class BayesOptParticleFiltering(BayesOptContinuous):
def __init__(self,
idx_simul,
low_b,
upp_b,
nb_it,
nb_ini,
refine,
load_save_flag,
max_loglkd=None):
if refine == True:
assert (max_loglkd is not None)
self.max_loglkd = max_loglkd
string = """
library(pomp)
square <- function(x) {
return(x^2)
}
pf_r <- function(Y, r = 0.1, sigma = 0.1, tau = 0.1)
{
gompertz.proc.sim <- function(x, t, params, delta.t,...) {
eps <- exp(rnorm(n = 1, mean = 0, sd = params["sigma"]))
S <- exp(-params["r"] * delta.t)
setNames(params["K"]^(1 - S) * x["X"]^S * eps, "X")
}
gompertz.meas.sim <- function(x, t, params, ...) {
setNames(rlnorm(n = 1, meanlog = log(x["X"]), sd = params["tau"]), "Y")
}
gompertz.meas.dens <- function(y, x, t, params, log, ...) {
dlnorm(x = y["Y"], meanlog = log(x["X"]), sdlog = params["tau"],
log = log)
}
gompertz <- pomp(data = data.frame(time = 1:100, Y = Y), times = "time",
rprocess = discrete.time.sim(step.fun = gompertz.proc.sim, delta.t = 1),
rmeasure = gompertz.meas.sim, t0 = 0)
# set parameters
theta <- c(r = r, K = 1, sigma = sigma, tau = tau, X.0 = 1)
############# INFERENCE #############
gompertz <- pomp(gompertz, dmeasure = gompertz.meas.dens)
# lkd estimator with pf
pf <- pfilter(gompertz, params = theta, Np = 20)
loglik.truth <- logLik(pf)
return(loglik.truth)
}
"""
self.n_dim = 3
self.refine = refine
self.idx_simul = idx_simul
BayesOptContinuous.__init__(self, self.n_dim)
self.powerpack = SignatureTranslatedAnonymousPackage(
string, "powerpack")
self.params = {}
self.params['n_iterations'] = nb_it # 200
self.params['n_init_samples'] = nb_ini # 100
self.params['n_iter_relearn'] = 1
self.params['verbose_level'] = 2
if self.refine:
self.params['noise'] = .01
self.params['crit_name'] = "cLCB"
else:
self.params['noise'] = 1
self.params['crit_name'] = "cEI"
self.params['l_type'] = "L_EMPIRICAL"
self.params['mean.name'] = "mZero"
self.params['l_all'] = True
self.params['sc_type'] = "SC_MTL"
self.params['surr_name'] = "sGaussianProcessML"
self.params['load_save_flag'] = load_save_flag
self.params['force_jump'] = 5
if load_save_flag == 2 and refine == 1:
self.params['save_filename'] = "logs/bayesopt_" + str(
idx_simul) + "_N_20_refine{0}_tmp.dat".format(refine * 1)
self.params['load_filename'] = "logs/bayesopt_" + str(
idx_simul) + "_N_20_refine{0}_tmp.dat".format(refine * 1)
else:
self.params['save_filename'] = "logs/bayesopt_" + str(
idx_simul) + "_N_20_refine{0}.dat".format(refine * 1)
self.params['load_filename'] = "logs/bayesopt_" + str(
idx_simul) + "_N_20_refine{0}.dat".format(refine * 1)
self.params['kernel_name'] = "kMaternARD5"
self.params['n_inner_iterations'] = 1000 # 500
self.params['lower_bound'] = low_b
self.params['upper_bound'] = upp_b
self.lower_bound = self.params['lower_bound']
self.upper_bound = self.params['upper_bound']
self.ub = self.upper_bound
self.lb = self.lower_bound
self.Y = get_Y(path='../simulations/simulation{0}.csv'.format(
self.idx_simul),
T=100)
print('Rough likelihood estimate is {0}'.format(
self.powerpack.pf_r(robjects.FloatVector(self.Y), .1, .1, .1)[0]))
def evaluateSample(self, Xin):
r = .5 * (10**Xin[0] - 10**0.) / (10**1. - 10**0.)
sigma = .5 * (10**Xin[1] - 10**0.) / (10**1. - 10**0.)
tau = .5 * (10**Xin[2] - 10**0.) / (10**1. - 10**0.)
print('parameters are : {0}, {1}, {2}'.format(r, sigma, tau))
if self.refine:
return -np.exp(
self.powerpack.pf_r(robjects.FloatVector(self.Y), r, sigma,
tau)[0] - self.max_loglkd)
else:
return -self.powerpack.pf_r(robjects.FloatVector(self.Y), r, sigma,
tau)[0]
gompertz <- pomp(gompertz, dmeasure = gompertz.meas.dens)
# lkd estimator with pf
pf <- pfilter(gompertz, params = theta, Np = 1000)
loglik.truth <- logLik(pf)
return(loglik.truth)
}
"""
def get_Y(path='arguments.csv', T=100):
warnings.warn('Only one space for csv')
output = []
csvfile = open(path, 'rb')
spamreader = csv.reader(csvfile, delimiter=' ', quotechar='|')
spamreader.next()
for line in spamreader:
output += line
output = np.asarray(output, dtype=np.float).ravel()
assert (len(output) == T)
return output
powerpack = SignatureTranslatedAnonymousPackage(string, "powerpack")
Y = get_Y()
lkd = powerpack.pf_r(robjects.FloatVector(Y))[0]
# rsum = robjects.r['square.R']
# r = robjects.r
# r.source()[0][0]
np.asarray(
powerpack.traj_r(
'N_20/iterated_filtering_4{0}'.format(idx_simul)))[:, 4][:-1])
trajectories_tau[1, idx_simul - 1] = np.exp(
np.asarray(
powerpack.traj_r(
'N_20/iterated_filtering_4{0}'.format(idx_simul)))[:, 5][:-1])
#indexes = np.asarray(get_traj_index('N_20/python_bayesopt{0}_4.dat'.format(idx_simul)))[:-1]
trajectories_r[0, idx_simul - 1] = get_params(
'N_20/python_bayesopt{0}_4.dat'.format(idx_simul))[0]
trajectories_sig[0, idx_simul - 1] = get_params(
'N_20/python_bayesopt{0}_4.dat'.format(idx_simul))[1]
trajectories_tau[0, idx_simul - 1] = get_params(
'N_20/python_bayesopt{0}_4.dat'.format(idx_simul))[2]
likelihoods[0, idx_simul - 1] = powerpack.pf_r(
robjects.FloatVector(Y), r, sigma, tau
)[0] # -info[0] # get_lkd('python_bayesopt_1{0}.dat'.format(idx_simul)) #
likelihoods[1, idx_simul - 1] = powerpack.pf_r(
robjects.FloatVector(Y), trajectories_r[1, idx_simul - 1, -1],
trajectories_sig[1, idx_simul - 1,
-1], trajectories_tau[1, idx_simul - 1, -1]
)[0] #powerpack.res_r(idx_simul)[0] # powerpack.res_r(idx_simul)[0] #
def to_normalized_weights(logWeights):
b = np.max(logWeights)
weights = [np.exp(logw - b) for logw in logWeights]
return weights / sum(weights)
idx_sim = 1
