Almost first draft for optim with W (still Compute_Omega to write)
authorBenjamin Auder <benjamin.auder@somewhere>
Sun, 8 Dec 2019 10:29:57 +0000 (11:29 +0100)
committerBenjamin Auder <benjamin.auder@somewhere>
Sun, 8 Dec 2019 10:29:57 +0000 (11:29 +0100)
pkg/R/optimParams.R
pkg/src/functions.c
reports/local_run.sh [new file with mode: 0644]

index 948167b..934a757 100644 (file)
@@ -12,8 +12,8 @@
 #'     \item β: regression matrix, size dxK
 #'     \item b: intercepts, size K
 #'   }
-#'   x0 is a vector containing respectively the K-1 first elements of p, then β by
-#'   columns, and finally b: \code{x0 = c(p[1:(K-1)],as.double(β),b)}.
+#'   θ0 is a vector containing respectively the K-1 first elements of p, then β by
+#'   columns, and finally b: \code{θ0 = c(p[1:(K-1)],as.double(β),b)}.
 #'
 #' @seealso \code{multiRun} to estimate statistics based on β, and
 #'   \code{generateSampleIO} for I/O random generation.
 #' io = generateSampleIO(10000, 1/2, matrix(c(1,-2,3,1),ncol=2), c(0,0), "logit")
 #' μ = computeMu(io$X, io$Y, list(K=2))
 #' o <- optimParams(io$X, io$Y, 2, "logit")
-#' x0 <- list(p=1/2, β=μ, b=c(0,0))
-#' par0 <- o$run(x0)
+#' θ0 <- list(p=1/2, β=μ, b=c(0,0))
+#' par0 <- o$run(θ0)
 #' # Compare with another starting point
-#' x1 <- list(p=1/2, β=2*μ, b=c(0,0))
-#' par1 <- o$run(x1)
+#' θ1 <- list(p=1/2, β=2*μ, b=c(0,0))
+#' par1 <- o$run(θ1)
 #' o$f( o$linArgs(par0) )
 #' o$f( o$linArgs(par1) )
 #' @export
@@ -67,9 +67,7 @@ setRefClass(
                li = "character", #link function
                X = "matrix",
                Y = "numeric",
-               M1 = "numeric",
-               M2 = "numeric", #M2 easier to process as a vector
-               M3 = "numeric", #same for M3
+    Mhat = "numeric", #vector of empirical moments
                # Dimensions
                K = "integer",
     n = "integer",
@@ -89,80 +87,90 @@ setRefClass(
 
       # Precompute empirical moments
       M <- computeMoments(optargs$X,optargs$Y)
-      M1 <<- as.double(M[[1]])
-      M2 <<- as.double(M[[2]])
-      M3 <<- as.double(M[[3]])
+      M1 <- as.double(M[[1]])
+      M2 <- as.double(M[[2]])
+      M3 <- as.double(M[[3]])
+      Mhat <<- matrix(c(M1,M2,M3), ncol=1)
 
                        n <<- nrow(X)
                        d <<- length(M1)
       W <<- diag(d+d^2+d^3) #initialize at W = Identity
                },
 
-               expArgs = function(x)
+               expArgs = function(v)
                {
-                       "Expand individual arguments from vector x"
+                       "Expand individual arguments from vector v into a list"
 
                        list(
                                # p: dimension K-1, need to be completed
-                               "p" = c(x[1:(K-1)], 1-sum(x[1:(K-1)])),
-                               "β" = matrix(x[K:(K+d*K-1)], ncol=K),
-                               "b" = x[(K+d*K):(K+(d+1)*K-1)])
+                               "p" = c(v[1:(K-1)], 1-sum(v[1:(K-1)])),
+                               "β" = matrix(v[K:(K+d*K-1)], ncol=K),
+                               "b" = v[(K+d*K):(K+(d+1)*K-1)])
                },
 
-               linArgs = function(o)
+               linArgs = function(L)
                {
-                       " Linearize vectors+matrices into a vector x"
+                       "Linearize vectors+matrices from list L into a vector"
 
-                       c(o$p[1:(K-1)], as.double(o$β), o$b)
+                       c(L$p[1:(K-1)], as.double(L$β), L$b)
                },
 
-    getOmega = function(theta)
+    computeW = function(θ)
     {
       dim <- d + d^2 + d^3
-      matrix( .C("Compute_Omega",
-        X=as.double(X), Y=as.double(Y), pn=as.integer(n), pd=as.integer(d),
-        p=as.double(theta$p), β=as.double(theta$β), b=as.double(theta$b),
-        W=as.double(W), PACKAGE="morpheus")$W, nrow=dim, ncol=dim)
+      W <<- solve( matrix( .C("Compute_Omega",
+        X=as.double(X), Y=as.double(Y), M=as.double(M(θ)),
+        pn=as.integer(n), pd=as.integer(d),
+        W=as.double(W), PACKAGE="morpheus")$W, nrow=dim, ncol=dim) )
+      NULL #avoid returning W
     },
 
-    f = function(theta)
+    M <- function(θ)
     {
-                       "Product t(Mi - hat_Mi) W (Mi - hat_Mi) with Mi(theta)"
+      "Vector of moments, of size d+d^2+d^3"
 
-      p <- theta$p
-                       β <- theta
+      p <- θ$p
+                       β <- θ
                        λ <- sqrt(colSums(β^2))
-                       b <- theta$b
+                       b <- θ$b
 
                        # Tensorial products β^2 = β2 and β^3 = β3 must be computed from current β1
                        β2 <- apply(β, 2, function(col) col %o% col)
                        β3 <- apply(β, 2, function(col) col %o% col %o% col)
 
-                       A <- matrix(c(
-                               β  %*% (p * .G(li,1,λ,b)) - M1,
-                               β2 %*% (p * .G(li,2,λ,b)) - M2,
-                               β3 %*% (p * .G(li,3,λ,b)) - M3), ncol=1)
+                       matrix(c(
+                               β  %*% (p * .G(li,1,λ,b)),
+                               β2 %*% (p * .G(li,2,λ,b)),
+                               β3 %*% (p * .G(li,3,λ,b))), ncol=1)
+    },
+
+    f = function(θ)
+    {
+                       "Product t(Mi - hat_Mi) W (Mi - hat_Mi) with Mi(theta)"
+
+                       A <- M(θ) - Mhat
       t(A) %*% W %*% A
     },
 
-               grad_f = function(x)
+               grad_f = function(θ)
                {
                        "Gradient of f, dimension (K-1) + d*K + K = (d+2)*K - 1"
 
-      # TODO: formula -2 t(grad M(theta)) . W . (Mhat - M(theta))
+      -2 * t(grad_M(θ)) %*% getW(θ) %*% (Mhat - M(θ))
     }
 
-    grad_M = function(theta)
+    grad_M = function(θ)
     {
-      # TODO: adapt code below for grad of d+d^2+d^3 vector of moments,
-      # instead of grad (sum(Mhat-M(theta)^2)) --> should be easier
+      "Gradient of the vector of moments, size (dim=)d+d^2+d^3 x K-1+K+d*K"
 
-      P <- expArgs(x)
-                       p <- P$p
-                       β <- P
+      L <- expArgs(θ)
+                       p <- L$p
+                       β <- L
                        λ <- sqrt(colSums(β^2))
                        μ <- sweep(β, 2, λ, '/')
-                       b <- P$b
+                       b <- L$b
+
+      res <- matrix(nrow=nrow(W), ncol=0)
 
                        # Tensorial products β^2 = β2 and β^3 = β3 must be computed from current β1
                        β2 <- apply(β, 2, function(col) col %o% col)
@@ -175,18 +183,13 @@ setRefClass(
                        G4 = .G(li,4,λ,b)
                        G5 = .G(li,5,λ,b)
 
-                       # (Mi - hat_Mi)^2 ' == (Mi - hat_Mi)' 2(Mi - hat_Mi) = Mi' Fi
-                       F1 = as.double( 2 * ( β  %*% (p * G1) - M1 ) )
-                       F2 = as.double( 2 * ( β2 %*% (p * G2) - M2 ) )
-                       F3 = as.double( 2 * ( β3 %*% (p * G3) - M3 ) )
-
+      # Gradient on p: K-1 columns, dim rows
                        km1 = 1:(K-1)
-                       grad <- #gradient on p
-                         t( sweep(as.matrix(β [,km1]), 2, G1[km1], '*') - G1[K] * β [,K] ) %*% F1 +
-                               t( sweep(as.matrix(β2[,km1]), 2, G2[km1], '*') - G2[K] * β2[,K] ) %*% F2 +
-                               t( sweep(as.matrix(β3[,km1]), 2, G3[km1], '*') - G3[K] * β3[,K] ) %*% F3
+                       res <- cbind(res, rbind(
+        t( sweep(as.matrix(β [,km1]), 2, G1[km1], '*') - G1[K] * β [,K] ),
+        t( sweep(as.matrix(β2[,km1]), 2, G2[km1], '*') - G2[K] * β2[,K] ),
+        t( sweep(as.matrix(β3[,km1]), 2, G3[km1], '*') - G3[K] * β3[,K] )))
 
-                       grad_β <- matrix(nrow=d, ncol=K)
                        for (i in 1:d)
                        {
                                # i determines the derivated matrix dβ[2,3]
@@ -213,46 +216,47 @@ setRefClass(
                                dβ3_right[block,] <- dβ3_right[block,] + β2
                                dβ3 <- dβ3_left + sweep(dβ3_right, 2, p * G3, '*')
 
-                               grad_β[i,] <- t(dβ) %*% F1 + t(dβ2) %*% F2 + t(dβ3) %*% F3
+                               res <- cbind(res, rbind(t(dβ), t(dβ2), t(dβ3)))
                        }
-                       grad <- c(grad, as.double(grad_β))
 
-                       grad = c(grad, #gradient on b
-                               t( sweep(β,  2, p * G2, '*') ) %*% F1 +
-                               t( sweep(β2, 2, p * G3, '*') ) %*% F2 +
-                               t( sweep(β3, 2, p * G4, '*') ) %*% F3 )
+      # Gradient on b
+                       res <- cbind(res, rbind(
+                               t( sweep(β,  2, p * G2, '*') ),
+                               t( sweep(β2, 2, p * G3, '*') ),
+                               t( sweep(β3, 2, p * G4, '*') )))
 
-                       grad
+                       res
                },
 
-    # TODO: rename x(0) into theta(0) --> θ
-               run = function(x0)
+               run = function(θ0)
                {
-                       "Run optimization from x0 with solver..."
-
-           if (!is.list(x0))
-                   stop("x0: list")
-      if (is.null(x0$β))
-        stop("At least x0$β must be provided")
-                       if (!is.matrix(x0$β) || any(is.na(x0$β)) || ncol(x0$β) != K)
-                               stop("x0$β: matrix, no NA, ncol == K")
-      if (is.null(x0$p))
-        x0$p = rep(1/K, K-1)
-      else if (length(x0$p) != K-1 || sum(x0$p) > 1)
-        stop("x0$p should contain positive integers and sum to < 1")
+                       "Run optimization from θ0 with solver..."
+
+           if (!is.list(θ0))
+                   stop("θ0: list")
+      if (is.null(θ0$β))
+        stop("At least θ0$β must be provided")
+                       if (!is.matrix(θ0$β) || any(is.na(θ0$β)) || ncol(θ0$β) != K)
+                               stop("θ0$β: matrix, no NA, ncol == K")
+      if (is.null(θ0$p))
+        θ0$p = rep(1/K, K-1)
+      else if (length(θ0$p) != K-1 || sum(θ0$p) > 1)
+        stop("θ0$p should contain positive integers and sum to < 1")
       # Next test = heuristic to detect missing b (when matrix is called "beta")
-      if (is.null(x0$b) || all(x0$b == x0$β))
-        x0$b = rep(0, K)
-      else if (any(is.na(x0$b)))
-        stop("x0$b cannot have missing values")
+      if (is.null(θ0$b) || all(θ0$b == θ0$β))
+        θ0$b = rep(0, K)
+      else if (any(is.na(θ0$b)))
+        stop("θ0$b cannot have missing values")
 
-                       op_res = constrOptim( linArgs(x0), .self$f, .self$grad_f,
+                       op_res = constrOptim( linArgs(θ0), .self$f, .self$grad_f,
                                ui=cbind(
                                        rbind( rep(-1,K-1), diag(K-1) ),
                                        matrix(0, nrow=K, ncol=(d+1)*K) ),
                                ci=c(-1,rep(0,K-1)) )
 
-      # We get a first non-trivial estimation of W: getOmega(theta)^{-1}
+      # debug:
+      print(computeW(expArgs(op_res$par)))
+      # We get a first non-trivial estimation of W
       # TODO: loop, this redefine f, so that we can call constrOptim again...
       # Stopping condition? N iterations? Delta <= ε ?
 
index 42bb134..41065bd 100644 (file)
@@ -54,9 +54,17 @@ void Moments_M3(double* X, double* Y, int* pn, int* pd, double* M3)
        }
 }
 
-void Compute_Omega(double* X, double* Y, int* pn, int* pd, double* W)
+void Compute_Omega(double* X, double* Y, double* M, int* pn, int* pd, double* W)
 {
+       int n=*pn, d=*pd;
+  //double* W = (double*)calloc(d+d*d+d*d*d,sizeof(double));
+
   // TODO: formula 1/N sum( t(g(Zi,theta)) g(Zi,theta) )
-  // = 1/N sum( t( (XiYi-...) - theta[i] ) ( ... ) )
+  // = 1/N sum( t( (XiYi-...) - M[i] ) ( ... ) )
   // --> similar to Moments_M2 and M3 above
+  for (int j=0; j<
+  for (int i=0; i<n; i++)
+  {
+    W[] += 
+  }
 }
diff --git a/reports/local_run.sh b/reports/local_run.sh
new file mode 100644 (file)
index 0000000..e2cff42
--- /dev/null
@@ -0,0 +1,12 @@
+#!/bin/bash
+
+N=100
+n=1e5
+nc=3
+nstart=5
+
+for d in 2 5 10; do
+       for link in "logit" "probit"; do
+               R --slave --args N=$N n=$n nc=$nc d=$d link=$link nstart=$nstart <multistart.R >out_${n}_${link}_${d}_${nstart} 2>&1
+       done
+done