draft Neighbors2; fix bug in Neighbors1
authorBenjamin Auder <benjamin.auder@somewhere>
Mon, 27 Mar 2017 01:22:55 +0000 (03:22 +0200)
committerBenjamin Auder <benjamin.auder@somewhere>
Mon, 27 Mar 2017 01:22:55 +0000 (03:22 +0200)
pkg/R/F_Neighbors.R
pkg/R/F_Neighbors2.R [new file with mode: 0644]

index 5b1f826..600c5c8 100644 (file)
@@ -103,7 +103,7 @@ NeighborsForecaster = R6::R6Class("NeighborsForecaster",
                                        # Require at least half of non-NA common values to compute the distance
                                        if ( !any( is.na(delta) ) )
                                                distances2[i] = mean(delta^2)
-                               }
+                               Centered}
 
                                sd_dist = sd(distances2)
                                if (sd_dist < .Machine$double.eps)
@@ -171,7 +171,7 @@ NeighborsForecaster = R6::R6Class("NeighborsForecaster",
 
                        prediction = rep(0, horizon)
                        for (i in seq_along(fdays))
-                               prediction = prediction + similarities[i] * data$getSerie(fdays[i]+1)[1:horizon]
+                               prediction = prediction + similarities[i] * data$getCenteredSerie(fdays[i]+1)[1:horizon]
                        prediction = prediction / sum(similarities, na.rm=TRUE)
 
                        if (final_call)
diff --git a/pkg/R/F_Neighbors2.R b/pkg/R/F_Neighbors2.R
new file mode 100644 (file)
index 0000000..e6addde
--- /dev/null
@@ -0,0 +1,149 @@
+#' @include Forecaster.R
+#'
+#' Neighbors2 Forecaster
+#'
+#' Predict tomorrow as a weighted combination of "futures of the past" days.
+#' Inherits \code{\link{Forecaster}}
+#'
+Neighbors2Forecaster = R6::R6Class("Neighbors2Forecaster",
+       inherit = Forecaster,
+
+       public = list(
+               predictShape = function(data, today, memory, horizon, ...)
+               {
+                       # (re)initialize computed parameters
+                       private$.params <- list("weights"=NA, "indices"=NA, "window"=NA)
+
+                       # Do not forecast on days with NAs (TODO: softer condition...)
+                       if (any(is.na(data$getCenteredSerie(today))))
+                               return (NA)
+
+                       # Determine indices of no-NAs days followed by no-NAs tomorrows
+                       fdays = getNoNA2(data, max(today-memory,1), today-1)
+
+                       # Get optional args
+                       kernel = ifelse(hasArg("kernel"), list(...)$kernel, "Gauss") #or "Epan"
+                       if (hasArg(h_window))
+                       {
+                               return ( private$.predictShapeAux(data,
+                                       fdays, today, horizon, list(...)$h_window, kernel, TRUE) )
+                       }
+
+
+                       # Indices of similar days for cross-validation; TODO: 45 = magic number
+                       # TODO: ici faut une sorte de "same_season==TRUE" --> mois similaires epandage
+                       sdays = getSimilarDaysIndices(today, limit=45, same_season=FALSE)
+
+
+                       # Function to optimize h : h |--> sum of prediction errors on last 45 "similar" days
+                       errorOnLastNdays = function(h, kernel)
+                       {
+                               error = 0
+                               nb_jours = 0
+                               for (i in intersect(fdays,sdays))
+                               {
+                                       # mix_strategy is never used here (simtype != "mix"), therefore left blank
+                                       prediction = private$.predictShapeAux(data, fdays, i, horizon, h, kernel, FALSE)
+                                       if (!is.na(prediction[1]))
+                                       {
+                                               nb_jours = nb_jours + 1
+                                               error = error +
+                                                       mean((data$getSerie(i+1)[1:horizon] - prediction)^2)
+                                       }
+                               }
+                               return (error / nb_jours)
+                       }
+
+                       # h :: only for endo in this variation
+                       h_best_endo = optimize(errorOnLastNdays, c(0,10), kernel=kernel)$minimum
+
+                       return (private$.predictShapeAux(data, fdays, today, horizon, h_best, kernel, TRUE))
+               }
+       ),
+       private = list(
+               # Precondition: "today" is full (no NAs)
+               .predictShapeAux = function(data, fdays, today, horizon, h, kernel, final_call)
+               {
+                       fdays = fdays[ fdays < today ]
+                       # TODO: 3 = magic number
+                       if (length(fdays) < 3)
+                               return (NA)
+
+                       # ENDO:: Distances from last observed day to days in the past
+                       distances2 = rep(NA, length(fdays))
+                       for (i in seq_along(fdays))
+                       {
+                               delta = data$getSerie(today) - data$getSerie(fdays[i])
+                               # Require at least half of non-NA common values to compute the distance
+                               if ( !any( is.na(delta) ) )
+                                       distances2[i] = mean(delta^2)
+                       }
+
+                       sd_dist = sd(distances2)
+                       if (sd_dist < .Machine$double.eps)
+                       {
+#                                      warning("All computed distances are very close: stdev too small")
+                               sd_dist = 1 #mostly for tests... FIXME:
+                       }
+                       simils_endo =
+                               if (kernel=="Gauss")
+                                       exp(-distances2/(sd_dist*h_endo^2))
+                               else
+                               {
+                                       # Epanechnikov
+                                       u = 1 - distances2/(sd_dist*h_endo^2)
+                                       u[abs(u)>1] = 0.
+                                       u
+                               }
+
+                       # EXOGENS: distances computations are enough
+                       # TODO: search among similar concentrations....... at this stage ?!
+                       M = matrix( nrow=1+length(fdays), ncol=1+length(data$getExo(today)) )
+                       M[1,] = c( data$getLevel(today), as.double(data$getExo(today)) )
+                       for (i in seq_along(fdays))
+                               M[i+1,] = c( data$getLevel(fdays[i]), as.double(data$getExo(fdays[i])) )
+
+                       sigma = cov(M) #NOTE: robust covariance is way too slow
+                       sigma_inv = solve(sigma) #TODO: use pseudo-inverse if needed?
+
+                       # Distances from last observed day to days in the past
+                       distances2 = rep(NA, nrow(M)-1)
+                       for (i in 2:nrow(M))
+                       {
+                               delta = M[1,] - M[i,]
+                               distances2[i-1] = delta %*% sigma_inv %*% delta
+                       }
+
+                       ppv <- sort(distances2, index.return=TRUE)$ix[1:10] #..............
+#PPV pour endo ?
+
+                       similarities =
+                               if (simtype == "exo")
+                                       simils_exo
+                               else if (simtype == "endo")
+                                       simils_endo
+                               else #mix
+                                       simils_endo * simils_exo
+
+                       prediction = rep(0, horizon)
+                       for (i in seq_along(fdays))
+                               prediction = prediction + similarities[i] * data$getSerie(fdays[i]+1)[1:horizon]
+                       prediction = prediction / sum(similarities, na.rm=TRUE)
+
+                       if (final_call)
+                       {
+                               private$.params$weights <- similarities
+                               private$.params$indices <- fdays
+                               private$.params$window <-
+                                       if (simtype=="endo")
+                                               h_endo
+                                       else if (simtype=="exo")
+                                               h_exo
+                                       else #mix
+                                               c(h_endo,h_exo)
+                       }
+
+                       return (prediction)
+               }
+       )
+)