svres: Support Vector Regression with Evolutionary Strategy
In ForAI: Forecasting with Artificial Intelligence

Description Usage Arguments Details Value Author(s) References Examples

svres is a hybrid algorithm which combines support vector regression and evolutionary strategy (uncorrelated mutation with p step sizes) to build predictive models.

  svres(X.train, Y.train, X.test, PercentValid = 20, Generations = 500, InitialGamma = 0.001, ErrorFunc = MSE, Step = FALSE, StepBoruta = FALSE, SplitRandom = FALSE, Trace = TRUE, dTrace = 10, earlyStop = 50, kfold = 1)

`X.train`	Data matrix (numeric) containing the input values (predictors) used to train the model.
`Y.train`	Response vector (numeric) used to train the model.
`X.test`	Data matrix (numeric) containing the input values (predictors) used to test the model.
`PercentValid`	Percentage of the data reserved for validation. Default is `20`%.
`Generations`	Number of generations considered to train the model. Default is `500` generations.
`InitialGamma`	Initial gamma hyperparameter. Default is `0.001`.
`ErrorFunc`	Error function to be minimized. The default is the function `MSE` from package `qualV` but the function could be customized.
`Step`	Option whether to use the stepwise regression to prescreening the input variables. Default is `FALSE`.
`StepBoruta`	Option whether to use the Boruta selection to prescreening the input variables. Default is `FALSE`.
`SplitRandom`	Option whether to split the train set randomly. Default is `FALSE`.
`Trace`	If `TRUE`, information is printed during the running of `svres`. Default is `TRUE`.
`dTrace`	Frequency which the information is pritend during the running of `svres`. Default is `10`.
`earlyStop`	Stopping criterium for the evolutionary algorithm using the number of consecutive generations without improvements. Default is `50`.
`kfold`	if a integer value k>1 is specified, a k-fold cross validation on the training data is performed.

To achieve better results, the use of a pre-processing technique (e.g. standardization of variables) is important. If there are multiple minima, try different values of InitialGamma.

`error.svm`	`TRUE` Hyper-parameter of the best trained model.
`forecast`	A vector of predicted values generated by the best trained model.
`svmf`	An object of class "svm" containing the fitted model.
`ffTrain`	Error value of the training based on `ErrorFunc`. (It is `NULL` if `kfold > 1`)
`ffValid`	Error value of the validation based on `ErrorFunc`.
`stepvars`	Variables selected by the prescreening methods when they are used.

Aranildo R. Lima and Alex J. Cannon

Aranildo R. Lima, Alex J. Cannon, William W. Hsieh, Nonlinear regression in environmental sciences by support vector machines combined with evolutionary strategy, Computers & Geosciences, Volume 50, January 2013, Pages 136-144, ISSN 0098-3004, http://www.sciencedirect.com/science/article/pii/S0098300412002269

Cherkassky, V. and Ma, Y., 2004. Practical selection of SVM parameters and noise estimation for SVM regression. Neural Netw. 17, 113-126.

     ## necessary libraries
     library(qualV)
     library(e1071)
     library(FNN)

      ## generating sets
      x1 <- rnorm(1000)
      x2 <- x1^2
      y  <- x1+x2

      x.fit <- cbind(x1,x2)[1:700,]
      x.test <- cbind(x1,x2)[701:1000,]
      y.fit <- y[1:700]
      y.test <- y[701:1000]

      ## running SVR-ES
      resul <- svres(x.fit, y.fit, x.test)

      ## points are the target and lines are the forecasting
      plot(y.test[1:100])
      lines(resul$forecast[1:100])