ki * mat.ones((1,n)) - mat.ones((n,1)) * Ki.T + k * mat.ones((n,n))
return (K, Kc)
def compute_merit(B, U, model, penalty_coeff):
xDim = model.xDim
T = model.T
merit = 0
for t in xrange(0,T-1):
x, s = belief.decompose_belief(B[:,t], model)
merit += model.alpha_belief*ml.trace(s*s)
merit += model.alpha_control*ml.sum(U[:,t].T*U[:,t])
merit += penalty_coeff*ml.sum(np.abs(B[:,t+1]-belief.belief_dynamics(B[:,t],U[:,t],None,model)))
x, s = belief.decompose_belief(B[:,T-1], model)
merit += model.alpha_final_belief*ml.trace(s*s)
return merit
def kernel_submatrix(self):
# Cache kernel evaluations between vectors in this dataset so we don't repeat this work every
# time we call Bhattacharrya
X = self.X
(n, d) = X.shape
K = mat.zeros((n, n))
for i in xrange(n):
for j in xrange(i + 1):
K[i, j] = self.kernel(X[i, :], X[j, :])
K[j, i] = K[i, j]
Ki = mat.sum(K, 1) / n
k = mat.sum(K) / (n * n)
Kc = K - Ki * mat.ones((1, n)) - mat.ones(
(n, 1)) * Ki.T + k * mat.ones((n, n))
return (K, Kc)
def compute_probability(S, K, model):
T = model.T
h = 10
Smin = np.zeros(T)
Smax = np.zeros(T)
P = ml.zeros([T, K])
for i in range(T):
Smin[i] = np.min(S[i, :])
Smax[i] = np.max(S[i, :])
#IPython.embed()
if (Smin[i] == Smax[i]):
Smax[i] = 1
for i in range(T):
P[i, :] = np.exp(-h * (S[i, :] - Smin[i]) / (Smax[i] - Smin[i]))
normalize = ml.sum(P[i, :])
# if(normalize == K):
# P[i,:] = 0.005;
P[i, :] = P[i, :] / normalize
return P
def cost_func(B, model, profile, profiler):
cost = ml.zeros([model.T, 1])
U = ml.zeros([model.uDim, model.T])
T = model.T
for t in range(T - 1):
U[:, t] = B[0:model.xDim, t + 1] - B[0:model.xDim, t]
B[:, t + 1] = belief.belief_dynamics(B[:, t], U[:, t], None, model)
if max(U[:, t]) > 1:
cost[t] = 1000
elif abs(B[0, t]) > model.xMax[0]:
cost[t] = 1000
elif abs(B[1, t]) > model.xMax[1]:
cost[t] = 1000
else:
null, s = belief.decompose_belief(B[:, t], model)
cost[t] = model.alpha_belief * ml.trace(
s * s) + model.alpha_control * ml.sum(U[:, t].T * U[:, t])
x, s = belief.decompose_belief(B[:, T - 1], model)
cost[T - 1] = model.alpha_final_belief * ml.trace(s * s)
return cost, B, U
def compute_probability(S,K,model):
T = model.T;
h = 10;
Smin = np.zeros(T)
Smax = np.zeros(T)
P = ml.zeros([T,K])
for i in range(T):
Smin[i] = np.min(S[i,:]);
Smax[i] = np.max(S[i,:]);
#IPython.embed()
if(Smin[i] == Smax[i]):
Smax[i] = 1;
for i in range(T):
P[i,:] = np.exp(-h*(S[i,:] - Smin[i])/(Smax[i]-Smin[i]));
normalize = ml.sum(P[i,:]);
# if(normalize == K):
# P[i,:] = 0.005;
P[i,:] = P[i,:]/normalize;
return P
def inf(self, x, meanonly=False):
x = np.asmatrix(x)
if x.shape[1] != self.d:
if x.shape[0] == self.d:
x = x.T
else:
raise Exception('Invalid test-set dimension -- '
'expected d = ' + str(self.d) + '.')
n = x.shape[0]
# Handle empty test set
if n == 0:
return (np.zeros((0, 1)), np.zeros((0, 1)))
ms = self.kernel.mean*np.ones((n, 1))
Kbb = self.kernel(x, diag=True)
# Handle empty training set
if len(self) == 0:
return (ms, np.asmatrix(Kbb))
Kba = self.kernel(x, self.x)
m = self.kernel.mean*np.ones((len(self), 1))
fm = ms + Kba*scipy.linalg.cho_solve((self.L, True), self.y - m,
overwrite_b=True)
if meanonly:
return fm
else:
W = scipy.linalg.cho_solve((self.L, True), Kba.T)
fv = np.asmatrix(Kbb - np.sum(np.multiply(Kba.T, W), axis=0).T)
# W = np.asmatrix(scipy.linalg.solve(self.L, Kba.T, lower=True))
# fv = np.asmatrix(Kbb - np.sum(np.power(W, 2), axis=0).T)
return (fm, fv)
def cost_func(B,model,profile,profiler):
cost = ml.zeros([model.T,1])
U = ml.zeros([model.uDim,model.T])
T = model.T
for t in range(T-1):
U[:,t] = B[0:model.xDim,t+1]-B[0:model.xDim,t];
B[:,t+1] = belief.belief_dynamics(B[:,t],U[:,t],None,model);
if max(U[:,t])> 1:
cost[t] = 1000
elif abs(B[0,t]) > model.xMax[0]:
cost[t] = 1000
elif abs(B[1,t]) > model.xMax[1]:
cost[t] = 1000
else:
null, s = belief.decompose_belief(B[:,t], model)
cost[t] = model.alpha_belief*ml.trace(s*s)+model.alpha_control*ml.sum(U[:,t].T*U[:,t])
x, s = belief.decompose_belief(B[:,T-1], model)
cost[T-1] = model.alpha_final_belief*ml.trace(s*s)
return cost, B, U
def _kmedoids(distmat, threshold, imedoids, verbose):
"""\
The *raw* version of k-medoids.
"""
# initialize J
Jprev = inf
# initialize iteration count
iter = 0
# iterations
while True:
# distance from medoids to all other points
dist = distmat[imedoids]
# assign x to nearst medoid
labels = dist.argmin(axis=0)
J = 0
# re-choose each medoids
for j in range(len(imedoids)):
idx_j = (labels == j).nonzero()[0]
distj = distmat[idx_j][:, idx_j]
distsum = ml.sum(distj, axis=1)
idxmin = distsum.argmin()
imedoids[j] = idx_j[idxmin]
J += distsum[idxmin]
iter += 1
if verbose:
print '[kmedoids] iter %d (J=%.4f)' % (iter, J)
if Jprev-J < threshold:
break
Jprev = J
return imedoids, labels, J
def bhatta(self,i,j):
"""Here is the actual Bhattacharrya algorithm"""
eta = self.eta
D1 = self.datasets[i]
D2 = self.datasets[j]
Beta1 = D1.Beta
Beta2 = D2.Beta
(n1, r) = Beta1.shape
(n2, r) = Beta2.shape
n = n1 + n2
Beta = mat.zeros((n,2*r))
Beta[0:n1,0:r] = Beta1
Beta[n1:n,r:2*r] = Beta2
(K, Kuc, Kc) = self.kernel_supermatrix(i,j)
# K = uncentered kernel matrix
# Kuc = Matrix between centered and uncentered vectors
# Kc = Centered kernel matrix
Omega = eig_ortho(Kc, Beta)
mu1 = mat.sum(Kuc[0:n1, :] * Omega, 0) / n1
mu2 = mat.sum(Kuc[n1:n, :] * Omega, 0) / n2
S1 = Omega.T * Kc[:,0:n1] * Kc[0:n1,:] * Omega / n1
S2 = Omega.T * Kc[:,n1:n] * Kc[n1:n,:] * Omega / n2
Eta = eta * mat.eye(2*r)
S1 += Eta
S2 += Eta
mu3 = .5 * (S1.I * mu1.T + S2.I * mu2.T).T
S3 = 2 * (S1.I + S2.I).I
d1 = la.det(S1) ** -.25
d2 = la.det(S2) ** -.25
d3 = la.det(S3) ** .5
e1 = exp(-mu1 * S1.I * mu1.T / 4)
e2 = exp(-mu2 * S2.I * mu2.T / 4)
e3 = exp(mu3 * S3 * mu3.T / 2)
dterm = d1 * d2 * d3
eterm = e1 * e2 * e3
rval = dterm * eterm
if math.isnan(rval):
rval = -1
print "Warning: Kernel failed on datasets ({},{})".format(i,j)
return rval
def bhatta(self, i, j):
"""Here is the actual Bhattacharrya algorithm"""
eta = self.eta
D1 = self.datasets[i]
D2 = self.datasets[j]
Beta1 = D1.Beta
Beta2 = D2.Beta
(n1, r) = Beta1.shape
(n2, r) = Beta2.shape
n = n1 + n2
Beta = mat.zeros((n, 2 * r))
Beta[0:n1, 0:r] = Beta1
Beta[n1:n, r:2 * r] = Beta2
(K, Kuc, Kc) = self.kernel_supermatrix(i, j)
# K = uncentered kernel matrix
# Kuc = Matrix between centered and uncentered vectors
# Kc = Centered kernel matrix
Omega = eig_ortho(Kc, Beta)
mu1 = mat.sum(Kuc[0:n1, :] * Omega, 0) / n1
mu2 = mat.sum(Kuc[n1:n, :] * Omega, 0) / n2
S1 = Omega.T * Kc[:, 0:n1] * Kc[0:n1, :] * Omega / n1
S2 = Omega.T * Kc[:, n1:n] * Kc[n1:n, :] * Omega / n2
Eta = eta * mat.eye(2 * r)
S1 += Eta
S2 += Eta
mu3 = .5 * (S1.I * mu1.T + S2.I * mu2.T).T
S3 = 2 * (S1.I + S2.I).I
d1 = la.det(S1)**-.25
d2 = la.det(S2)**-.25
d3 = la.det(S3)**.5
e1 = exp(-mu1 * S1.I * mu1.T / 4)
e2 = exp(-mu2 * S2.I * mu2.T / 4)
e3 = exp(mu3 * S3 * mu3.T / 2)
dterm = d1 * d2 * d3
eterm = e1 * e2 * e3
rval = dterm * eterm
if math.isnan(rval):
rval = -1
print "Warning: Kernel failed on datasets ({},{})".format(i, j)
return rval
def kmedoids(X, k, observer=None, threshold=1e-15, maxiter=300):
'''
Cluster and show data X
X: N * 2 array, data to be clustered
k: k means
observer: the plot function
'''
N = len(X)
labels = zeros(N, dtype=int)
centers = array(random.sample(X, k)) # Choose k unique elements from X
iter = 0
def calc_J():
sum = 0
for i in xrange(N):
sum += norm(X[i]-centers[labels[i]])
return sum
def distmat(X, Y):
n = len(X)
m = len(Y)
xx = ml.sum(X*X, axis=1)
yy = ml.sum(Y*Y, axis=1)
xy = ml.dot(X, Y.T)
return tile(xx, (m, 1)).T+tile(yy, (n, 1)) - 2*xy
Jprev = calc_J()
while True:
# notify the observer
if observer is not None:
observer(iter, labels, centers)
# calculate distance from x to each center
# distance_matrix is only available in scipy newer than 0.7
# dist = distance_matrix(X, centers)
dist = distmat(X, centers)
# assign x to nearst center
labels = dist.argmin(axis=1)
# re-calculate each center
for j in range(k):
idx_j = (labels == j).nonzero()
distj = distmat(X[idx_j], X[idx_j])
distsum = ml.sum(distj, axis=1)
icenter = distsum.argmin()
centers[j] = X[idx_j[0][icenter]]
J = calc_J()
iter += 1
if Jprev-J < threshold:
break
Jprev = J
if iter >= maxiter:
break
# final notification
if observer is not None:
observer(iter, labels, centers)
def __init__(self, points, masses):
"""
Inputs:
points: list of 3 dimensional points
masses: list of float (same length as points)
"""
self.nb_points = len(points)
if self.nb_points == 0:
raise (Exception, "Provide at least one point")
if len(masses) != self.nb_points:
raise (Exception,
"Same number of points and masses must be provided")
self.masses = masses
self.initial_points = [Vect(p) for p in points]
# Center of mass p and its velocity
self.center_position = npmat.sum(self.initial_points,
axis=0) / self.nb_points
self.center_velocity = npmat.zeros((3, 1))
# Position of the points relative to the center of mass, stacked into a big matrix
self.centered_initial_points_hstack = npmat.hstack(
self.initial_points) - self.center_position
# Total mass and its inverse
self.total_mass = npmat.sum(self.masses)
self.total_mass_inverse = 1. / self.total_mass
# Inertia matrix and its inverse
self.initial_inertia_matrix = Solid._inertia_matrix(
self.masses, self.centered_initial_points_hstack)
self.initial_inertia_matrix_inverse = npmat.linalg.pinv(
self.initial_inertia_matrix)
# Rotation quaternion (used to construct matrix representing orientation) and angular momentum (sigma)
self.rotation_quaternion = Vect([1., 0., 0., 0.])
self.rotation_matrix = rotation_matrix_from_quaternion(
self.rotation_quaternion)
self.angular_momentum = npmat.zeros((3, 1))
self._compute_AABB()
def dist_matrix(pts):
"""Calculate the euclidean distance matrix (EDM) for a set of points.
Parameters
----------
pts : np.ndarray (shape = (2,))
Returns
-------
dist_mat : np.ndarray
The distance matrix as 2d ndarray.
Implementation Details
----------------------
Uses two auxiliary matrixes to easily calculate the distance from each
point to every other point in the list using this approach:
(1) aux matrixes:
repmat(l, n1, n2): l is repeated n1 times, along axis 1, and n2 times along
axis 2, so repmat(pts, len(pts), 1) =
array( [ [1, 2], [4, 6], [1, 2], [4, 6] ] )
repeat(l, n, a): each element of l is repeated n times along axis a (w/o
'a' a plain list is generated), so repeat(pts, 2, 1) =
array( [ [1, 2], [1, 2], [4, 6], [4, 6] ] )
(2) Pythagoras:
Then, the element-wise difference of the generated matrixes is calculated
each value is squared:
array( [ [ 0, 0], [ 9, 16], [ 9, 16], [ 0, 0] ] )
These squares are then summed up (linewise) using sum(..., axis=1):
array([ 0, 25, 25, 0])
Finally the square root is taken for each element:
array([ 0., 5., 5., 0.])
To transform the list into a distance matrix reshape() is used.
Example
-------
>>> dist_matrix([ [1, 2], [4, 6] ])
array([[ 0., 5.],
[ 5., 0.]])
>>> dist_matrix([ [1.8, 4.1, 4.0], [2.8, 4.7, 4.5], [5.2, 4.2, 4.7],
... [4.1, 4.5, 4.6], [5.7, 3.4, 4.5]])
array([[ 0. , 1.26885775, 3.47275107, 2.41039416, 3.99374511],
[ 1.26885775, 0. , 2.45967478, 1.3190906 , 3.17804972],
[ 3.47275107, 2.45967478, 0. , 1.14455231, 0.96436508],
[ 2.41039416, 1.3190906 , 1.14455231, 0. , 1.94422221],
[ 3.99374511, 3.17804972, 0.96436508, 1.94422221, 0. ]])
"""
dist_mat = scipy.sqrt(
matlib.sum(
(
matlib.repmat(pts, len(pts), 1) -
matlib.repeat(pts, len(pts), axis=0)
) ** 2,
axis=1
)
)
return dist_mat.reshape((len(pts), len(pts)))
def BhattaFromDataset(D1, D2, eta):
assert D1.r == D2.r
kernel = D1.kernel
Beta1 = D1.Beta
Beta2 = D2.Beta
(n1, r) = Beta1.shape
(n2, r) = Beta2.shape
n = n1 + n2
Beta = mat.zeros((n,2*r))
Beta[0:n1,0:r] = Beta1
Beta[n1:n,r:2*r] = Beta2
(K, Kuc, Kc) = kernel_supermatrix(D1,D2)
Omega = eig_ortho(Kc, Beta)
mu1 = mat.sum(Kuc[0:n1, :] * Omega, 0) / n1
mu2 = mat.sum(Kuc[n1:n, :] * Omega, 0) / n2
S1 = Omega.T * Kc[:,0:n1] * Kc[0:n1,:] * Omega / n1
S2 = Omega.T * Kc[:,n1:n] * Kc[n1:n,:] * Omega / n2
Eta= eta * mat.eye(2*r)
S1 += Eta
S2 += Eta
mu3 = .5 * (S1.I * mu1.T + S2.I * mu2.T).T
S3 = 2 * (S1.I + S2.I).I
d1 = la.det(S1) ** -.25
d2 = la.det(S2) ** -.25
d3 = la.det(S3) ** .5
e1 = math.exp(-mu1 * S1.I * mu1.T / 4)
e2 = math.exp(-mu2 * S2.I * mu2.T / 4)
e3 = math.exp(mu3 * S3 * mu3.T / 2)
dterm = d1*d2*d3
eterm = e1*e2*e3
rval = dterm*eterm
if math.isnan(rval):
rval = -1
print "Warning: Kernel failed on datasets ({},{})".format(i,j)
return rval
def BhattaFromDataset(D1, D2, eta):
assert D1.r == D2.r
kernel = D1.kernel
Beta1 = D1.Beta
Beta2 = D2.Beta
(n1, r) = Beta1.shape
(n2, r) = Beta2.shape
n = n1 + n2
Beta = mat.zeros((n, 2 * r))
Beta[0:n1, 0:r] = Beta1
Beta[n1:n, r:2 * r] = Beta2
(K, Kuc, Kc) = kernel_supermatrix(D1, D2)
Omega = eig_ortho(Kc, Beta)
mu1 = mat.sum(Kuc[0:n1, :] * Omega, 0) / n1
mu2 = mat.sum(Kuc[n1:n, :] * Omega, 0) / n2
S1 = Omega.T * Kc[:, 0:n1] * Kc[0:n1, :] * Omega / n1
S2 = Omega.T * Kc[:, n1:n] * Kc[n1:n, :] * Omega / n2
Eta = eta * mat.eye(2 * r)
S1 += Eta
S2 += Eta
mu3 = .5 * (S1.I * mu1.T + S2.I * mu2.T).T
S3 = 2 * (S1.I + S2.I).I
d1 = la.det(S1)**-.25
d2 = la.det(S2)**-.25
d3 = la.det(S3)**.5
e1 = math.exp(-mu1 * S1.I * mu1.T / 4)
e2 = math.exp(-mu2 * S2.I * mu2.T / 4)
e3 = math.exp(mu3 * S3 * mu3.T / 2)
dterm = d1 * d2 * d3
eterm = e1 * e2 * e3
rval = dterm * eterm
if math.isnan(rval):
rval = -1
print "Warning: Kernel failed on datasets ({},{})".format(i, j)
return rval
def EMForMoG(datam, n, pl, muvl, covml, max_iters=1):
assert len(pl)==len(muvl) and len(muvl)==len(covml)
d=datam.shape[0]
N=datam.shape[1]
#start EM posteriorm n*N
posteriorm=matlib.zeros((n, N), dtype=np.float32)
for it in xrange(max_iters):
print 'Iter: '+str(it)
#create Gaussian kernels
NormalFl=[makeNormalF(muv, covm) for muv, covm in zip(muvl, covml)]
#probability
px=matlib.zeros((1, N), dtype=np.float32)
posteriorm.fill(0)
#caculate posteriorm
for j in xrange(N):
cur_data=datam[:, j]
for i in xrange(n):
#print i, j
posteriorm[i, j]=pl[i]*NormalFl[i](cur_data)
px[0, j]+=posteriorm[i, j]
# print 'px:', px
posteriorm/=px
#update parameters
#soft num n*1
softnum=matlib.sum(posteriorm, 1)
print softnum
softnum_inv=1.0/softnum
pl=np.array((softnum/N)).reshape(-1).tolist()
mum=datam*posteriorm.T*matlib.diag(np.array(softnum_inv).reshape(-1))
muvl=[mum[:, k] for k in range(n)]
mum=[]#release
for k in range(n):
datam_temp=datam-muvl[k]
covml[k]=softnum_inv[k, 0]*datam_temp*matlib.diag(np.array(posteriorm[k, :]).reshape(-1))\
*datam_temp.T
return pl, muvl, covml
def verify_kernel_matrix():
n1 = 10
n2 = 10
n = n1+n2
d = 5
degree = 3
X = randn(n,d)
Phi = poly.phi(X, degree)
(K, Kuc, Kc) = kernel_matrix(X, polyk(degree), n1, n2)
P1 = Phi[0:n1,:]
P2 = Phi[n1:n,:]
mu1 = mat.sum(P1,0) / n1
mu2 = mat.sum(P2,0) / n2
P1c = P1 - mat.tile(mu1, (n1,1))
P2c = P2 - mat.tile(mu2, (n2,1))
Pc = bmat('P1c; P2c')
KP = mat.zeros((n,n))
for i in xrange(n):
for j in xrange(i+1):
KP[i,j] = dotp(Phi[i,:], Phi[j,:])
KP[j,i] = KP[i,j]
KucP = mat.zeros((n,n))
for i in xrange(n):
for j in xrange(n):
KucP[i,j] = dotp(Phi[i,:], Pc[j,:])
KcP = mat.zeros((n,n))
for i in xrange(n):
for j in xrange(n):
KcP[i,j] = dotp(Pc[i,:], Pc[j,:])
#KcP[j,i] = KcP[i,j]
#debug()
print "Div1: " + str(sum(abs(K-KP)))
print "Div2: " + str(sum(abs(Kuc-KucP)))
print "Div3: " + str(sum(abs(Kc-KcP)))
def kernel_supermatrix(self, i, j):
kernel = self.kernel
D1 = self.datasets[i]
D2 = self.datasets[j]
X1 = D1.X
X2 = D2.X
(n1, d) = X1.shape
(n2, d) = X2.shape
n = n1 + n2
X = mat.bmat('X1; X2')
K1 = D1.K
K2 = D2.K
K = mat.zeros((n, n))
K[0:n1, 0:n1] = K1
K[n1:n, n1:n] = K2
for i in xrange(n1):
for j in xrange(n1, n):
K[i, j] = kernel(X[i, :], X[j, :])
K[j, i] = K[i, j]
# Inelegant - improve later
U1 = mat.sum(K[0:n1, :], 0) / n1
U2 = mat.sum(K[n1:n, :], 0) / n2
U1m = mat.tile(U1, (n1, 1))
U2m = mat.tile(U2, (n2, 1))
U = mat.bmat('U1m; U2m')
m1m1 = mat.sum(K[0:n1, 0:n1]) / (n1 * n1)
m1m2 = mat.sum(K[0:n1, n1:n]) / (n1 * n2)
m2m2 = mat.sum(K[n1:n, n1:n]) / (n2 * n2)
mumu = mat.zeros((n, n))
mumu[0:n1, 0:n1] = m1m1
mumu[0:n1, n1:n] = m1m2
mumu[n1:n, 0:n1] = m1m2
mumu[n1:n, n1:n] = m2m2
Kcu = K - U
Kuc = Kcu.T
N = mat.ones((n, n)) / n
Kc = K - U - U.T + mumu
return (K, Kuc, Kc)
def kernel_supermatrix(self, i, j):
kernel = self.kernel
D1 = self.datasets[i]
D2 = self.datasets[j]
X1 = D1.X
X2 = D2.X
(n1, d) = X1.shape
(n2, d) = X2.shape
n = n1 + n2
X = mat.bmat('X1; X2')
K1 = D1.K
K2 = D2.K
K = mat.zeros((n,n))
K[0:n1, 0:n1] = K1
K[n1:n, n1:n] = K2
for i in xrange(n1):
for j in xrange(n1, n):
K[i,j] = kernel(X[i,:], X[j,:])
K[j,i] = K[i,j]
# Inelegant - improve later
U1 = mat.sum(K[0:n1,:],0) / n1
U2 = mat.sum(K[n1:n,:],0) / n2
U1m = mat.tile(U1, (n1,1))
U2m = mat.tile(U2, (n2,1))
U = mat.bmat('U1m; U2m')
m1m1 = mat.sum(K[0:n1, 0:n1]) / (n1*n1)
m1m2 = mat.sum(K[0:n1, n1:n]) / (n1*n2)
m2m2 = mat.sum(K[n1:n, n1:n]) / (n2*n2)
mumu = mat.zeros((n,n))
mumu[0:n1, 0:n1] = m1m1
mumu[0:n1, n1:n] = m1m2
mumu[n1:n, 0:n1] = m1m2
mumu[n1:n, n1:n] = m2m2
Kcu = K - U
Kuc = Kcu.T
N = mat.ones((n,n))/n
Kc = K - U - U.T + mumu
return (K, Kuc, Kc)
def kmeans(X, k, observer=None, threshold=1e-15, maxiter=300): N = len(X) labels = zeros(N, dtype=int) centers = array(random.sample(X, k)) iter = 0 def calc_J(): sum = 0 for i in xrange(N): sum += norm(X[i]-centers[labels[i]]) return sum def distmat(X, Y): n = len(X) m = len(Y) xx = ml.sum(X*X, axis=1) yy = ml.sum(Y*Y, axis=1) xy = ml.dot(X, Y.T) return tile(xx, (m, 1)).T+tile(yy, (n, 1)) - 2*xy Jprev = calc_J() while True: #notify the observer if observer is not None: observer(iter, labels, centers) dist = distmat(X, centers) labels = dist.argmin(axis=1) for j in range(k): idx_j = (labels == j).nonzero() distj = distmat(X[idx_j], X[idx_j]) distsum = ml.sum(distj, axis=1) icenter = distsum.argmin() centers[j] = X[idx_j[0][icenter]] J = calc_J() iter += 1 if Jprev-J < threshold: break Jprev = J if iter >= maxiter: break if observer is not None: observer(iter, labels, centers)
def compute_forward_simulated_cost(b, U, model):
T = model.T
cost = 0
b_t = b
for t in xrange(0,T-1):
x_t, s_t = decompose_belief(b_t, model)
cost += model.alpha_belief*ml.trace(s_t*s_t)
cost += model.alpha_control*ml.sum(U[:,t].T*U[:,t])
b_t = belief_dynamics(b_t, U[:,t], None, model)
x_T, s_T = decompose_belief(b_t, model)
cost += model.alpha_final_belief*ml.trace(s_T*s_T)
return cost
def compute_forward_simulated_cost(b, U, model):
T = model.T
cost = 0
b_t = b
for t in xrange(0, T - 1):
x_t, s_t = decompose_belief(b_t, model)
cost += model.alpha_belief * ml.trace(s_t * s_t)
cost += model.alpha_control * ml.sum(U[:, t].T * U[:, t])
b_t = belief_dynamics(b_t, U[:, t], None, model)
x_T, s_T = decompose_belief(b_t, model)
cost += model.alpha_final_belief * ml.trace(s_T * s_T)
return cost
def constraints_satisfied(B, U, model, tolerance):
xDim = model.xDim
xMax = model.xMax
xMin = model.xMin
uMax = model.uMax
uMin = model.uMin
T = model.T
done = True
constraint_violations = 0
for t in xrange(0, T-1):
constraint_violations += ml.sum(np.abs(B[:,t+1] - belief.belief_dynamics(B[:,t],U[:,t],None,model)))
done &= constraint_violations < tolerance
done &= np.max(B[1:xDim,t] <= xMax)
done &= np.min(B[1:xDim,t] >= xMin)
done &= np.max(U[:,t] <= uMax)
done &= np.min(U[:,t] >= uMin)
if not done:
break
#print('Constraint violations: %g' % constraint_violations)
return done
def inf(self, x, meanonly=False):
x = np.asmatrix(x)
assert x.shape[1] == self.d
n = x.shape[0]
# Handle empty test set
if n == 0:
return (np.zeros((0, 1)), np.zeros((0, 1)))
ms = self.kernel.mean*np.ones((n, 1))
Kbb = self.kernel(x, diag=True)
# Handle empty training set
if len(self) == 0:
return (ms, np.asmatrix(np.diag(Kbb)).T)
Kba = self.kernel(x, self.x)
m = self.kernel.mean*np.ones((len(self), 1))
fm = ms + Kba*scipy.linalg.cho_solve((self.L, True), self.y - m,
overwrite_b=True)
if meanonly:
return fm
else:
W = scipy.linalg.cho_solve((self.L, True), Kba.T)
fv = np.asmatrix(Kbb - np.sum(np.multiply(Kba.T, W), axis=0).T)
# W = np.asmatrix(scipy.linalg.solve(self.L, Kba.T, lower=True))
# fv = np.asmatrix(Kbb - np.sum(np.power(W, 2), axis=0).T)
return (fm, fv)
def eig_bhatta(X1, X2, kernel, eta, r):
# Tested. Verified:
# Poly-kernel RKHS representations of all objects are roughly equal to eigenbasis representations (slight differences for S3)
# Correctness for X1 ~= X2
# Close results to empirical bhatta in test_suite_1
# Remaining issues: Eigendecomposition of centered kernel matrices
# occasionally produces negative-value eigenvalues
(n1, d1) = X1.shape
(n2, d2) = X2.shape
assert d1==d2
n = n1+n2
X = mat.bmat("X1;X2")
(K, Kuc, Kc) = kernel_matrix(X, kernel, n1, n2)
Kc1 = Kc[0:n1, 0:n1]
Kc2 = Kc[n1:n, n1:n]
(Lam1, Alpha1) = eigsh(Kc1, r)
(Lam2, Alpha2) = eigsh(Kc2, r)
Alpha1 = matrix(Alpha1)
Alpha2 = matrix(Alpha2)
Lam1 = Lam1 / n1
Lam2 = Lam2 / n2
Beta1 = mat.zeros((n,r))
Beta2 = mat.zeros((n,r))
for i in xrange(r):
Beta1[0:n1, i] = Alpha1[:,i] / (n1 * Lam1[i])**.5
Beta2[n1:n, i] = Alpha2[:,i] / (n2 * Lam2[i])**.5
#Eta = mat.eye((gamma, gamma)) * eta
Beta = mat.bmat('Beta1, Beta2')
assert not(any(math.isnan(Beta)))
Omega = eig_ortho(Kc, Beta)
mu1_w = mat.sum(Kuc[0:n1, :] * Omega, 0) / n1
mu2_w = mat.sum(Kuc[n1:n, :] * Omega, 0) / n2
Eta_w = eta * mat.eye(2*r)
S1_w = Omega.T * Kc[:,0:n1] * Kc[0:n1,:] * Omega / n1
S2_w = Omega.T * Kc[:,n1:n] * Kc[n1:n,:] * Omega / n2
S1_w += Eta_w
S2_w += Eta_w
mu3_w = .5 * (S1_w.I * mu1_w.T + S2_w.I * mu2_w.T).T
S3_w = 2 * (S1_w.I + S2_w.I).I
d1 = la.det(S1_w) ** -.25
d2 = la.det(S2_w) ** -.25
e1 = exp(-mu1_w * S1_w.I * mu1_w.T / 4)
e2 = exp(-mu2_w * S2_w.I * mu2_w.T / 4)
d3 = la.det(S3_w) ** .5
e3 = exp(mu3_w * S3_w * mu3_w.T / 2)
dterm = d1*d2*d3
eterm = e1*e2*e3
rval = float(dterm*eterm)
if math.isnan(rval):
rval = -1
return rval
def kmeans(X, k, observer=None, threshold=1e-15, maxiter=300, style="kmeans"):
N = len(X)
labels = np.zeros(N, dtype=int)
centers = X[np.random.choice(len(X), k)]
itr = 0
def calc_J():
"""
计算所有点距离和
"""
sums = 0
for i in range(N):
sums += norm(X[i] - centers[labels[i]])
return sums
def distmat(X, Y):
"""
计算距离
"""
n = len(X)
m = len(Y)
xx = ml.sum(X * X, axis=1)
yy = ml.sum(Y * Y, axis=1)
xy = ml.dot(X, Y.T)
return np.tile(xx, (m, 1)).T + np.tile(yy, (n, 1)) - 2 * xy
Jprev = calc_J()
while True:
# 绘图
observer(itr, labels, centers)
dist = distmat(X, centers)
labels = dist.argmin(axis=1)
# 再次绘图
observer(itr, labels, centers)
# 重新计算聚类中心
if style == "kmeans":
for j in range(k):
idx_j = (labels == j).nonzero()
centers[j] = X[idx_j].mean(axis=0)
elif style == "kmedoids":
for j in range(k):
idx_j = (labels == j).nonzero()
distj = distmat(X[idx_j], X[idx_j])
distsum = ml.sum(distj, axis=1)
icenter = distsum.argmin()
centers[j] = X[idx_j[0][icenter]]
J = calc_J()
itr += 1
if Jprev - J < threshold:
"""
当中心不再变化停止迭代
"""
break
Jprev = J
if itr >= maxiter:
break
def eig_bhatta(X1, X2, kernel, eta, r):
# Tested. Verified:
# Poly-kernel RKHS representations of all objects are roughly equal to eigenbasis representations (slight differences for S3)
# Correctness for X1 ~= X2
# Close results to empirical bhatta in test_suite_1
# Remaining issues: Eigendecomposition of centered kernel matrices
# occasionally produces negative-value eigenvalues
(n1, d1) = X1.shape
(n2, d2) = X2.shape
assert d1 == d2
n = n1 + n2
X = mat.bmat("X1;X2")
(K, Kuc, Kc) = kernel_matrix(X, kernel, n1, n2)
Kc1 = Kc[0:n1, 0:n1]
Kc2 = Kc[n1:n, n1:n]
(Lam1, Alpha1) = eigsh(Kc1, r)
(Lam2, Alpha2) = eigsh(Kc2, r)
Alpha1 = matrix(Alpha1)
Alpha2 = matrix(Alpha2)
Lam1 = Lam1 / n1
Lam2 = Lam2 / n2
Beta1 = mat.zeros((n, r))
Beta2 = mat.zeros((n, r))
for i in xrange(r):
Beta1[0:n1, i] = Alpha1[:, i] / (n1 * Lam1[i])**.5
Beta2[n1:n, i] = Alpha2[:, i] / (n2 * Lam2[i])**.5
#Eta = mat.eye((gamma, gamma)) * eta
Beta = mat.bmat('Beta1, Beta2')
assert not (any(math.isnan(Beta)))
Omega = eig_ortho(Kc, Beta)
mu1_w = mat.sum(Kuc[0:n1, :] * Omega, 0) / n1
mu2_w = mat.sum(Kuc[n1:n, :] * Omega, 0) / n2
Eta_w = eta * mat.eye(2 * r)
S1_w = Omega.T * Kc[:, 0:n1] * Kc[0:n1, :] * Omega / n1
S2_w = Omega.T * Kc[:, n1:n] * Kc[n1:n, :] * Omega / n2
S1_w += Eta_w
S2_w += Eta_w
mu3_w = .5 * (S1_w.I * mu1_w.T + S2_w.I * mu2_w.T).T
S3_w = 2 * (S1_w.I + S2_w.I).I
d1 = la.det(S1_w)**-.25
d2 = la.det(S2_w)**-.25
e1 = exp(-mu1_w * S1_w.I * mu1_w.T / 4)
e2 = exp(-mu2_w * S2_w.I * mu2_w.T / 4)
d3 = la.det(S3_w)**.5
e3 = exp(mu3_w * S3_w * mu3_w.T / 2)
dterm = d1 * d2 * d3
eterm = e1 * e2 * e3
rval = float(dterm * eterm)
if math.isnan(rval):
rval = -1
return rval
def SPIRIT(A, lamb, energy, k0=1, holdOffTime=0, reorthog=False, evalMetrics="F"):
A = np.mat(A)
n = A.shape[1]
totalTime = A.shape[0]
Proj = npm.ones((totalTime, n)) * np.nan
recon = npm.zeros((totalTime, n))
# initialize w_i to unit vectors
W = npm.eye(n)
d = 0.01 * npm.ones((n, 1))
m = k0 # number of eigencomponents
relErrors = npm.zeros((totalTime, 1))
sumYSq = 0.0
E_t = []
sumXSq = 0.0
E_dash_t = []
res = {}
k_hist = []
W_hist = []
anomalies = []
# incremental update W
lastChangeAt = 0
for t in range(totalTime):
k_hist.append(m)
# update W for each y_t
x = A[t, :].T # new data as column vector
for j in range(m):
W[:, j], d[j], x = updateW(x, W[:, j], d[j], lamb)
Wj = W[:, j]
# Grams smit reorthog
if reorthog == True:
W[:, :m], R = npm.linalg.qr(W[:, :m])
# compute low-D projection, reconstruction and relative error
Y = W[:, :m].T * A[t, :].T # project to m-dimensional space
xActual = A[t, :].T # actual vector of the current time
xProj = W[:, :m] * Y # reconstruction of the current time
Proj[t, :m] = Y.T
recon[t, :] = xProj.T
xOrth = xActual - xProj
relErrors[t] = npm.sum(npm.power(xOrth, 2)) / npm.sum(npm.power(xActual, 2))
# update energy
sumYSq = lamb * sumYSq + npm.sum(npm.power(Y, 2))
E_dash_t.append(sumYSq)
sumXSq = lamb * sumXSq + npm.sum(npm.power(A[t, :], 2))
E_t.append(sumXSq)
# Record RSRE
if t == 0:
top = 0.0
bot = 0.0
top = top + npm.power(npm.linalg.norm(xActual - xProj), 2)
bot = bot + npm.power(npm.linalg.norm(xActual), 2)
new_RSRE = top / bot
if t == 0:
RSRE = new_RSRE
else:
RSRE = npm.vstack((RSRE, new_RSRE))
### Metric EVALUATION ###
# deviation from truth
if evalMetrics == "T":
Qt = W[:, :m]
if t == 0:
res["subspace_error"] = npm.zeros((totalTime, 1))
res["orthog_error"] = npm.zeros((totalTime, 1))
res["angle_error"] = npm.zeros((totalTime, 1))
Cov_mat = npm.zeros([n, n])
# Calculate Covarentce Matrix of data up to time t
Cov_mat = lamb * Cov_mat + npm.dot(xActual, xActual.T)
# Get eigenvalues and eigenvectors
WW, V = npm.linalg.eig(Cov_mat)
# Use this to sort eigenVectors in according to deccending eigenvalue
eig_idx = WW.argsort() # Get sort index
eig_idx = eig_idx[::-1] # Reverse order (default is accending)
# v_r = highest r eigen vectors (accoring to thier eigenvalue if sorted).
V_k = V[:, eig_idx[:m]]
# Calculate subspace error
C = npm.dot(V_k, V_k.T) - npm.dot(Qt, Qt.T)
res["subspace_error"][t, 0] = 10 * np.log10(npm.trace(npm.dot(C.T, C))) # frobenius norm in dB
# Calculate angle between projection matrixes
D = npm.dot(npm.dot(npm.dot(V_k.T, Qt), Qt.T), V_k)
eigVal, eigVec = npm.linalg.eig(D)
angle = npm.arccos(np.sqrt(max(eigVal)))
res["angle_error"][t, 0] = angle
# Calculate deviation from orthonormality
F = npm.dot(Qt.T, Qt) - npm.eye(m)
res["orthog_error"][t, 0] = 10 * np.log10(npm.trace(npm.dot(F.T, F))) # frobenius norm in dB
# Energy thresholding
######################
# check the lower bound of energy level
if sumYSq < energy[0] * sumXSq and lastChangeAt < t - holdOffTime and m < n:
lastChangeAt = t
m = m + 1
anomalies.append(t)
# print 'Increasing m to %d at time %d (ratio %6.2f)\n' % (m, t, 100 * sumYSq/sumXSq)
# check the upper bound of energy level
elif sumYSq > energy[1] * sumXSq and lastChangeAt < t - holdOffTime and m < n and m > 1:
lastChangeAt = t
m = m - 1
# print 'Decreasing m to %d at time %d (ratio %6.2f)\n' % (m, t, 100 * sumYSq/sumXSq)
W_hist.append(W[:, :m])
# set outputs
# Grams smit reorthog
if reorthog == True:
W[:, :m], R = npm.linalg.qr(W[:, :m])
# Data Stores
res2 = {
"hidden": Proj, # Array for hidden Variables
"E_t": np.array(E_t), # total energy of data
"E_dash_t": np.array(E_dash_t), # hidden var energy
"e_ratio": np.array(E_dash_t) / np.array(E_t), # Energy ratio
"rel_orth_err": relErrors, # orthoX error
"RSRE": RSRE, # Relative squared Reconstruction error
"recon": recon, # reconstructed data
"r_hist": k_hist, # history of r values
"W_hist": W_hist, # history of Weights
"anomalies": anomalies,
}
res.update(res2)
return res
def SPIRIT(streams, energyThresh, lamb, evalMetrics):
# Make
if type(streams) == np.ndarray:
streams_iter = iter(streams)
# Max No. Streams
if streams.ndim == 1:
streams = np.expand_dims(streams, axis=1)
num_streams = streams.shape[1]
else:
num_streams = streams.shape[1]
count_over = 0
count_under = 0
#===============================================================================
# Initalise k, w and d, lamb
#===============================================================================
k = 1 # Hidden Variables, initialise to one
# Weights
pc_weights = npm.zeros(num_streams)
pc_weights[0, 0] = 1
# initialise outputs
res = {}
all_weights = []
k_hist = []
anomalies = []
x_dash = npm.zeros((1,num_streams))
Eng = mat([0.00000001, 0.00000001])
E_xt = 0 # Energy of X at time t
E_rec_i = mat([0.000000000000001]) # Energy of reconstruction
Y = npm.zeros(num_streams)
timeSteps = streams.shape[0]
#===============================================================================
# Main Loop
#===============================================================================
for t in range(1, timeSteps + 1): # t = 1,...,200
k_hist.append(k)
x_t_plus_1 = mat(streams_iter.next()) # Read in next signals
d_i = E_rec_i * t
# Step 1 - Update Weights
pc_weights, y_t_i, error = track_W(x_t_plus_1,
k, pc_weights, d_i,
num_streams,
lamb)
# Record hidden variables
padding = num_streams - k
y_bar_t = npm.hstack((y_t_i, mat([nan] * padding)))
Y = npm.vstack((Y,y_bar_t))
# Record Weights
all_weights.append(pc_weights)
# Record reconstrunted z and RSRE
x_dash = npm.vstack((x_dash, y_t_i * pc_weights))
# Record RSRE
if t == 1:
top = 0.0
bot = 0.0
top = top + (norm(x_t_plus_1 - x_dash) ** 2 )
bot = bot + (norm(x_t_plus_1) ** 2)
new_RSRE = top / bot
if t == 1:
RSRE = new_RSRE
else:
RSRE = npm.vstack((RSRE, new_RSRE))
### FOR EVALUATION ###
#deviation from truth
if evalMetrics == 'T' :
Qt = pc_weights.T
if t == 1 :
res['subspace_error'] = npm.zeros((timeSteps,1))
res['orthog_error'] = npm.zeros((timeSteps,1))
res['angle_error'] = npm.zeros((timeSteps,1))
Cov_mat = npm.zeros([num_streams,num_streams])
# Calculate Covarentce Matrix of data up to time t
Cov_mat = lamb * Cov_mat + npm.dot(x_t_plus_1, x_t_plus_1.T)
# Get eigenvalues and eigenvectors
W , V = eig(Cov_mat)
# Use this to sort eigenVectors in according to deccending eigenvalue
eig_idx = W.argsort() # Get sort index
eig_idx = eig_idx[::-1] # Reverse order (default is accending)
# v_r = highest r eigen vectors (accoring to thier eigenvalue if sorted).
V_k = V[:, eig_idx[:k]]
# Calculate subspace error
C = npm.dot(V_k , V_k.T) - npm.dot(Qt , Qt.T)
res['subspace_error'][t-1,0] = 10 * np.log10(npm.trace(npm.dot(C.T , C))) #frobenius norm in dB
# Calculate angle between projection matrixes
D = npm.dot(npm.dot(npm.dot(V_k.T, Qt), Qt.T), V_k)
eigVal, eigVec = eig(D)
angle = npm.arccos(np.sqrt(max(eigVal)))
res['angle_error'][t-1,0] = angle
# Calculate deviation from orthonormality
F = npm.dot(Qt.T , Qt) - npm.eye(k)
res['orthog_error'][t-1,0] = 10 * np.log10(npm.trace(npm.dot(F.T , F))) #frobenius norm in dB
# Step 2 - Update Energy estimate
E_xt = ((lamb * (t-1) * E_xt) + norm(x_t_plus_1) ** 2) / t
for i in range(k):
E_rec_i[0, i] = ((lamb * (t-1) * E_rec_i[0, i]) + (y_t_i[0, i] ** 2)) / t
# Step 3 - Estimate the retained energy
E_retained = npm.sum(E_rec_i,1)
# Record Energy
Eng_new = npm.hstack((E_xt, E_retained[0,0]))
Eng = npm.vstack((Eng, Eng_new))
if E_retained < energyThresh[0] * E_xt:
if k != num_streams:
k = k + 1
# Initalise Ek+1 <-- 0
E_rec_i = npm.hstack((E_rec_i, mat([0])))
# Initialise W_i+1
new_weight_vec = npm.zeros(num_streams)
new_weight_vec[0, k-1] = 1
pc_weights = npm.vstack((pc_weights, new_weight_vec))
anomalies.append(t -1)
else:
count_over += 1
elif E_retained > energyThresh[1] * E_xt:
if k > 1 :
k = k - 1
# discard w_k and error
pc_weights = delete(pc_weights, -1, 0)
# Discard E_rec_i[k]
E_rec_i = delete(E_rec_i, -1)
else:
count_under += 1
# Data Stores
res2 = {'hidden' : Y, # Array for hidden Variables
'weights' : all_weights,
'E_t' : Eng[:,0], # total energy of data
'E_dash_t' : Eng[:,1], # hidden var energy
'e_ratio' : np.divide(Eng[:,1], Eng[:,0]), # Energy ratio
'RSRE' : RSRE, # Relative squared Reconstruction error
'recon' : x_dash, # reconstructed data
'r_hist' : k_hist, # history of r values
'anomalies' : anomalies}
res.update(res2)
return res, all_weights
def SPIRIT(A,
lamb,
energy,
k0=1,
holdOffTime=0,
reorthog=False,
evalMetrics='F'):
A = np.mat(A)
n = A.shape[1]
totalTime = A.shape[0]
Proj = npm.ones((totalTime, n)) * np.nan
recon = npm.zeros((totalTime, n))
# initialize w_i to unit vectors
W = npm.eye(n)
d = 0.01 * npm.ones((n, 1))
m = k0 # number of eigencomponents
relErrors = npm.zeros((totalTime, 1))
sumYSq = 0.
E_t = []
sumXSq = 0.
E_dash_t = []
res = {}
k_hist = []
W_hist = []
anomalies = []
# incremental update W
lastChangeAt = 0
for t in range(totalTime):
k_hist.append(m)
# update W for each y_t
x = A[t, :].T # new data as column vector
for j in range(m):
W[:, j], d[j], x = updateW(x, W[:, j], d[j], lamb)
Wj = W[:, j]
# Grams smit reorthog
if reorthog == True:
W[:, :m], R = npm.linalg.qr(W[:, :m])
# compute low-D projection, reconstruction and relative error
Y = W[:, :m].T * A[t, :].T # project to m-dimensional space
xActual = A[t, :].T # actual vector of the current time
xProj = W[:, :m] * Y # reconstruction of the current time
Proj[t, :m] = Y.T
recon[t, :] = xProj.T
xOrth = xActual - xProj
relErrors[t] = npm.sum(npm.power(xOrth, 2)) / npm.sum(
npm.power(xActual, 2))
# update energy
sumYSq = lamb * sumYSq + npm.sum(npm.power(Y, 2))
E_dash_t.append(sumYSq)
sumXSq = lamb * sumXSq + npm.sum(npm.power(A[t, :], 2))
E_t.append(sumXSq)
# Record RSRE
if t == 0:
top = 0.0
bot = 0.0
top = top + npm.power(npm.linalg.norm(xActual - xProj), 2)
bot = bot + npm.power(npm.linalg.norm(xActual), 2)
new_RSRE = top / bot
if t == 0:
RSRE = new_RSRE
else:
RSRE = npm.vstack((RSRE, new_RSRE))
### Metric EVALUATION ###
#deviation from truth
if evalMetrics == 'T':
Qt = W[:, :m]
if t == 0:
res['subspace_error'] = npm.zeros((totalTime, 1))
res['orthog_error'] = npm.zeros((totalTime, 1))
res['angle_error'] = npm.zeros((totalTime, 1))
Cov_mat = npm.zeros([n, n])
# Calculate Covarentce Matrix of data up to time t
Cov_mat = lamb * Cov_mat + npm.dot(xActual, xActual.T)
# Get eigenvalues and eigenvectors
WW, V = npm.linalg.eig(Cov_mat)
# Use this to sort eigenVectors in according to deccending eigenvalue
eig_idx = WW.argsort() # Get sort index
eig_idx = eig_idx[::-1] # Reverse order (default is accending)
# v_r = highest r eigen vectors (accoring to thier eigenvalue if sorted).
V_k = V[:, eig_idx[:m]]
# Calculate subspace error
C = npm.dot(V_k, V_k.T) - npm.dot(Qt, Qt.T)
res['subspace_error'][t, 0] = 10 * np.log10(
npm.trace(npm.dot(C.T, C))) #frobenius norm in dB
# Calculate angle between projection matrixes
D = npm.dot(npm.dot(npm.dot(V_k.T, Qt), Qt.T), V_k)
eigVal, eigVec = npm.linalg.eig(D)
angle = npm.arccos(np.sqrt(max(eigVal)))
res['angle_error'][t, 0] = angle
# Calculate deviation from orthonormality
F = npm.dot(Qt.T, Qt) - npm.eye(m)
res['orthog_error'][t, 0] = 10 * np.log10(
npm.trace(npm.dot(F.T, F))) #frobenius norm in dB
# Energy thresholding
######################
# check the lower bound of energy level
if sumYSq < energy[
0] * sumXSq and lastChangeAt < t - holdOffTime and m < n:
lastChangeAt = t
m = m + 1
anomalies.append(t)
# print 'Increasing m to %d at time %d (ratio %6.2f)\n' % (m, t, 100 * sumYSq/sumXSq)
# check the upper bound of energy level
elif sumYSq > energy[
1] * sumXSq and lastChangeAt < t - holdOffTime and m < n and m > 1:
lastChangeAt = t
m = m - 1
# print 'Decreasing m to %d at time %d (ratio %6.2f)\n' % (m, t, 100 * sumYSq/sumXSq)
W_hist.append(W[:, :m])
# set outputs
# Grams smit reorthog
if reorthog == True:
W[:, :m], R = npm.linalg.qr(W[:, :m])
# Data Stores
res2 = {
'hidden': Proj, # Array for hidden Variables
'E_t': np.array(E_t), # total energy of data
'E_dash_t': np.array(E_dash_t), # hidden var energy
'e_ratio': np.array(E_dash_t) / np.array(E_t), # Energy ratio
'rel_orth_err': relErrors, # orthoX error
'RSRE': RSRE, # Relative squared Reconstruction error
'recon': recon, # reconstructed data
'r_hist': k_hist, # history of r values
'W_hist': W_hist, # history of Weights
'anomalies': anomalies
}
res.update(res2)
return res
def MMAPPH1FCFS(D, sigma, S, *argv):
"""
Returns various performane measures of a MMAP[K]/PH[K]/1
first-come-first-serve queue, see [1]_.
Parameters
----------
D : list of matrices of shape (N,N), length (K+1)
The D0...DK matrices of the arrival process.
sigma : list of row vectors, length (K)
The list containing the initial probability vectors of the service
time distributions of the various customer types. The length of the
vectors does not have to be the same.
S : list of square matrices, length (K)
The transient generators of the phase type distributions representing
the service time of the jobs belonging to various types.
further parameters :
The rest of the function parameters specify the options
and the performance measures to be computed.
The supported performance measures and options in this
function are:
+----------------+--------------------+----------------------------------------+
| Parameter name | Input parameters | Output |
+================+====================+========================================+
| "ncMoms" | Number of moments | The moments of the number of customers |
+----------------+--------------------+----------------------------------------+
| "ncDistr" | Upper limit K | The distribution of the number of |
| | | customers from level 0 to level K-1 |
+----------------+--------------------+----------------------------------------+
| "stMoms" | Number of moments | The sojourn time moments |
+----------------+--------------------+----------------------------------------+
| "stDistr" | A vector of points | The sojourn time distribution at the |
| | | requested points (cummulative, cdf) |
+----------------+--------------------+----------------------------------------+
| "stDistrME" | None | The vector-matrix parameters of the |
| | | matrix-exponentially distributed |
| | | sojourn time distribution |
+----------------+--------------------+----------------------------------------+
| "stDistrPH" | None | The vector-matrix parameters of the |
| | | matrix-exponentially distributed |
| | | sojourn time distribution, converted |
| | | to a continuous PH representation |
+----------------+--------------------+----------------------------------------+
| "prec" | The precision | Numerical precision used as a stopping |
| | | condition when solving the Riccati |
| | | equation |
+----------------+--------------------+----------------------------------------+
| "classes" | Vector of integers | Only the performance measures |
| | | belonging to these classes are |
| | | returned. If not given, all classes |
| | | are analyzed. |
+----------------+--------------------+----------------------------------------+
(The quantities related to the number of customers in
the system include the customer in the server, and the
sojourn time related quantities include the service
times as well)
Returns
-------
Ret : list of the performance measures
Each entry of the list corresponds to a performance
measure requested. Each entry is a matrix, where the
columns belong to the various job types.
If there is just a single item,
then it is not put into a list.
References
----------
.. [1] Qiming He, "Analysis of a continuous time
SM[K]/PH[K]/1/FCFS queue: Age process, sojourn times,
and queue lengths", Journal of Systems Science and
Complexity, 25(1), pp 133-155, 2012.
"""
K = len(D) - 1
# parse options
eaten = []
precision = 1e-14
classes = np.arange(0, K)
for i in range(len(argv)):
if argv[i] == "prec":
precision = argv[i + 1]
eaten.append(i)
eaten.append(i + 1)
elif argv[i] == "classes":
classes = np.array(argv[i + 1]) - 1
eaten.append(i)
eaten.append(i + 1)
if butools.checkInput and not CheckMMAPRepresentation(D):
raise Exception(
'MMAPPH1FCFS: The arrival process is not a valid MMAP representation!'
)
if butools.checkInput:
for k in range(K):
if not CheckPHRepresentation(sigma[k], S[k]):
raise Exception(
'MMAPPH1FCFS: the vector and matrix describing the service times is not a valid PH representation!'
)
# some preparation
D0 = D[0]
N = D0.shape[0]
Ia = ml.eye(N)
Da = ml.zeros((N, N))
for q in range(K):
Da += D[q + 1]
theta = CTMCSolve(D0 + Da)
beta = [CTMCSolve(S[k] + ml.sum(-S[k], 1) * sigma[k]) for k in range(K)]
lambd = [np.sum(theta * D[k + 1]) for k in range(K)]
mu = [np.sum(beta[k] * (-S[k])) for k in range(K)]
Nsk = [S[k].shape[0] for k in range(K)]
ro = np.sum(np.array(lambd) / np.array(mu))
alpha = theta * Da / sum(lambd)
D0i = (-D0).I
Sa = S[0]
sa = [ml.zeros(sigma[0].shape)] * K
sa[0] = sigma[0]
ba = [ml.zeros(beta[0].shape)] * K
ba[0] = beta[0]
sv = [ml.zeros((Nsk[0], 1))] * K
sv[0] = ml.sum(-S[0], 1)
Pk = [D0i * D[q + 1] for q in range(K)]
for k in range(1, K):
Sa = la.block_diag(Sa, S[k])
for q in range(K):
if q == k:
sa[q] = ml.hstack((sa[q], sigma[k]))
ba[q] = ml.hstack((ba[q], beta[k]))
sv[q] = ml.vstack((sv[q], -np.sum(S[k], 1)))
else:
sa[q] = ml.hstack((sa[q], ml.zeros(sigma[k].shape)))
ba[q] = ml.hstack((ba[q], ml.zeros(beta[k].shape)))
sv[q] = ml.vstack((sv[q], ml.zeros((Nsk[k], 1))))
Sa = ml.matrix(Sa)
P = D0i * Da
iVec = ml.kron(D[1], sa[0])
for k in range(1, K):
iVec += ml.kron(D[k + 1], sa[k])
Ns = Sa.shape[0]
Is = ml.eye(Ns)
# step 1. solve the age process of the queue
# ==========================================
# solve Y0 and calculate T
Y0 = FluidFundamentalMatrices(ml.kron(Ia, Sa), ml.kron(Ia, -ml.sum(Sa, 1)),
iVec, D0, "P", precision)
T = ml.kron(Ia, Sa) + Y0 * iVec
# calculate pi0 and v0
pi0 = ml.zeros((1, T.shape[0]))
for k in range(K):
pi0 += ml.kron(theta * D[k + 1], ba[k] / mu[k])
pi0 = -pi0 * T
iT = (-T).I
oa = ml.ones((N, 1))
# step 2. calculate performance measures
# ======================================
Ret = []
for k in classes:
argIx = 0
clo = iT * ml.kron(oa, sv[k])
while argIx < len(argv):
if argIx in eaten:
argIx += 1
continue
elif type(argv[argIx]) is str and argv[argIx] == "stMoms":
numOfSTMoms = argv[argIx + 1]
rtMoms = []
for m in range(1, numOfSTMoms + 1):
rtMoms.append(
math.factorial(m) * np.sum(pi0 * iT**m * clo /
(pi0 * clo)))
Ret.append(rtMoms)
argIx += 1
elif type(argv[argIx]) is str and argv[argIx] == "stDistr":
stCdfPoints = argv[argIx + 1]
cdf = []
for t in stCdfPoints:
pr = 1 - np.sum(pi0 * la.expm(T * t) * clo / (pi0 * clo))
cdf.append(pr)
Ret.append(np.array(cdf))
argIx += 1
elif type(argv[argIx]) is str and argv[argIx] == "stDistrME":
Bm = SimilarityMatrixForVectors(clo / (pi0 * clo),
ml.ones((N * Ns, 1)))
Bmi = Bm.I
A = Bm * T * Bmi
alpha = pi0 * Bmi
Ret.append(alpha)
Ret.append(A)
elif type(argv[argIx]) is str and argv[argIx] == "stDistrPH":
vv = pi0 * iT
ix = np.arange(N * Ns)
nz = ix[vv.flat > precision]
delta = Diag(vv[:, nz])
cl = -T * clo / (pi0 * clo)
alpha = cl[nz, :].T * delta
A = delta.I * T[nz, :][:, nz].T * delta
Ret.append(alpha)
Ret.append(A)
elif type(argv[argIx]) is str and argv[argIx] == "ncDistr":
numOfQLProbs = argv[argIx + 1]
argIx += 1
values = np.empty(numOfQLProbs)
jm = ml.zeros((Ns, 1))
jm[np.sum(Nsk[0:k]):np.sum(Nsk[0:k + 1]), :] = 1
jmc = ml.ones((Ns, 1))
jmc[np.sum(Nsk[0:k]):np.sum(Nsk[0:k + 1]), :] = 0
LmCurr = la.solve_sylvester(T, ml.kron(D0 + Da - D[k + 1], Is),
-ml.eye(N * Ns))
values[0] = 1 - ro + np.sum(pi0 * LmCurr * ml.kron(oa, jmc))
for i in range(1, numOfQLProbs):
LmPrev = LmCurr
LmCurr = la.solve_sylvester(
T, ml.kron(D0 + Da - D[k + 1], Is),
-LmPrev * ml.kron(D[k + 1], Is))
values[i] = np.sum(pi0 * LmCurr * ml.kron(oa, jmc) +
pi0 * LmPrev * ml.kron(oa, jm))
Ret.append(values)
elif type(argv[argIx]) is str and argv[argIx] == "ncMoms":
numOfQLMoms = argv[argIx + 1]
argIx += 1
jm = ml.zeros((Ns, 1))
jm[np.sum(Nsk[0:k]):np.sum(Nsk[0:k + 1]), :] = 1
ELn = [
la.solve_sylvester(T, ml.kron(D0 + Da, Is),
-ml.eye(N * Ns))
]
qlMoms = []
for n in range(1, numOfQLMoms + 1):
bino = 1
Btag = ml.zeros((N * Ns, N * Ns))
for i in range(n):
Btag += bino * ELn[i]
bino *= (n - i) / (i + 1)
ELn.append(
la.solve_sylvester(T, ml.kron(D0 + Da, Is),
-Btag * ml.kron(D[k + 1], Is)))
qlMoms.append(
np.sum(pi0 * ELn[n]) +
np.sum(pi0 * Btag * ml.kron(oa, jm)))
Ret.append(qlMoms)
else:
raise Exception("MMAPPH1FCFS: Unknown parameter " +
str(argv[argIx]))
argIx += 1
if len(Ret) == 1:
return Ret[0]
else:
return Ret
def SPIRIT(streams, energyThresh, lamb, evalMetrics):
# Make
if type(streams) == np.ndarray:
streams_iter = iter(streams)
# Max No. Streams
if streams.ndim == 1:
streams = np.expand_dims(streams, axis=1)
num_streams = streams.shape[1]
else:
num_streams = streams.shape[1]
count_over = 0
count_under = 0
#===============================================================================
# Initalise k, w and d, lamb
#===============================================================================
k = 1 # Hidden Variables, initialise to one
# Weights
pc_weights = npm.zeros(num_streams)
pc_weights[0, 0] = 1
# initialise outputs
res = {}
all_weights = []
k_hist = []
anomalies = []
x_dash = npm.zeros((1, num_streams))
Eng = mat([0.00000001, 0.00000001])
E_xt = 0 # Energy of X at time t
E_rec_i = mat([0.000000000000001]) # Energy of reconstruction
Y = npm.zeros(num_streams)
timeSteps = streams.shape[0]
#===============================================================================
# Main Loop
#===============================================================================
for t in range(1, timeSteps + 1): # t = 1,...,200
k_hist.append(k)
x_t_plus_1 = mat(streams_iter.next()) # Read in next signals
d_i = E_rec_i * t
# Step 1 - Update Weights
pc_weights, y_t_i, error = track_W(x_t_plus_1, k, pc_weights, d_i,
num_streams, lamb)
# Record hidden variables
padding = num_streams - k
y_bar_t = npm.hstack((y_t_i, mat([nan] * padding)))
Y = npm.vstack((Y, y_bar_t))
# Record Weights
all_weights.append(pc_weights)
# Record reconstrunted z and RSRE
x_dash = npm.vstack((x_dash, y_t_i * pc_weights))
# Record RSRE
if t == 1:
top = 0.0
bot = 0.0
top = top + (norm(x_t_plus_1 - x_dash)**2)
bot = bot + (norm(x_t_plus_1)**2)
new_RSRE = top / bot
if t == 1:
RSRE = new_RSRE
else:
RSRE = npm.vstack((RSRE, new_RSRE))
### FOR EVALUATION ###
#deviation from truth
if evalMetrics == 'T':
Qt = pc_weights.T
if t == 1:
res['subspace_error'] = npm.zeros((timeSteps, 1))
res['orthog_error'] = npm.zeros((timeSteps, 1))
res['angle_error'] = npm.zeros((timeSteps, 1))
Cov_mat = npm.zeros([num_streams, num_streams])
# Calculate Covarentce Matrix of data up to time t
Cov_mat = lamb * Cov_mat + npm.dot(x_t_plus_1, x_t_plus_1.T)
# Get eigenvalues and eigenvectors
W, V = eig(Cov_mat)
# Use this to sort eigenVectors in according to deccending eigenvalue
eig_idx = W.argsort() # Get sort index
eig_idx = eig_idx[::-1] # Reverse order (default is accending)
# v_r = highest r eigen vectors (accoring to thier eigenvalue if sorted).
V_k = V[:, eig_idx[:k]]
# Calculate subspace error
C = npm.dot(V_k, V_k.T) - npm.dot(Qt, Qt.T)
res['subspace_error'][t - 1, 0] = 10 * np.log10(
npm.trace(npm.dot(C.T, C))) #frobenius norm in dB
# Calculate angle between projection matrixes
D = npm.dot(npm.dot(npm.dot(V_k.T, Qt), Qt.T), V_k)
eigVal, eigVec = eig(D)
angle = npm.arccos(np.sqrt(max(eigVal)))
res['angle_error'][t - 1, 0] = angle
# Calculate deviation from orthonormality
F = npm.dot(Qt.T, Qt) - npm.eye(k)
res['orthog_error'][t - 1, 0] = 10 * np.log10(
npm.trace(npm.dot(F.T, F))) #frobenius norm in dB
# Step 2 - Update Energy estimate
E_xt = ((lamb * (t - 1) * E_xt) + norm(x_t_plus_1)**2) / t
for i in range(k):
E_rec_i[0, i] = ((lamb * (t - 1) * E_rec_i[0, i]) +
(y_t_i[0, i]**2)) / t
# Step 3 - Estimate the retained energy
E_retained = npm.sum(E_rec_i, 1)
# Record Energy
Eng_new = npm.hstack((E_xt, E_retained[0, 0]))
Eng = npm.vstack((Eng, Eng_new))
if E_retained < energyThresh[0] * E_xt:
if k != num_streams:
k = k + 1
# Initalise Ek+1 <-- 0
E_rec_i = npm.hstack((E_rec_i, mat([0])))
# Initialise W_i+1
new_weight_vec = npm.zeros(num_streams)
new_weight_vec[0, k - 1] = 1
pc_weights = npm.vstack((pc_weights, new_weight_vec))
anomalies.append(t - 1)
else:
count_over += 1
elif E_retained > energyThresh[1] * E_xt:
if k > 1:
k = k - 1
# discard w_k and error
pc_weights = delete(pc_weights, -1, 0)
# Discard E_rec_i[k]
E_rec_i = delete(E_rec_i, -1)
else:
count_under += 1
# Data Stores
res2 = {
'hidden': Y, # Array for hidden Variables
'weights': all_weights,
'E_t': Eng[:, 0], # total energy of data
'E_dash_t': Eng[:, 1], # hidden var energy
'e_ratio': np.divide(Eng[:, 1], Eng[:, 0]), # Energy ratio
'RSRE': RSRE, # Relative squared Reconstruction error
'recon': x_dash, # reconstructed data
'r_hist': k_hist, # history of r values
'anomalies': anomalies
}
res.update(res2)
return res, all_weights
def MMAPPH1FCFS(D, sigma, S, *argv):
"""
Returns various performane measures of a MMAP[K]/PH[K]/1
first-come-first-serve queue, see [1]_.
Parameters
----------
D : list of matrices of shape (N,N), length (K+1)
The D0...DK matrices of the arrival process.
sigma : list of row vectors, length (K)
The list containing the initial probability vectors of the service
time distributions of the various customer types. The length of the
vectors does not have to be the same.
S : list of square matrices, length (K)
The transient generators of the phase type distributions representing
the service time of the jobs belonging to various types.
further parameters :
The rest of the function parameters specify the options
and the performance measures to be computed.
The supported performance measures and options in this
function are:
+----------------+--------------------+----------------------------------------+
| Parameter name | Input parameters | Output |
+================+====================+========================================+
| "ncMoms" | Number of moments | The moments of the number of customers |
+----------------+--------------------+----------------------------------------+
| "ncDistr" | Upper limit K | The distribution of the number of |
| | | customers from level 0 to level K-1 |
+----------------+--------------------+----------------------------------------+
| "stMoms" | Number of moments | The sojourn time moments |
+----------------+--------------------+----------------------------------------+
| "stDistr" | A vector of points | The sojourn time distribution at the |
| | | requested points (cummulative, cdf) |
+----------------+--------------------+----------------------------------------+
| "stDistrME" | None | The vector-matrix parameters of the |
| | | matrix-exponentially distributed |
| | | sojourn time distribution |
+----------------+--------------------+----------------------------------------+
| "stDistrPH" | None | The vector-matrix parameters of the |
| | | matrix-exponentially distributed |
| | | sojourn time distribution, converted |
| | | to a continuous PH representation |
+----------------+--------------------+----------------------------------------+
| "prec" | The precision | Numerical precision used as a stopping |
| | | condition when solving the Riccati |
| | | equation |
+----------------+--------------------+----------------------------------------+
| "classes" | Vector of integers | Only the performance measures |
| | | belonging to these classes are |
| | | returned. If not given, all classes |
| | | are analyzed. |
+----------------+--------------------+----------------------------------------+
(The quantities related to the number of customers in
the system include the customer in the server, and the
sojourn time related quantities include the service
times as well)
Returns
-------
Ret : list of the performance measures
Each entry of the list corresponds to a performance
measure requested. Each entry is a matrix, where the
columns belong to the various job types.
If there is just a single item,
then it is not put into a list.
References
----------
.. [1] Qiming He, "Analysis of a continuous time
SM[K]/PH[K]/1/FCFS queue: Age process, sojourn times,
and queue lengths", Journal of Systems Science and
Complexity, 25(1), pp 133-155, 2012.
"""
K = len(D)-1
# parse options
eaten = []
precision = 1e-14;
classes = np.arange(0,K)
for i in range(len(argv)):
if argv[i]=="prec":
precision = argv[i+1]
eaten.append(i)
eaten.append(i+1)
elif argv[i]=="classes":
classes = np.array(argv[i+1])-1
eaten.append(i)
eaten.append(i+1)
if butools.checkInput and not CheckMMAPRepresentation(D):
raise Exception('MMAPPH1FCFS: The arrival process is not a valid MMAP representation!')
if butools.checkInput:
for k in range(K):
if not CheckPHRepresentation(sigma[k],S[k]):
raise Exception('MMAPPH1FCFS: the vector and matrix describing the service times is not a valid PH representation!')
# some preparation
D0 = D[0]
N = D0.shape[0]
Ia = ml.eye(N);
Da = ml.zeros((N,N))
for q in range(K):
Da += D[q+1]
theta = CTMCSolve(D0+Da)
beta = [CTMCSolve(S[k]+ml.sum(-S[k],1)*sigma[k]) for k in range(K)]
lambd = [np.sum(theta*D[k+1]) for k in range(K)]
mu = [np.sum(beta[k]*(-S[k])) for k in range(K)]
Nsk = [S[k].shape[0] for k in range(K)]
ro = np.sum(np.array(lambd)/np.array(mu))
alpha = theta*Da/sum(lambd)
D0i = (-D0).I
Sa = S[0];
sa = [ml.zeros(sigma[0].shape)]*K
sa[0] = sigma[0]
ba = [ml.zeros(beta[0].shape)]*K
ba[0] = beta[0]
sv = [ml.zeros((Nsk[0],1))]*K
sv[0] = ml.sum(-S[0],1)
Pk = [D0i*D[q+1] for q in range(K)]
for k in range(1,K):
Sa = la.block_diag(Sa, S[k])
for q in range(K):
if q==k:
sa[q] = ml.hstack((sa[q], sigma[k]))
ba[q] = ml.hstack((ba[q], beta[k]))
sv[q] = ml.vstack((sv[q], -np.sum(S[k],1)))
else:
sa[q] = ml.hstack((sa[q], ml.zeros(sigma[k].shape)))
ba[q] = ml.hstack((ba[q], ml.zeros(beta[k].shape)))
sv[q] = ml.vstack((sv[q], ml.zeros((Nsk[k],1))))
Sa = ml.matrix(Sa)
P = D0i*Da
iVec = ml.kron(D[1],sa[0])
for k in range(1,K):
iVec += ml.kron(D[k+1],sa[k])
Ns = Sa.shape[0]
Is = ml.eye(Ns)
# step 1. solve the age process of the queue
# ==========================================
# solve Y0 and calculate T
Y0 = FluidFundamentalMatrices (ml.kron(Ia,Sa), ml.kron(Ia,-ml.sum(Sa,1)), iVec, D0, "P", precision)
T = ml.kron(Ia,Sa) + Y0 * iVec
# calculate pi0 and v0
pi0 = ml.zeros((1,T.shape[0]))
for k in range(K):
pi0 += ml.kron(theta*D[k+1],ba[k]/mu[k])
pi0 = - pi0 * T
iT = (-T).I
oa = ml.ones((N,1))
# step 2. calculate performance measures
# ======================================
Ret = []
for k in classes:
argIx = 0
clo = iT*ml.kron(oa,sv[k])
while argIx<len(argv):
if argIx in eaten:
argIx += 1
continue
elif type(argv[argIx]) is str and argv[argIx]=="stMoms":
numOfSTMoms = argv[argIx+1]
rtMoms = []
for m in range(1,numOfSTMoms+1):
rtMoms.append(math.factorial(m) * np.sum(pi0 * iT**m * clo / (pi0*clo)))
Ret.append(rtMoms)
argIx += 1
elif type(argv[argIx]) is str and argv[argIx]=="stDistr":
stCdfPoints = argv[argIx+1]
cdf = [];
for t in stCdfPoints:
pr = 1 - np.sum(pi0 * la.expm(T*t) * clo / (pi0*clo))
cdf.append(pr)
Ret.append(np.array(cdf))
argIx += 1
elif type(argv[argIx]) is str and argv[argIx]=="stDistrME":
Bm = SimilarityMatrixForVectors(clo/(pi0*clo),ml.ones((N*Ns,1)))
Bmi = Bm.I
A = Bm * T * Bmi
alpha = pi0 * Bmi
Ret.append(alpha)
Ret.append(A)
elif type(argv[argIx]) is str and argv[argIx]=="stDistrPH":
vv = pi0*iT
ix = np.arange(N*Ns)
nz = ix[vv.flat>precision]
delta = Diag(vv[:,nz])
cl = -T*clo/(pi0*clo)
alpha = cl[nz,:].T*delta
A = delta.I*T[nz,:][:,nz].T*delta
Ret.append(alpha)
Ret.append(A)
elif type(argv[argIx]) is str and argv[argIx]=="ncDistr":
numOfQLProbs = argv[argIx+1]
argIx += 1
values = np.empty(numOfQLProbs)
jm = ml.zeros((Ns,1))
jm[np.sum(Nsk[0:k]):np.sum(Nsk[0:k+1]),:] = 1
jmc = ml.ones((Ns,1))
jmc[np.sum(Nsk[0:k]):np.sum(Nsk[0:k+1]),:] = 0
LmCurr = la.solve_sylvester(T, ml.kron(D0+Da-D[k+1],Is), -ml.eye(N*Ns))
values[0] = 1-ro+np.sum(pi0*LmCurr*ml.kron(oa,jmc))
for i in range(1,numOfQLProbs):
LmPrev = LmCurr
LmCurr = la.solve_sylvester(T, ml.kron(D0+Da-D[k+1],Is), -LmPrev*ml.kron(D[k+1],Is))
values[i] = np.sum(pi0*LmCurr*ml.kron(oa,jmc) + pi0*LmPrev*ml.kron(oa,jm));
Ret.append(values)
elif type(argv[argIx]) is str and argv[argIx]=="ncMoms":
numOfQLMoms = argv[argIx+1]
argIx += 1
jm = ml.zeros((Ns,1))
jm[np.sum(Nsk[0:k]):np.sum(Nsk[0:k+1]),:] = 1
ELn = [la.solve_sylvester(T, ml.kron(D0+Da,Is), -ml.eye(N*Ns))]
qlMoms = []
for n in range(1,numOfQLMoms+1):
bino = 1
Btag = ml.zeros((N*Ns,N*Ns))
for i in range(n):
Btag += bino * ELn[i]
bino *= (n-i) / (i+1)
ELn.append(la.solve_sylvester(T, ml.kron(D0+Da,Is), -Btag*ml.kron(D[k+1],Is)))
qlMoms.append(np.sum(pi0*ELn[n]) + np.sum(pi0*Btag*ml.kron(oa,jm)))
Ret.append(qlMoms)
else:
raise Exception("MMAPPH1FCFS: Unknown parameter "+str(argv[argIx]))
argIx += 1
if len(Ret)==1:
return Ret[0]
else:
return Ret
