def computeParticleWeights(self):
if self.debug == 2:print "\t***computeParticleWeights"
for k in range(self.nparticles):
this_model = self.model_curr[k]
this_param = self.parameters_curr[k]
# calculate model prior probility
mprob = self.modelprior[ this_model ]
np = len(self.parameters_curr[k])
# particle prior probability
pprob = 1
for n in range(0,np):
x = 1.0
if self.models[ this_model ].prior[n][0]==0:
x=1
if self.models[ this_model ].prior[n][0]==1:
x=statistics.getPdfGauss(self.models[ this_model ].prior[n][1],
numpy.sqrt(self.models[ this_model ].prior[n][2]),
this_param[n])
if self.models[ this_model ].prior[n][0]==2:
x=statistics.getPdfUniform(self.models[ this_model ].prior[n][1],
self.models[ this_model ].prior[n][2],
this_param[n])
if self.models[ this_model ].prior[n][0]==3:
x=statistics.getPdfLognormal(self.models[ this_model ].prior[n][1],
numpy.sqrt(self.models[ this_model ].prior[n][2]),
this_param[n])
pprob = pprob*x
numer = self.b[k] * mprob * pprob
denom_m = 0
for i in range(self.nmodel):
denom_m = denom_m + self.margins_prev[i]*getPdfModelKernel(this_model, i, self.modelKernel, self.nmodel, self.dead_models)
denom = 0
# print "Calculating denom\t", selected_model, sampleParameters
for j in range(self.nparticles):
if(int(this_model) == int(self.model_prev[j]) ):
# print "\t", j, model_prev[j], weights_prev[j], parameters_prev[j]
if self.debug == 2:
print "\tj, weights_prev, kernelpdf", j, self.weights_prev[j],
self.kernelpdffn(this_param, self.parameters_prev[j], self.models[this_model].prior, self.kernels[this_model], self.kernel_aux[j], self.kernel_type )
denom = denom + self.weights_prev[j] * self.kernelpdffn(this_param, self.parameters_prev[j], self.models[this_model].prior, self.kernels[this_model], self.kernel_aux[j], self.kernel_type )
if self.debug == 2: print "\tnumer/denom_m/denom/m(t-1) : ", numer,denom_m, denom, self.margins_prev[this_model]
self.weights_curr[k] = numer/(denom_m*denom/self.margins_prev[this_model])
def getPdfParameterKernel(params, params0, priors, kernel, auxilliary,
kernel_type):
if kernel_type == 1:
prob = 1
# n refers to the index of the parameter (integer between 0 and np-1)
# ind is an integer between 0 and len(kernel[0])-1 which enables to determine the kernel to use
ind = 0
for n in kernel[0]:
kern = statistics.getPdfUniform(params0[n] + kernel[2][ind][0],
params0[n] + kernel[2][ind][1],
params[n])
prob = prob * kern
ind += 1
return prob
elif kernel_type == 2:
# n refers to the index of the parameter (integer between 0 and np-1)
# ind is an integer between 0 and len(kernel[0])-1 which enables to determine the kernel to use
prob = 1
ind = 0
for n in kernel[0]:
mean = params0[n]
scale = numpy.sqrt(kernel[2][ind])
kern = statistics.getPdfGauss(mean, scale, params[n])
kern = kern / auxilliary[n]
prob = prob * kern
ind += 1
return prob
elif kernel_type == 3:
p0 = list()
p = list()
for n in kernel[0]:
p0.append(params0[n])
p.append(params[n])
kern = statistics.getPdfMultinormal(p0, kernel[2], p)
kern = kern / auxilliary
return kern
elif (kernel_type == 4 or kernel_type == 5):
p0 = list()
p = list()
D = kernel[2]
for n in kernel[0]:
p0.append(params0[n])
p.append(params[n])
kern = statistics.getPdfMultinormal(p0, D[str(params0)], p)
kern = kern / auxilliary
return kern
def getPdfParameterKernel(params, params0, priors, kernel, auxilliary, kernel_type):
if kernel_type == 1:
prob = 1
# n refers to the index of the parameter (integer between 0 and np-1)
# ind is an integer between 0 and len(kernel[0])-1 which enables to determine the kernel to use
ind = 0
for n in kernel[0]:
kern = statistics.getPdfUniform(params0[n] + kernel[2][ind][0], params0[n] + kernel[2][ind][1], params[n])
prob = prob * kern
ind += 1
return prob
elif kernel_type == 2:
# n refers to the index of the parameter (integer between 0 and np-1)
# ind is an integer between 0 and len(kernel[0])-1 which enables to determine the kernel to use
prob = 1
ind = 0
for n in kernel[0]:
mean = params0[n]
scale = numpy.sqrt(kernel[2][ind])
kern = statistics.getPdfGauss(mean, scale, params[n])
kern = kern / auxilliary[n]
prob = prob * kern
ind += 1
return prob
elif kernel_type == 3:
p0 = list()
p = list()
for n in kernel[0]:
p0.append(params0[n])
p.append(params[n])
kern = statistics.getPdfMultinormal(p0, kernel[2], p)
kern = kern / auxilliary
return kern
elif kernel_type == 4 or kernel_type == 5:
p0 = list()
p = list()
D = kernel[2]
for n in kernel[0]:
p0.append(params0[n])
p.append(params[n])
kern = statistics.getPdfMultinormal(p0, D[str(params0)], p)
kern = kern / auxilliary
return kern
def perturbParticle(params, priors, kernel, kernel_type, special_cases):
np = len(priors)
prior_prob = 1
if special_cases == 1:
# this is the case where kernel is uniform and all priors are uniform
ind = 0
for n in kernel[0]:
lflag = (params[n] + kernel[2][ind][0]) < priors[n][1]
uflag = (params[n] + kernel[2][ind][1]) > priors[n][2]
lower = kernel[2][ind][0]
upper = kernel[2][ind][1]
if lflag == True:
lower = -(params[n] - priors[n][1])
if uflag == True:
upper = priors[n][2] - params[n]
delta = 0
positive = False
if lflag == False and uflag == False:
# proceed as normal
delta = rnd.uniform(low=kernel[2][ind][0],
high=kernel[2][ind][1])
else:
# decide if the particle is to be perturbed positively or negatively
positive = rnd.uniform(0,
1) > abs(lower) / (abs(lower) + upper)
if positive == True:
# theta = theta + U(0, min(prior,kernel) )
delta = rnd.uniform(low=0, high=upper)
else:
# theta = theta + U( max(prior,kernel), 0 )
delta = rnd.uniform(low=lower, high=0)
params[n] = params[n] + delta
ind += 1
# this is not the actaul value of the pdf but we only require it to be non zero
return 1.0
else:
if kernel_type == 1:
ind = 0
# n refers to the index of the parameter (integer between 0 and np-1)
# ind is an integer between 0 and len(kernel[0])-1 which enables to determine the kernel to use
for n in kernel[0]:
params[n] = params[n] + rnd.uniform(low=kernel[2][ind][0],
high=kernel[2][ind][1])
ind += 1
if kernel_type == 2:
ind = 0
# n refers to the index of the parameter (integer between 0 and np-1)
# ind is an integer between 0 and len(kernel[0])-1 which enables to determine the kernel to use
for n in kernel[0]:
params[n] = rnd.normal(params[n], numpy.sqrt(kernel[2][ind]))
ind += 1
if kernel_type == 3:
mean = list()
for n in kernel[0]:
mean.append(params[n])
tmp = statistics.mvnd_gen(mean, kernel[2])
ind = 0
for n in kernel[0]:
params[n] = tmp[ind]
ind = ind + 1
if (kernel_type == 4 or kernel_type == 5):
mean = list()
for n in kernel[0]:
mean.append(params[n])
D = kernel[2]
tmp = statistics.mvnd_gen(mean, D[str(params)])
ind = 0
for n in kernel[0]:
params[n] = tmp[ind]
ind = ind + 1
# compute the likelihood
prior_prob = 1
for n in range(np):
x = 1.0
#if priors[n][0]==1:
# x=statistics.getPdfGauss(priors[n][1], numpy.sqrt(priors[n][2]), params[n])
# if we do not care about the value of prior_prob, then here: x=1.0
if priors[n][0] == 2:
x = statistics.getPdfUniform(priors[n][1], priors[n][2],
params[n])
#if priors[n][0]==3:
# x=statistics.getPdfLognormal(priors[n][1],priors[n][2],params[n])
# if we do not care about the value of prior_prob, then here: x=1.0 if params[n]>=0 and 0 otherwise
prior_prob = prior_prob * x
return prior_prob
def perturbParticle(params, priors, kernel, kernel_type, special_cases):
np = len(priors)
prior_prob = 1
if special_cases == 1:
# this is the case where kernel is uniform and all priors are uniform
ind = 0
for n in kernel[0]:
lflag = (params[n] + kernel[2][ind][0]) < priors[n][1]
uflag = (params[n] + kernel[2][ind][1]) > priors[n][2]
lower = kernel[2][ind][0]
upper = kernel[2][ind][1]
if lflag == True:
lower = -(params[n] - priors[n][1])
if uflag == True:
upper = priors[n][2] - params[n]
delta = 0
positive = False
if lflag == False and uflag == False:
# proceed as normal
delta = rnd.uniform(low=kernel[2][ind][0], high=kernel[2][ind][1])
else:
# decide if the particle is to be perturbed positively or negatively
positive = rnd.uniform(0, 1) > abs(lower) / (abs(lower) + upper)
if positive == True:
# theta = theta + U(0, min(prior,kernel) )
delta = rnd.uniform(low=0, high=upper)
else:
# theta = theta + U( max(prior,kernel), 0 )
delta = rnd.uniform(low=lower, high=0)
params[n] = params[n] + delta
ind += 1
# this is not the actaul value of the pdf but we only require it to be non zero
return 1.0
else:
if kernel_type == 1:
ind = 0
# n refers to the index of the parameter (integer between 0 and np-1)
# ind is an integer between 0 and len(kernel[0])-1 which enables to determine the kernel to use
for n in kernel[0]:
params[n] = params[n] + rnd.uniform(low=kernel[2][ind][0], high=kernel[2][ind][1])
ind += 1
if kernel_type == 2:
ind = 0
# n refers to the index of the parameter (integer between 0 and np-1)
# ind is an integer between 0 and len(kernel[0])-1 which enables to determine the kernel to use
for n in kernel[0]:
params[n] = rnd.normal(params[n], numpy.sqrt(kernel[2][ind]))
ind += 1
if kernel_type == 3:
mean = list()
for n in kernel[0]:
mean.append(params[n])
tmp = statistics.mvnd_gen(mean, kernel[2])
ind = 0
for n in kernel[0]:
params[n] = tmp[ind]
ind = ind + 1
if kernel_type == 4 or kernel_type == 5:
mean = list()
for n in kernel[0]:
mean.append(params[n])
D = kernel[2]
tmp = statistics.mvnd_gen(mean, D[str(params)])
ind = 0
for n in kernel[0]:
params[n] = tmp[ind]
ind = ind + 1
# compute the likelihood
prior_prob = 1
for n in range(np):
x = 1.0
# if priors[n][0]==1:
# x=statistics.getPdfGauss(priors[n][1], numpy.sqrt(priors[n][2]), params[n])
# if we do not care about the value of prior_prob, then here: x=1.0
if priors[n][0] == 2:
x = statistics.getPdfUniform(priors[n][1], priors[n][2], params[n])
# if priors[n][0]==3:
# x=statistics.getPdfLognormal(priors[n][1],priors[n][2],params[n])
# if we do not care about the value of prior_prob, then here: x=1.0 if params[n]>=0 and 0 otherwise
prior_prob = prior_prob * x
return prior_prob
def computeParticleWeights(self):
if self.debug == 2: print "\t***computeParticleWeights"
for k in range(self.nparticles):
this_model = self.model_curr[k]
this_param = self.parameters_curr[k]
# calculate model prior probility
mprob = self.modelprior[this_model]
np = len(self.parameters_curr[k])
# particle prior probability
pprob = 1
for n in range(0, np):
x = 1.0
if self.models[this_model].prior[n][0] == 0:
x = 1
if self.models[this_model].prior[n][0] == 1:
x = statistics.getPdfGauss(
self.models[this_model].prior[n][1],
numpy.sqrt(self.models[this_model].prior[n][2]),
this_param[n])
if self.models[this_model].prior[n][0] == 2:
x = statistics.getPdfUniform(
self.models[this_model].prior[n][1],
self.models[this_model].prior[n][2], this_param[n])
if self.models[this_model].prior[n][0] == 3:
x = statistics.getPdfLognormal(
self.models[this_model].prior[n][1],
numpy.sqrt(self.models[this_model].prior[n][2]),
this_param[n])
pprob = pprob * x
numer = self.b[k] * mprob * pprob
denom_m = 0
for i in range(self.nmodel):
denom_m = denom_m + self.margins_prev[i] * getPdfModelKernel(
this_model, i, self.modelKernel, self.nmodel,
self.dead_models)
denom = 0
# print "Calculating denom\t", selected_model, sampleParameters
for j in range(self.nparticles):
if (int(this_model) == int(self.model_prev[j])):
# print "\t", j, model_prev[j], weights_prev[j], parameters_prev[j]
if self.debug == 2:
print "\tj, weights_prev, kernelpdf", j, self.weights_prev[
j],
self.kernelpdffn(this_param, self.parameters_prev[j],
self.models[this_model].prior,
self.kernels[this_model],
self.kernel_aux[j], self.kernel_type)
denom = denom + self.weights_prev[j] * self.kernelpdffn(
this_param, self.parameters_prev[j], self.
models[this_model].prior, self.kernels[this_model],
self.kernel_aux[j], self.kernel_type)
if self.debug == 2:
print "\tnumer/denom_m/denom/m(t-1) : ", numer, denom_m, denom, self.margins_prev[
this_model]
self.weights_curr[k] = numer / (denom_m * denom /
self.margins_prev[this_model])
