replace "AF_ConVar" by "AF_UEI".

DESCRIPTION

@@ -1,7 +1,7 @@
 Package: EPBO
 Title: Bayesian Optimization via Exact Penalty
-Version: 0.0.9
-Date: 2023-11-03
+Version: 0.0.11
+Date: 2023-11-20
 Authors@R: 
   person(given = "Jiangyan", 
          family = "Zhao", 

NAMESPACE

@@ -11,6 +11,8 @@ export(AF_ScaledEI_Kernel)
 export(AF_ScaledEI_MC)
 export(AF_TS)
 export(AF_UEI)
+export(AF_UEI_Kernel)
+export(AF_UEI_MC)
 export(AF_VEI)
 export(bilog)
 export(normalize)

R/AF_ConVar.R

@@ -47,14 +47,13 @@ AF_ConVar = function(x, fgpi, Cgpi, equal)
     mu_C = pred_C$mean
     sigma_C = sqrt(pred_C$s2)
     if(equal[j]){
-      LCB[,j] = abs(mu_C) - 3*sigma_C
+      LCB[,j] = abs(mu_C) - 6*sigma_C
     }else{
-      LCB[,j] = mu_C - 3*sigma_C
+      LCB[,j] = mu_C - 6*sigma_C
     }
   }
 
   ## Acquaisition function
-  AF = sigma_f
   infeasible = apply(LCB > 0, 1, any)
   if(all(infeasible)){
     AF = sigma_f

R/AF_LCB.R

@@ -9,8 +9,6 @@
 #' 
 #' @param Cgpi description
 #' 
-#' @param epbest description
-#' 
 #' @param rho description
 #' 
 #' @param equal an optional vector containing zeros and ones, whose length equals the number of
@@ -37,7 +35,7 @@
 #' 
 #' 
 
-AF_LCB = function(x, fgpi, fmean, fsd, Cgpi, Cmean, Csd, epbest, rho, equal)
+AF_LCB = function(x, fgpi, fmean, fsd, Cgpi, rho, equal)
 {
   if(is.null(nrow(x))) x = matrix(x, nrow=1)
   ncand = nrow(x) # number of the candidate points
@@ -52,15 +50,15 @@ AF_LCB = function(x, fgpi, fmean, fsd, Cgpi, Cmean, Csd, epbest, rho, equal)
   mu_C = sigma_C = omega = matrix(NA, nc, ncand)
   for (j in 1:nc) {
     pred_C = predGPsep(Cgpi[j], x, lite=TRUE)
-    mu_C[j,] = pred_C$mean * Csd[j] + Cmean[j]
-    sigma_C[j,] = sqrt(pred_C$s2) * Csd[j]
+    mu_C[j,] = pred_C$mean
+    sigma_C[j,] = sqrt(pred_C$s2)
     omega[j,] = (equal[j]+1)*pnorm(mu_C[j,]/sigma_C[j,]) - equal[j]
   }
 
   ## Acquaisition function
   mu_ep = mu_f + rho%*%(omega*mu_C)
   sigma_ep = sqrt(sigma_f^2 + (rho^2)%*%((omega*sigma_C)^2))
-  LCB = mu_ep - sigma_ep
+  LCB = mu_ep - 3*sigma_ep
   return(LCB)
 }
 

R/AF_UEI.R

@@ -36,15 +36,15 @@
 #' 
 #' 
 
-AF_UEI = function(x, fgpi, Cgpi, epbest, rho, equal)
+AF_UEI = function(x, fgpi, fmean, fsd, Cgpi, epbest, rho, equal, beta=3)
 {
   if(is.null(nrow(x))) x = matrix(x, nrow=1)
   ncand = nrow(x) # number of the candidate points
 
   ## objective
   pred_f = predGPsep(fgpi, x, lite=TRUE)
-  mu_f = pred_f$mean
-  sigma_f = sqrt(pred_f$s2)
+  mu_f = pred_f$mean * fsd + fmean
+  sigma_f = sqrt(pred_f$s2) * fsd
 
   ## constraints
   nc = length(Cgpi) # number of the constraint
@@ -62,7 +62,8 @@ AF_UEI = function(x, fgpi, Cgpi, epbest, rho, equal)
   d = (epbest - mu_ep)/sigma_ep
   EI = sigma_ep * (d*pnorm(d) + dnorm(d)) # expected improvement
   VI = sigma_ep^2 * ((d^2+1)*pnorm(d) + d*dnorm(d)) - EI^2 # variance of the improvement (remove sigma_ep)
-  VI = pmax(0, VI)
-  UEI = EI + 2*sqrt(VI) # Scaled expected improvement
+  VI = pmax(.Machine$double.xmin, VI)
+  UEI = EI + beta*sqrt(VI)
+  EI[is.nan(UEI)] = 0
   return(UEI)
 }

R/AF_UEI_Kernel.R

@@ -0,0 +1,69 @@
+#' @title UEI acquisition function
+#' 
+#' @description Scaled expected improvement
+#' 
+#' @param x description
+#' 
+#' @param fgpi description
+#' 
+#' @param Cgpi description
+#' 
+#' @param epbest description
+#' 
+#' @param rho description
+#' 
+#' @param equal an optional vector containing zeros and ones, whose length equals the number of
+#' constraints, specifying which should be treated as equality constraints (\code{1}) and 
+#' which as inequality (\code{0}) 
+#' 
+#' 
+#' @returns AF 
+#' 
+#' @author Jiangyan Zhao \email{[email protected]}
+#' 
+#' @references Noe, U. and D. Husmeier (2018). On a new improvement-based acquisition function 
+#' for Bayesian optimization. \emph{arXiv:1808.06918}.
+#' 
+#' @import laGP
+#' @importFrom stats dnorm 
+#' @importFrom stats pnorm 
+#' 
+#' 
+#' @export
+#' 
+#' @examples 
+#' B = rbind(c(0, 1), c(0, 1)) 
+#' 
+#' 
+
+AF_UEI_Kernel = function(x, fgpi, fmean, fsd, Cgpi, epbest, rho, equal, beta=3, type="UK")
+{
+  if(is.null(nrow(x))) x = matrix(x, nrow=1)
+  ncand = nrow(x) # number of the candidate points
+
+  ## objective
+  pred_f = predict(object=fgpi, newdata=x, type=type, checkNames = FALSE, light.return = TRUE)
+  mu_f = pred_f$mean * fsd + fmean
+  sigma_f = pred_f$sd * fsd
+
+  ## constraints
+  nc = length(Cgpi) # number of the constraint
+  mu_C = sigma_C = omega = matrix(NA, nc, ncand)
+  for (j in 1:nc) {
+    pred_C = predict(object=Cgpi[[j]], newdata=x, type=type, checkNames = FALSE, light.return = TRUE)
+    mu_C[j,] = pred_C$mean
+    sigma_C[j,] = pred_C$sd
+    omega[j,] = (equal[j]+1)*pnorm(mu_C[j,]/sigma_C[j,]) - equal[j]
+  }
+
+  ## Acquaisition function
+  mu_ep = mu_f + rho%*%(omega*mu_C)
+  sigma_ep = sqrt(sigma_f^2 + (rho^2)%*%((omega*sigma_C)^2))
+  d = (epbest - mu_ep)/sigma_ep
+  EI = sigma_ep * (d*pnorm(d) + dnorm(d)) # expected improvement
+  VI = sigma_ep^2 * ((d^2+1)*pnorm(d) + d*dnorm(d)) - EI^2 # variance of the improvement (remove sigma_ep)
+  VI = pmax(.Machine$double.xmin, VI)
+  UEI = EI + beta*sqrt(VI)
+  UEI[is.nan(UEI)] = 0
+  return(UEI)
+}

R/AF_UEI_MC.R

@@ -0,0 +1,83 @@
+#' @title UEI acquisition function
+#' 
+#' @description Scaled expected improvement
+#' 
+#' @param x description
+#' 
+#' @param fgpi description
+#' 
+#' @param fmean description
+#' 
+#' @param fsd description
+#' 
+#' @param Cgpi description
+#' 
+#' @param epbest description
+#' 
+#' @param rho description
+#' 
+#' @param equal an optional vector containing zeros and ones, whose length equals the number of
+#' constraints, specifying which should be treated as equality constraints (\code{1}) and 
+#' which as inequality (\code{0}) 
+#' 
+#' @param N description
+#' 
+#' @param beta description
+#' 
+#' @returns AF 
+#' 
+#' @author Jiangyan Zhao \email{[email protected]}
+#' 
+#' @references Noe, U. and D. Husmeier (2018). On a new improvement-based acquisition function 
+#' for Bayesian optimization. \emph{arXiv:1808.06918}.
+#' 
+#' @import laGP
+#' @importFrom stats dnorm 
+#' @importFrom stats pnorm 
+#' 
+#' 
+#' @export
+#' 
+#' @examples 
+#' B = rbind(c(0, 1), c(0, 1)) 
+#' 
+#' 
+
+AF_UEI_MC = function(x, fgpi, fmean, fsd, Cgpi, epbest, rho, equal, N=1000, beta=3)
+{
+  if(is.null(nrow(x))) x = matrix(x, nrow=1)
+  ncand = nrow(x) # number of the candidate points
+
+  ## objective
+  pred_f = predGPsep(fgpi, x, lite=TRUE)
+  mu_f = pred_f$mean * fsd + fmean
+  sigma_f = sqrt(pred_f$s2) * fsd
+
+  ## constraints
+  nc = length(Cgpi) # number of the constraint
+  mu_C = sigma_C = matrix(NA, nc, ncand)
+  for (j in 1:nc) {
+    pred_C = predGPsep(Cgpi[j], x, lite=TRUE)
+    mu_C[j,] = pred_C$mean
+    sigma_C[j,] = sqrt(pred_C$s2)
+  }
+
+  ## expected improvement
+  EP = matrix(NA, nrow = N, ncol = ncand)
+  for (n in 1:N) {
+    EP[n,] = rnorm(ncand, mu_f, sigma_f)
+    for (j in 1:nc) {
+      if(equal[j]){
+        EP[n,] = EP[n,] + rho[j]*abs(rnorm(ncand, mu_C[j,], sigma_C[j,]))
+      }else{
+        EP[n,] = EP[n,] + rho[j]*pmax(0, rnorm(ncand, mu_C[j,], sigma_C[j,]))
+      }
+    }
+  }
+  improvement = matrix(pmax(0, epbest-EP), nrow = N)
+  EI = colMeans(improvement) # expected improvement
+  SI = colSds(improvement)   # Standard deviation of the improvement
+  UEI = EI + beta * SI
+  EI[is.nan(UEI)] = 0
+  return(UEI)
+}