| 1 | constructionModelesLassoMLE = function(phiInit,rhoInit,piInit,gamInit,mini,maxi,gamma,glambda, |
| 2 | X,Y,seuil,tau,A1,A2) |
| 3 | { |
| 4 | n = dim(X)[1]; |
| 5 | p = dim(phiInit)[1] |
| 6 | m = dim(phiInit)[2] |
| 7 | k = dim(phiInit)[3] |
| 8 | L = length(glambda) |
| 9 | |
| 10 | #output parameters |
| 11 | phi = array(0, dim=c(p,m,k,L)) |
| 12 | rho = array(0, dim=c(m,m,k,L)) |
| 13 | pi = matrix(0, k, L) |
| 14 | llh = matrix(0, L, 2) #log-likelihood |
| 15 | |
| 16 | for(lambdaIndex in 1:L) |
| 17 | { |
| 18 | a = A1[,1,lambdaIndex] |
| 19 | a = a[a!=0] |
| 20 | if(length(a)==0) |
| 21 | next |
| 22 | |
| 23 | res = EMGLLF(phiInit[a,,],rhoInit,piInit,gamInit,mini,maxi,gamma,0.,X[,a],Y,tau) |
| 24 | |
| 25 | for (j in 1:length(a)) |
| 26 | phi[a[j],,,lambdaIndex] = res$phi[j,,] |
| 27 | rho[,,,lambdaIndex] = res$rho |
| 28 | pi[,lambdaIndex] = res$pi |
| 29 | |
| 30 | dimension = 0 |
| 31 | for (j in 1:p) |
| 32 | { |
| 33 | b = A2[j,2:dim(A2)[2],lambdaIndex] |
| 34 | b = b[b!=0] |
| 35 | if (length(b) > 0) |
| 36 | phi[A2[j,1,lambdaIndex],b,,lambdaIndex] = 0. |
| 37 | c = A1[j,2:dim(A1)[2],lambdaIndex] |
| 38 | dimension = dimension + sum(c!=0) |
| 39 | } |
| 40 | |
| 41 | #on veut calculer l'EMV avec toutes nos estimations |
| 42 | densite = matrix(0, nrow=n, ncol=L) |
| 43 | for (i in 1:n) |
| 44 | { |
| 45 | for (r in 1:k) |
| 46 | { |
| 47 | delta = Y[i,]%*%rho[,,r,lambdaIndex] - (X[i,a]%*%phi[a,,r,lambdaIndex]); |
| 48 | densite[i,lambdaIndex] = densite[i,lambdaIndex] + pi[r,lambdaIndex] * |
| 49 | det(rho[,,r,lambdaIndex])/(sqrt(2*base::pi))^m * exp(-tcrossprod(delta)/2.0) |
| 50 | } |
| 51 | } |
| 52 | llh[lambdaIndex,1] = sum(log(densite[,lambdaIndex])) |
| 53 | llh[lambdaIndex,2] = (dimension+m+1)*k-1 |
| 54 | } |
| 55 | return (list("phi"=phi, "rho"=rho, "pi"=pi, "llh" = llh)) |
| 56 | } |