R/CoorGateP.R

In fchamroukhi/HDME: LASSO Regularized Mixture of Experts Models

CoorGateP = function(X, wk, tau, Gamma, rho)
{#using Proximal Newton method
# source("Fs.R")
# source("CoorLQk.R")
# source("Pik.R")
n = dim(X)[1]
K = ncol(tau)
P_k = c(rep(0,n))
d_k = c(rep(0,n))
c_k = c(rep(0,n))
Stepsize = 0.5
esp = 10^-5 #threshold for Q value
wk_new = wk
repeat
{
  wk_old = wk_new
  Q_old = Fs(X, tau, Gamma, rho, wk_old)
  for(k in 1:(K-1))
  { #First: compute the quadratic approximation w.r.t (w_k): L_Qk
    P = Pik(n, K, X, wk_new) #value of Pi_k
    P_k = P[,k]
    d_k = P_k*(1-P_k)
    c_k = X%*%(as.matrix(wk_new[k,]))+(tau[,k]-P_k)/d_k
    #Second: coordinate descent for maximizing L_Qk
    wk_new[k,] = CoorLQk(X, c_k, d_k, wk_new[k,], Gamma[k], rho)
  }
  Q_new = Fs(X, tau, Gamma, rho, wk_new)
  #-------------BACKTRACKING LINE SEARCH
  t = 1
  while (Q_new < Q_old)
  {
    t = t*Stepsize
    wk_new = wk_new*t + wk_old*(1-t)
    Q_new = Fs(X, tau, Gamma, rho, wk_new)
  }
  if((Q_new - Q_old) < esp) break
}
return(wk_new)
}