def cost(self, x):
u = self.forward(x)
diff = (self.obs - u)
negative_log_likelihood = 0.5 * np.dot(
np.dot(diff, self.Gamma_noise_inv), diff)
controls = self.unflatten(x)
penalty_controls = 0. * np.sum(
np.mean(ewm(
1.,
np.power(
np.abs(np.diff(controls[0], axis=0)) -
np.diff(controls[0], axis=0), 1.)),
axis=0))
penalty_controls += np.sum(
np.mean(ewm(1., np.power(np.abs(np.diff(controls[0], axis=0)),
1.)),
axis=0))
# # # penalization on HFR, beta, delta, kappa, sigma, eta_I, eta_Q, mu, gamma_I, gamma_A, gamma_H, gamma_Q from mean
# result = 0.
# for j in range(0, len(parameters)):
# result = result + 0.5/500 * np.sum(np.power(np.divide(np.abs(controls[j+1] - parameters[j]), parameters[j]/2), 1))
#
# penalty_controls += result
print("cost, penalty = ", negative_log_likelihood,
10. * penalty_controls)
return negative_log_likelihood + 10. * penalty_controls
def penalization(controls, coeffs, kappas):
result = 0.
# # penalization on alpha, q, tau from decreasing
# increasing = [0.01, 1., 1.]
# for i in range(1):
# result = result + increasing[i]*np.sum(np.mean(ewm(coeffs[i][1:], np.power(np.abs(np.diff(controls[i], axis=0))-np.diff(controls[i], axis=0), 1.)), axis=0))
result = np.sum(np.mean(ewm(coeffs[0][1:], np.power(np.abs(np.diff(controls[0], axis=0))-np.diff(controls[0], axis=0), 1.)), axis=0))
# result = result + 1.*np.sum(np.mean(ewm(coeffs[3][1:], np.power(np.abs(np.diff(controls[3], axis=0))+np.diff(controls[3], axis=0), 1.)), axis=0))
# # # penalization on HFR, beta, delta, kappa, sigma, eta_I, eta_Q, mu, gamma_I, gamma_A, gamma_H, gamma_Q from mean
# for j in range(2, len(parameters)):
# result = result + 0.5/500 * np.sum(np.power(np.divide(controls[j+3] - parameters[j], parameters[j]/2), 2))
return result
# # data from China, https://science.sciencemag.org/content/sci/suppl/2020/03/13/science.abb3221.DC1/abb3221_Li_SM_rev.pdf
# theta = 0.5
# # data from EU https://www.imperial.ac.uk/media/imperial-college/medicine/mrc-gida/2020-04-23-COVID19-Report-16.pdf
# theta = 0.1 # a big range investigated from 0 - 1
# ratio of pre-isolation infectious period, only in action if quarantine is used
# data from EU, https://www.imperial.ac.uk/media/imperial-college/medicine/mrc-gida/2020-04-23-COVID19-Report-16.pdf
delta = 0.5
# #### initial seeding and simulation time step
# proportion of each age and risk group
population_proportion = 1.
# # susceptible population in each age and risk group
S = np.array([ewm(N_total, population_proportion)])
# exposed at the beginning
E = np.array([100.])
# S, E, Q, A, I, H, R, D, Tc, Tu # Tc/Tu is the total reported/unreported positive cases
zeros = np.zeros(1)
# Tc = np.ones(1) * data_confirmed[0]
y0 = np.array([S, E, zeros, zeros, zeros, zeros, zeros, zeros, zeros, zeros]).flatten("C")
# # set data from solution to fitted death data
# data = np.load("data/find_initial.npz")
# solution_opt = data["solution_opt"]
# today = data["today"]
# y0 = solution_opt[today, :]
# #### control variable for social distancing, isolution, and quarantine, in [0, 1], 0 means no control, 1 full control
def seir(self, y, t, parameters, controls, stochastic=False):
# define the right hand side of the ode systems given state y, time t, parameters, and controls
if self.number_group > 1:
y = y.reshape((10, self.number_group))
S, E, Q, A, I, H, R, D, Tc, Tu = y
q, tau, HFR, kappa, beta, delta, sigma, eta_I, eta_Q, mu, gamma_I, gamma_A, gamma_H, gamma_Q = parameters
# _, _, _, delta, sigma, eta_I, eta_Q, mu, gamma_I, gamma_A, gamma_H, gamma_Q = parameters
alpha, = controls
alpha = (np.tanh(alpha) + 1) / 2
alpha = self.interpolation(t, self.t_control, alpha)
# alpha, q, tau, HFR, kappa = [self.interpolation(t, self.t_control, controls[i]) for i in range(self.number_time_dependent_controls)]
# tau_p = self.interpolation(t - np.max(1./sigma), self.t_control, controls[2])
# tau_p = self.interpolation(t, self.t_control, controls[2])
# tau_p = tau
# IHR = np.divide(kappa, np.max(kappa) + tau_p)
IHR = kappa
QHR = ewm(tau, IHR)
# gamma_A = gamma_I
pi = self.proportion2factor(IHR, eta_I, gamma_I)
nu = self.proportion2factor(HFR, mu, gamma_H)
rho = self.proportion2factor(QHR, eta_Q, gamma_Q)
contact = self.contact_rate(t)
# theta_I = 2 - tau
# theta_A = 1 - tau
theta_I = 1. - 0 * tau
theta_A = 1. - 0 * tau
delta = 1. + 0 * delta
C_E = ewm(
1 - alpha,
ewm(
1 - q,
ewm(delta,
np.dot(contact, ewm(theta_I, np.divide(I, self.N_total)))))
+ np.dot(contact, ewm(theta_A, np.divide(A, self.N_total))))
C_Q = ewm(
1 - alpha,
ewm(
q,
ewm(delta,
np.dot(contact, ewm(theta_I, np.divide(I,
self.N_total))))))
if stochastic:
zeros = np.zeros(np.size(S))
S = np.max([zeros, S], axis=0)
E = np.max([zeros, E], axis=0)
Q = np.max([zeros, Q], axis=0)
A = np.max([zeros, A], axis=0)
I = np.max([zeros, I], axis=0)
H = np.max([zeros, H], axis=0)
P1 = ewm(beta, ewm(C_E, S))
P2 = ewm(beta, ewm(C_Q, S))
P3 = ewm(tau, ewm(sigma, E))
P4 = ewm(1 - tau, ewm(sigma, E))
P5 = ewm(rho, ewm(eta_Q, Q))
P6 = ewm(1 - rho, ewm(gamma_Q, Q))
P7 = ewm(gamma_A, A)
P8 = ewm(pi, ewm(eta_I, I))
P9 = ewm(1 - pi, ewm(gamma_I, I))
P10 = ewm(nu, ewm(mu, H))
P11 = ewm(1 - nu, ewm(gamma_H, H))
if stochastic:
P1 = np.random.poisson(P1)
P2 = np.random.poisson(P2)
P3 = np.random.poisson(P3)
P4 = np.random.poisson(P4)
P5 = np.random.poisson(P5)
P6 = np.random.poisson(P6)
P7 = np.random.poisson(P7)
P8 = np.random.poisson(P8)
P9 = np.random.poisson(P9)
P10 = np.random.poisson(P10)
P11 = np.random.poisson(P11)
dS = -P1 - P2
dE = P1 - P3 - P4
dQ = P2 - P5 - P6
dA = P4 - P7
dI = P3 - P8 - P9
dH = P8 + P5 - P10 - P11
dR = P7 + P9 + P11 + P6
dD = P10
dTc = P3 + P2 # + quarantined, P2
dTu = P4
dydt = np.array([dS, dE, dQ, dA, dI, dH, dR, dD, dTc,
dTu]).flatten("C")
return dydt
def proportion2factor(self, proportion, r1, r2):
# define a function to transform proportion to factor, e.g., from prop_EI to tau with r1=sigma_I, r2 = sigma_A
factor = np.divide(ewm(r2, proportion), r1 + ewm(r2 - r1, proportion))
return factor
def reproduction(self, t, parameters, controls, solution):
# effective reproduction number at time t
# https://royalsocietypublishing.org/doi/pdf/10.1098/rsif.2009.0386
# The construction of next-generation matrices for compartmental epidemic models
t_set = t
number_time = len(t)
if number_time == 1:
t_set = [t]
Rt = np.zeros(number_time)
for n_t in range(number_time):
t = t_set[n_t]
q, tau, HFR, kappa, beta, delta, sigma, eta_I, eta_Q, mu, gamma_I, gamma_A, gamma_H, gamma_Q = parameters
# alpha, q, tau, HFR, kappa, beta, delta, sigma, eta_I, eta_Q, mu, gamma_I, gamma_A, gamma_H, gamma_Q = controls
# alpha, q, tau, HFR, kappa, beta, _, _, _, _, _, _, _, _, _ = controls
alpha = self.interpolation(t, self.t_control, controls)
# tau_p = self.interpolation(t - np.max(1./sigma), self.t_control, controls[2])
# tau_p = self.interpolation(t, self.t_control, controls[2])
# IHR = np.divide(kappa, np.max(kappa) + tau_p)
IHR = kappa
QHR = ewm(tau, IHR)
pi = self.proportion2factor(IHR, eta_I, gamma_I)
# transform the parameter in the vector format
ones = np.ones(self.number_group)
alpha, q, tau, pi, beta, delta, kappa, sigma, eta_I, eta_Q, mu, gamma_I, gamma_A, gamma_H, gamma_Q = \
ewm(alpha,ones), ewm(q,ones), ewm(tau,ones), ewm(pi,ones), ewm(beta,ones), ewm(delta,ones), ewm(kappa,ones), \
ewm(sigma,ones), ewm(eta_I,ones), ewm(eta_Q,ones), ewm(mu,ones), ewm(gamma_I,ones), ewm(gamma_A,ones), ewm(gamma_H,ones), ewm(gamma_Q,ones)
# theta_I = 2 - tau
# theta_A = 1 - tau
theta_I = 1. - 0 * tau
theta_A = 1. - 0 * tau
delta = 1. + 0 * delta
S = solution[n_t, :self.number_group]
zeros = np.zeros((self.number_group, self.number_group))
ItoS = np.diag(beta) * np.diag(1.-alpha) * np.diag(1.-q) * np.diag(delta) * np.dot(self.contact_rate(t), np.diag(theta_I)) * np.diag(np.divide(S, self.N_total))
AtoS = np.diag(beta) * np.diag(1.-alpha) * np.dot(self.contact_rate(t), np.diag(theta_A)) * np.diag(np.divide(S, self.N_total))
T = np.block([
[zeros, AtoS, ItoS],
[zeros, zeros, zeros],
[zeros, zeros, zeros]
])
Sigma = np.block([
[-np.diag(sigma), zeros, zeros],
[np.diag(1.-tau)*np.diag(sigma), -np.diag(gamma_A), zeros],
[np.diag(tau)*np.diag(sigma), zeros, -np.diag(pi)*np.diag(eta_I)-np.diag(1.-pi)*np.diag(gamma_I)]
])
w, _ = np.linalg.eig(-np.dot(T, np.linalg.inv(Sigma)))
Rt[n_t] = np.max(w)
return Rt
# # data from China, https://science.sciencemag.org/content/sci/suppl/2020/03/13/science.abb3221.DC1/abb3221_Li_SM_rev.pdf
# theta = 0.5
# # data from EU https://www.imperial.ac.uk/media/imperial-college/medicine/mrc-gida/2020-04-23-COVID19-Report-16.pdf
# theta = 0.1 # a big range investigated from 0 - 1
# ratio of pre-isolation infectious period, only in action if quarantine is used
# data from EU, https://www.imperial.ac.uk/media/imperial-college/medicine/mrc-gida/2020-04-23-COVID19-Report-16.pdf
delta = 0.5
# #### initial seeding and simulation time step
# proportion of each age and risk group
population_proportion = 1.
# # susceptible population in each age and risk group
S = np.array([ewm(N_total, population_proportion)])
# exposed at the beginning
E = np.array([100.])
# S, E, Q, A, I, H, R, D, Tc, Tu # Tc/Tu is the total reported/unreported positive cases
zeros = np.zeros(1)
# Tc = np.ones(1) * data_confirmed[0]
y0 = np.array(
[S, E, zeros, zeros, zeros, zeros, zeros, zeros, zeros,
zeros]).flatten("C")
# # set data from solution to fitted death data
# data = np.load("data/find_initial.npz")
# solution_opt = data["solution_opt"]
# today = data["today"]
# y0 = solution_opt[today, :]
