learn.models <- function(obj, learner, learner.pars){
  percents <- list()
  
  # for PERCENTS .1,.2, .3, .4, .5
  for(i in 1:length(obj$data)){
    pred <- real <- list()
    
    reg.data <- obj$data[[i]]$reg.data
    
    # for each 30 data variations
    for(j in 1:length(reg.data)){
      aux.data <- reg.data[[j]]
      ind.na <- which(is.na(aux.data[,"target"]))
      test.data <- aux.data[ind.na,-(1:3)] # filter x, y, target
      train.data <- aux.data[-ind.na,-(1:2)] # filter x, y
      
      # learn model
      model <- train.model(train.data, learner, learner.pars)
      # predict test data, excluding 1 column which contains NA
      pred[[length(pred)+1]] <- predict(model, data.frame(test.data))

      # get real data
      real[[length(real)+1]] <- mapply(function(x,y) obj$orig.data[x,y], aux.data[ind.na, "x"], aux.data[ind.na, "y"]) # get x, y of the NA
    }#end for 30 data variation
    percents[[length(percents)+1]] <- list(pred=pred, real=real)
    
  }#end for PERCENTS
  percents
}

train.model <- function(data,learner,learner.pars){
  if(any(is.na(data)))
    stop("data has NAs")
  
  cat('*')
  model <- NULL

  new.pars <- list()	
  if(learner %in% c("rpart","rpartXse")){
    new.pars$form <- formula(paste(colnames(data)[1]," ~ .",sep=""))	
  }else{
    new.pars$formula <- formula(paste(colnames(data)[1]," ~ .",sep=""))
  }
  new.pars$data <- data.frame(data)
  new.pars <- c(new.pars,learner.pars) 
  # trainning
  model <- do.call(learner,new.pars)
  
  model
}#.train.model

# in the future put in other file
# tree.x0 <- tune.model(data=x0, learner="rpartXse", params=list(se=c(0,.5,1,1.5)))
# svm.x0 <- tune.model(data=x0, learner="svm", params=list(cost=c(1,5,10,50,100),gamma=c(.001,.01,.05,.1)))
# rf.x0 <- tune.model(data=x0, learner="randomForest", params=list(ntree=c(500,1000,1500)))
tune.model <- function(data, learner, params){
  model <- list()
	
  pars <- permuta.list(params)
  for(i in 1:length(pars)){
    p <- pars[[i]]
    if(DEBUG)
      print(p)
    model[[paste(learner, print.list(p))]] <- learn.models(obj=data,learner=learner,learner.pars=p)
  }
  model
}


# m1.x0 <- learn.base.models(obj=x0, type="m1")
# m2.x0 <- learn.base.models(obj=x0, type="m2")
# m3.x0 <- learn.base.models(obj=x0, type="m3")
# w1.x0 <- learn.base.models(obj=x0, type="w1")
# w2.x0 <- learn.base.models(obj=x0, type="w2")
# w3.x0 <- learn.base.models(obj=x0, type="w3")
learn.base.models <- function(obj, type="m1"){
  if(! type %in% c("m1","m2","m3","w1","w2","w3"))
    stop("type must be one of ")
  
  percents <- list()
  
  # for PERCENTS .1,.2, .3, .4, .5
  for(i in 1:length(obj$data)){
    pred <- real <- list()
    
    reg.data <- obj$data[[i]]$reg.data
    
    # for each 30 data variations
    for(j in 1:length(reg.data)){
      aux.data <- reg.data[[j]]
      ind.na <- which(is.na(aux.data[,"target"]))

      # selecting the corresponding column...
      pred[[length(pred)+1]] <- aux.data[ind.na,type] 

      # get real data
      ind.real <- aux.data[ind.na, c("x","y")] # get x, y of the NA
      real[[length(real)+1]] <- mapply(function(x,y) obj$orig.data[x,y], ind.real[,1], ind.real[,2])
    }#end for 30 data variation
    percents[[length(percents)+1]] <- list(pred=pred, real=real)
    
  }#end for PERCENTS
  percents
}


learn.autokrige <- function(obj, nmax=NULL){
  percents <- list()
  coordIn <- obj$allCoordIn
  
  # for PERCENTS .1,.2, .3, .4, .5
  for(i in 1:length(obj$data)){
    cat(i,'0%\n')
    pred <- real <- kcs <- list()

    gaps <- obj$data[[i]]$gaps
    
    # for each 30 data variations
    for(j in 1:length(gaps)){
      cat(j,"\n")
      gap <- gaps[[j]]

      z <- cbind(merge(1:nrow(obj$orig.data),1:ncol(obj$orig.data)), gap)
      
      colnames(z) <- c("x","y","valor")

      ind.na <- which(is.na(z$valor))
      loci <- as.data.frame(z[-ind.na,])
      loci$x <- as.numeric(loci$x)
      loci$y <- as.numeric(loci$y)
      na.data <- z[ind.na, c("x","y")]
      na.data$x <- as.numeric(na.data$x)
      na.data$y <- as.numeric(na.data$y)
      # prepara for krige
      coordinates(loci) <- ~ x + y
      coordinates(na.data) <- ~ x + y

      # calcula o krige
      # adicionei o verbose, pra imprimir algumas infos dos modelos krige treinados
      # adicionei o nmax por causa de erros nos modelos x2, x7 e x9. Restringe o numero de vizinhos nas previsões...
      if(is.null(nmax))
        nmax <- Inf
      
      kc <- autoKrige(formula=valor~1, input_data=loci, new_data=na.data, verbose=T, nmax=nmax)
      # predict 
      pred[[as.character(j)]] <- kc$krige_output$var1.pred
      # real data
      real[[as.character(j)]] <- obj$vect.orig.data[ind.na] #mapply(function(x,y) obj$orig.data[x,y], na.data[,"x"], na.data[,"y"])
      kcs[[as.character(j)]] <- kc
    }#end for 30 data variation
    percents[[as.character(i)]] <- list(pred=pred, real=real, kcs=kcs)    
  }#end for PERCENTS
  percents
}