Skip to content
Snippets Groups Projects
rls_optim.R 4.87 KiB
Newer Older
  • Learn to ignore specific revisions
  • pbac's avatar
    pbac committed
    # Do this in a separate file to see the generated help:
    #library(devtools)
    #document()
    #load_all(as.package("../../onlineforecast"))
    #?rls_optim
    
    
    #' Optimize parameters (transformation stage) of RLS model
    #'
    #' This is a wrapper for \code{\link{optim}} to enable easy use of bounds and caching in the optimization.
    #' 
    #' @title Optimize parameters for onlineforecast model fitted with RLS
    #' @param model The onlineforecast model, including inputs, output, kseq, p
    #' @param data The data.list including the variables used in the model.
    #' @param scorefun The function to be score used for calculating the score to be optimized.
    #' @param cachedir A character specifying the path (and prefix) of the cache file name. If set to \code{""}, then no cache will be loaded or written.
    #' @param printout A logical determining if the score function is printed out in each iteration of the optimization.
    #' @param method The method argument for \code{\link{optim}}.
    #' @param ... Additional parameters to \code{\link{optim}}
    #' @return Result object of optim().
    #' Parameters resulting from the optimization can be found from \code{result$par}
    #' @seealso \code{link{optim}} for how to control the optimization.
    #' @examples
    #'
    #' # Take data (See vignette ??(ref) for better model and more details)
    
    pbac's avatar
    pbac committed
    #' D <- subset(Dbuilding, c("2010-12-15", "2011-01-01"))
    
    pbac's avatar
    pbac committed
    #' D$y <- D$heatload
    #' # Define a model 
    #' model <- forecastmodel$new()
    #' model$add_inputs(Ta = "Ta", mu = "ones()")
    #' model$add_regprm("rls_prm(lambda=0.99)")
    #'
    #' # Before fitting the model, define which points to include in the evaluation of the score function
    #' D$scoreperiod <- in_range("2010-12-20", D$t)
    #' # And the sequence of horizons to fit for
    #' model$kseq <- 1:6
    #' # Now we can fit the model and get the score, as it is
    #' rls_fit(model=model, data=D, scorefun=rmse, returnanalysis=FALSE)
    #' # Or we can change the lambda
    #' rls_fit(c(lambda=0.9), model, D, rmse, returnanalysis=FALSE)
    #'
    
    #' # This could be passed to optim() (or any optimizer, see forecastmodel$insert_prm()).
    #' optim(c(lambda=0.98), rls_fit, model=model, data=D, scorefun=rmse, returnanalysis=FALSE,
    #'       lower=c(lambda=0.9), upper=c(lambda=0.999), method="L-BFGS-B")
    
    pbac's avatar
    pbac committed
    #'
    #' # rls_optim is simply a helper, it's makes using bounds easiere and enables caching of the results
    #' # First add bounds for lambda (lower, init, upper)
    #' model$add_prmbounds(lambda = c(0.9, 0.98, 0.999))
    #'
    #' # Now the same optimization as above can be done by
    #' val <- rls_optim(model, D)
    #' val
    #' 
    #' # Caching can be done by providing a path (try rerunning and see the file in "cache" folder)
    #' val <- rls_optim(model, D, cachedir="cache")
    #' val
    #'
    #' # If anything affecting the results are changed, then the cache is not loaded
    #' model$add_prmbounds(lambda = c(0.89, 0.98, 0.999))
    #' val <- rls_optim(model, D, cachedir="cache")
    #' 
    #' # To delete the cache
    #' file.remove(dir("cache", full.names=TRUE))
    #' file.remove("cache")
    #' 
    #' @export
    rls_optim <- function(model, data, scorefun = rmse, cachedir="", printout=TRUE, method="L-BFGS-B", ...){
        # Take the parameters bounds from the parameter bounds set in the model
        init <- model$get_prmbounds("init")
        lower <- model$get_prmbounds("lower")
        upper <- model$get_prmbounds("upper")
        # If bounds are NA, then set
        if(any(is.na(lower))){ lower[is.na(lower)] <- -Inf}
        if(any(is.na(upper))){ lower[is.na(upper)] <- Inf}
    
        # Caching the results based on some of the function arguments
        if(cachedir != ""){
            # Have to insert the parameters in the expressions to get the right state of the model for unique checksum
            model$insert_prm(init)
            # Give all the elements needed to calculate the unique cache name
            # This is maybe smarter, don't have to calculate the transformation of the data: cnm <- cache_name(model$regprm, getse(model$inputs, nms="expr"), model$output, model$prmbounds, model$kseq, data, objfun, init, lower, upper, cachedir = cachedir)
            # Have to reset the state first to remove dependency of previous calls
            model$reset_state()
            cnm <- cache_name(rls_fit, rls_optim, model$outputrange, model$regprm, model$transform_data(data), data[[model$output]], scorefun, init, lower, upper, cachedir = cachedir)
            # Maybe load the cached result
            if(file.exists(cnm)){ return(readRDS(cnm)) }
        }
    
        # Run the optimization
        res <- optim(par = init,
                     fn = rls_fit,
                     # Parameters to pass to rls_fit
                     model = model,
                     data = data,
                     scorefun = scorefun,
                     printout = printout,
                     returnanalysis = FALSE,
                     # Parameters to pass to optim
                     lower = lower,
                     upper = upper,
                     method =  method,
                     ...)
        
        # Save the result in the cachedir
        if(cachedir != ""){ cache_save(res, cnm) }
        # Return the result
        return(res)
    }