X-Git-Url: https://git.auder.net/%3C?a=blobdiff_plain;f=pkg%2FR%2Fm_ExponentialWeights.R;fp=pkg%2FR%2Fm_ExponentialWeights.R;h=e59e28f524d016837e8f0747bb64977a83e58963;hb=357b8be00388d07c04c10d4bf7f503fd947185d2;hp=0000000000000000000000000000000000000000;hpb=a961f8a15492bf4b18d24bc117358d1f412dd078;p=aggexp.git

diff --git a/pkg/R/m_ExponentialWeights.R b/pkg/R/m_ExponentialWeights.R
new file mode 100644
index 0000000..e59e28f
--- /dev/null
+++ b/pkg/R/m_ExponentialWeights.R
@@ -0,0 +1,55 @@
+#' @include b_LinearAlgorithm.R
+
+#' @title Exponential Weights Algorithm
+#'
+#' @description Exponential Weights Algorithm.
+#' Inherits \code{\link{LinearAlgorithm}}
+#'
+#' @field alpha Importance of weights redistribution, in [0,1]. Default: 0
+#' @field grad Whether to use or not the (sub)gradient trick. Default: FALSE
+#'
+ExponentialWeights = setRefClass(
+	Class = "ExponentialWeights",
+
+	fields = c(
+		alpha = "numeric",
+		grad = "logical"
+	),
+
+	contains = "LinearAlgorithm",
+
+	methods = list(
+		initialize = function(...)
+		{
+			callSuper(...)
+			if (length(alpha) == 0 || alpha < 0. || alpha > 1.)
+				alpha <<- 0. #no redistribution
+			if (length(grad) == 0)
+				grad <<- FALSE
+		},
+		predict_noNA = function(XY, x)
+		{
+			K = ncol(XY) - 1
+			if (K == 1)
+			{
+				#shortcut: nothing to combine
+				finalWeight = 1.
+			}
+
+			else
+			{
+				X = XY[,names(XY) != "Measure"]
+				Y = XY[,"Measure"]
+				finalWeight = .C("ew_predict_noNA", X = as.double(t(X)), Y = as.double(Y), n = as.integer(nrow(XY)), 
+					K = as.integer(K), alpha=as.double(alpha), grad = as.integer(grad), weight=double(K))$weight
+			}
+
+			appendWeight(finalWeight)
+			return (matricize(x) %*% finalWeight)
+#			M = matricize(x)
+#			M[M<=30] = -1
+#			M[M>30] = 1
+#			return (30 + M%*%finalWeight)
+		}
+	)
+)