Package 'sisal'

Title: Sequential Input Selection Algorithm
Description: Implements the SISAL algorithm by Tikka and Hollmén. It is a sequential backward selection algorithm which uses a linear model in a cross-validation setting. Starting from the full model, one variable at a time is removed based on the regression coefficients. From this set of models, a parsimonious (sparse) model is found by choosing the model with the smallest number of variables among those models where the validation error is smaller than a threshold. Also implements extensions which explore larger parts of the search space and/or use ridge regression instead of ordinary least squares.
Authors: Mikko Korpela [aut, cre]
Maintainer: Mikko Korpela <[email protected]>
License: GPL (>= 2)
Version: 0.49
Built: 2024-10-27 12:33:35 UTC
Source: CRAN

Help Index


sisal: Sequential input selection algorithm

Description

Implements the SISAL algorithm by Tikka and Hollmén. It is a sequential backward selection algorithm which uses a linear model in a cross-validation setting. Starting from the full model, one variable at a time is removed based on the regression coefficients. From this set of models, a parsimonious (sparse) model is found by choosing the model with the smallest number of variables among those models where the validation error is smaller than a threshold. Also implements extensions which explore larger parts of the search space and/or use ridge regression instead of ordinary least squares.

Details

Package: sisal
Depends: R (>= 3.1.2)
Imports: graphics, grDevices, grid, methods, stats, utils,
boot, lattice, mgcv, digest, R.matlab, R.methodsS3
Suggests: graph, Rgraphviz, testthat (>= 0.8)
License: GPL (>= 2)
LazyData: yes

Index:

bootMSE                 Bootstrap Estimate of Mean Squared Error Using
                        SISAL Object
dynTextGrob             Create Text with Changing Size
laggedData              Create Input Matrix and Output Vector for Time
                        Series Prediction
plot.sisal              Plotting Sequential Input Selection Results
plotSelected.sisal      Plotting Sets of Inputs Produced by Sequential
                        Input Selection
print.sisal             Printing Sequential Input Selection Objects
sisal                   Sequential Input Selection Algorithm (SISAL)
sisal-package           sisal: Sequential input selection algorithm in
                        R
sisalData               Download External Datasets for SISAL
sisalTable              Draw Table with Equally Sized Cells
summary.sisal           Summarizing Sequential Input Selection Results
testSisal               Testing the Sequential Input Selection
                        Algorithm
toy.learn               Toy Data for SISAL (Learning Set)
toy.test                Toy Data for SISAL (Test Set)
tsToy.learn             Toy Time Series Data for SISAL (Learning Set)
tsToy.test              Toy Time Series Data for SISAL (Test Set)

Run input selection on your own data with sisal. For demo purposes, use testSisal to run the algorithm on example data sets. After input selection, compute bootstrap MSE in test data with bootMSE.

Author(s)

Mikko Korpela [email protected]

References

Tikka, J. and Hollmén, J. (2008) Sequential input selection algorithm for long-term prediction of time series. Neurocomputing, 71(13–15):2604–2615.


Bootstrap Estimate of Mean Squared Error Using SISAL Object

Description

Using a linear model produced by sisal, computes a bootstrap estimate of MSE in test data.

Usage

bootMSE(object, dataset = NULL, R = 1000,
        inputs = c("L.f", "L.v", "full"),
        method = c("OLS", "magic"), standardize = "inherit",
        stepsAhead = NULL, noiseSd = NULL, verbose = 1, ...)

Arguments

object

an object of class "sisal", containing the results of input selection and the corresponding ordinary least squares and ridge regression models. Must be compatible with dataset. See ‘Details’.

dataset

dataset to work on. A character string, a numeric vector or a list with components "X" and "y". When the default value NULL is used, the function attempts to detect the dataset from attributes of object. See ‘Details’.

R

the number of bootstrap replicates. Usually a single positive integral number. See boot::boot.

inputs

a character string. Which set of input variables to use. Choices are "L.f" (smallest set with error under threshold), "L.v" (minimum validation error) and "full" (full model). See sisal.

method

a character string. "OLS" for ordinary least squares regression or "magic" for a ridge regression model with an automatically selected regularization parameter. See sisal.

standardize

"inherit" or a logical flag. If TRUE, standardizes the data to zero mean and unit variance. If FALSE, uses original data. If "inherit" (the default), the value of this argument is copied from object. This affects the scale of the results.

stepsAhead

If doing time series prediction, this indicates how many steps ahead to predict. A non-negative integral value or NULL. If NULL (the default), the value is copied from an attribute of object, put there by testSisal.

noiseSd

standard deviation of the noise to be added to the dependent variable when dataset is "toy". The noise is a saved dataset. Thus it is always identical, only scaled by noiseSd. If NULL (the default), the value is copied from object.

verbose

verbosity level. A single numeric value. If 0, output is disabled. If greater than 0, shows some information about what the function is doing. Currently there is only one non-zero verbosity level (the default).

...

arguments passed to boot::boot.

Details

Four types of values are supported in dataset.

  1. Use one of "laser", "poland", "toy" and "tsToy" to work on the test part of a dataset included in or specifically supported by the package. The first two options will load their respective datasets over a network connection. See sisalData, toy.test and tsToy.test.

  2. Use a numeric vector to work with time series data. The use of the "laser" and "poland" datasets is recognized. Loading the datasets in advance reduces unnecessary network traffic when doing multiple repeats with the same dataset.

  3. Use a list with a numeric matrix "X" and a numeric vector "y" to supply inputs "X" and output "y". This is appropriate when using your own data for something else than time series prediction based on past values of the same time series.

  4. Use NULL (the default value) for automatic detection of the dataset. This works if object was created with testSisal.

When using time series data, the names of the inputs used in object must match the regular expression "lag\.\d+", i.e. "lag" followed by a dot and an integer without spaces or any other formatting. This is automatically taken care of by laggedData and testSisal.

When using other than time series data, the user-supplied dataset must contain all the input variables used in the selected linear model (i.e. full model or a subset of inputs) of object.

Value

An object of class "boot", as returned by boot::boot.

Author(s)

Mikko Korpela

See Also

boot::boot, sisal, testSisal

Examples

foo <- testSisal(dataset="toy", Mtimes=10)
bootMSE(foo)

Create Text with Changing Size

Description

This function creates a text object. When drawn, its size changes automatically according to the space available.

Usage

dynTextGrob(label, x = 0.5, y = 0.5, width = 1, height = 1,
            default.units = "npc", just = c(0.5, 0.5),
            hjust = NULL, vjust = NULL, rot = 0, rotJust = TRUE,
            rotHjust = NULL, rotVjust = NULL, resize = TRUE,
            sizingWidth = NULL, sizingHeight = NULL,
            adjustJust = TRUE, takeMeasurements = FALSE,
            name = NULL, gp = gpar(), vp = NULL)

Arguments

label

a character or expression vector, or a list containing both character strings and mathematical expressions. These are the text items to be drawn.

x

a numeric vector or unit of x locations for the labels.

y

a numeric vector or unit of y locations for the labels.

width

the space available for the labels in the width direction of the viewport. Used for computing the fontsize.

height

the space available for the labels in the height direction of the viewport. Used for computing the fontsize.

default.units

default unit to use when dimensions or locations are unitless numbers. See unit.

just

a numeric or character vector with one or two elements for setting the same justification for all labels. See textGrob.

hjust

a numeric vector for setting horizontal justification of individual labels. If given, overrides just.

vjust

a numeric vector for setting vertical justification of individual labels. If given, overrides just.

rot

a numeric vector for setting the rotation angle of individual labels in degrees.

rotJust

a logical vector which affects the justification of individual labels. If an element is FALSE, the corresponding label is first justified according to hjust (reading direction) and vjust (the perpendicular direction), then rotated. This is the way a textGrob works. If an element is TRUE, the concept is: align the label with the other labels according to rotHjust (reading direction) and rotVjust (the perpendicular direction), then rotate, and finally justify in the width and height directions of the viewport with hjust and vjust, respectively.

rotHjust

a numeric vector or NULL. When the corresponding element of rotJust is TRUE, rotHjust sets the justification of a label in the reading direction. If NULL or an NA element is encountered, an automatic value will be computed based on rotation angle (rot) and justification along the viewport axes (just, hjust and vjust).

rotVjust

a numeric vector or NULL. Set the justification of labels perpendicular to the reading direction when rotJust is TRUE. See rotHjust.

resize

a logical flag. If TRUE (the default), the fontsize of the labels will be adjusted according to the space available. If FALSE, the size will remain constant, even if the graphical object is drawn in a viewport with a different setting for the "cex" graphical parameter.

sizingWidth

If resize is TRUE, a numeric value given here sets the width of the grob used when calculating fontsize at drawing time. If NULL (the default), the size is computed from the actual dimensions of the labels.

sizingHeight

See sizingWidth, only height instead of width.

adjustJust

A logical flag. If TRUE (the default), adjustments are made to the justification of the labels instead of passing the justification settings straight to the underlying textGrob(s). The justification of labels given in expression form will be unified with the justification of character labels, meaning that a setting of vjust = 0 will align the baselines of the labels and vjust = 1 will align the labels at lineheight, or at a multiple of lineheight in case of multiline character labels. The labels will also be shifted so that there is room for descenders.

takeMeasurements

A logical flag. If TRUE, only measurements of labels will be returned instead of a graphical object. An example of where this might be useful is when several labels should have the same fontsize but different graphical parameters such as color, or when the labels should be drawn in different viewports. See the source of sisalTable, particularly makeContent.sisalTable, for an example. If FALSE (the default) a graphical object will be returned.

name

a character string identifier for the graphical object returned by the function. If NULL (the default), a name will be assigned automatically.

gp

graphical parameters. See gpar.

vp

a "viewport" object, the name of a viewport object, a vpPath object pointing to a viewport or NULL (the default). If not NULL, this graphical object will be drawn in the given viewport. The name or the path must point to a descendant of the viewport that is current at drawing time. See current.vpPath, current.vpTree, downViewport and grid.draw.

Details

The number of labels created is the maximum of the lengths of x and y. Variables are recycled to that length if necessary.

All labels of one "dynText" grob have the same fontsize.

Value

If takeMeasurements is FALSE (the default), returns a grob of class "dynText". It can be drawn with grid.draw.

If takeMeasurements is TRUE, returns a list containing measurements of the labels.

Author(s)

Mikko Korpela

See Also

See function textGrob in package grid.

Examples

library(grid)
grid.newpage()
grid.draw(dynTextGrob("Hello", vjust = 0, y = 0))
grid.draw(dynTextGrob(list(expression(y==x^2),
                           "Hello,\ntry resizing me!"),
                      x = rep(1, 2), y = 1, rot = -45,
                      hjust = 1, vjust = 1,
                      rotHjust = c(0, 1), rotVjust = 1))

Create Input Matrix and Output Vector for Time Series Prediction

Description

Given a time series vector, produces the input matrix and output vector for a time series prediction task. The other parameters are the lags to include and the number of steps ahead to predict.

Usage

laggedData(x, lags = 0:9, stepsAhead = 1)

Arguments

x

an atomic vector representing a (uniformly sampled) time series. Any attributes are ignored.

lags

which lags to use for prediction. A vector of non-negative integral values.

stepsAhead

how many steps ahead to predict. A non-negative integral value (integer or numeric).

Details

The default parameters correspond to predicting one step ahead (position t+1) using the ten most recent values (positions t ... t-9).

Value

A list with two components:

X

The (length(x) - max(lags) - stepsAhead) rows by length(lags) columns input matrix with the same type as x.

y

The output vector with length(x) - max(lags) - stepsAhead elements. Same type as x.

Author(s)

Mikko Korpela

Examples

laggedData(1:20)

Plotting Sequential Input Selection Results

Description

A plot method for class "sisal". Supports 3 plot types: error as a function of the number of variables, search graph, and color key of the search graph.

Usage

## S3 method for class 'sisal'
plot(x, which = 1, standardize = "inherit", ...,
     plotArgs = list(list(), list(mai = rep(0.1, 4))),
     xlim = c(x[["d"]], 0), ylim = NULL, ask = TRUE,
     dev.set = !ask, draw.node.labels = TRUE,
     draw.edge.labels = TRUE, draw.selected.labels = TRUE,
     rankdir = c("TB", "LR", "BT", "RL"),
     fillcolor.normal = "deepskyblue",
     fillcolor.pruned = "deeppink",
     fillcolor.selected = "chartreuse",
     fillcolor.levelbest = "gold",
     fillcolor.small = "moccasin", fillcolor.large = "black",
     fillcolor.NA = "white",
     bordercolor.normal = "black",
     bordercolor.special.levelbest = fillcolor.levelbest,
     bordercolor.special.selected = fillcolor.selected,
     color.by.error = FALSE,
     ramp.space = c("Lab", "rgb"), ramp.size = 128,
     error.limits = c(NA_real_, NA_real_),
     category.labels =
         c(normal = gettext("Other", domain="R-sisal"),
           pruned = gettext("Pruned", domain="R-sisal"),
           levelbest = gettext("Best\nin class", domain="R-sisal"),
           selected = gettext("Selected", domain="R-sisal"),
           special.levelbest = gettext("Best\n(no branching)",
                                       domain="R-sisal"),
           special.selected = gettext("Selected\n(no branching)",
                                      domain="R-sisal"),
           shape.normal=gettext("Other", domain="R-sisal"),
           shape.highlighted=gettext("Highlighted", domain="R-sisal")),
     integrate.colorkey = TRUE, colorkey.gap = 0.1,
     colorkey.space = c("right", "bottom", "left", "top"),
     colorkey.title.gp = gpar(fontface = "bold"),
     nodesep = 0.25, ranksep = 0.5,
     graph.attributes = character(0),
     node.attributes = character(0),
     edge.attributes = character(0))

Arguments

x

an object of class "sisal".

which

which plots to draw. A numeric vector containing a subset of the following numbers:

1

error vs. number of inputs.

2

search graph. A directed acyclic graph (DAG).

3

node shape and color keys for the search graph. Requires that plot 2 is drawn, too.

The default is to draw plot number 1. For drawing plot number 2, Bioconductor packages "graph" and "Rgraphviz" must be installed.

Some other arguments of this method only apply to specific plots.

standardize

"inherit" or a logical flag. If TRUE, the error values in plot 1 correspond to standardized data (see standardize in sisal). If FALSE, the original scale of the data is used instead. If "inherit" (the default), the value of this argument is copied from x.

...

arguments passed to plot and matplot. These are used in all plots where plot or matplot do the actual drawing. For more fine-grained control and passing arguments to other graphical functions, use the plotArgs argument.

plotArgs

arguments passed to graphical functions. A list where plotArgs[[k]] (if present) are named lists of arguments passed to plot number k. See ‘Details’.

xlim

the x limits c(x1, x2) of plot 1. A numeric vector. Defaults to showing the whole range, i.e. everything between no input variables at all (except possibly an intercept) and the maximum number of inputs.

ylim

the y limits c(x1, x2) of plot 1. A numeric vector. If NULL (the default), adjusts to the range of y values corresponding to x values delimited by xlim.

ask

a logical flag. If TRUE (the default) and !dev.set, prompts the user before replacing a plot drawn with this function with another one. The user will not be alerted as long as there are free slots in the plot layout (see mfrow and mfcol in par).

dev.set

a logical flag. If TRUE, the function calls dev.set for switching to the next available graphical device when it runs out of free slots in the plot layout. If FALSE, the same graphical device is used for all the plots. The default value is !ask.

draw.node.labels

a logical flag. If TRUE, label the nodes of the search graph plot representing one input variable.

draw.edge.labels

a logical flag. If TRUE, label the edges of the search graph plot with the identity of the removed input variable.

draw.selected.labels

a logical flag. If TRUE, label the nodes of the search graph plot representing the L.v and L.f input variable sets of sisal.

rankdir

the drawing direction of plot number 2 (search graph). A character string, one of "TB" (top to bottom, the default), "LR" (left to right), "BT" (bottom to top), or "RL" (right to left).

fillcolor.normal

fill color for normal nodes in plot number 2.

fillcolor.pruned

fill color for pruned (unevaluated) nodes in plot 2. If color.by.error is TRUE, this color is used as the border color.

fillcolor.selected

fill color for nodes representing the L.v and L.f input variable sets of sisal in plot 2. If color.by.error is TRUE, this color is used as the border color.

fillcolor.levelbest

fill color for nodes with the smallest validation error using a given number of input variables in plot 2. If color.by.error is TRUE, this color is used as the border color.

fillcolor.small

if color.by.error is TRUE, fill color for nodes with small validation error in plot 2.

fillcolor.large

if color.by.error is TRUE, fill color for nodes with large validation error in plot 2.

fillcolor.NA

if color.by.error is TRUE, fill color for pruned (unevaluated) nodes in plot 2.

bordercolor.normal

border color for normal nodes in plot 2.

bordercolor.special.levelbest

border color for special nodes in plot 2. If branching (hbranches > 1) reduces validation error with a given number of input variables, the “no branching” node is marked with this border color. If pruning.keep.best is FALSE, the comparison may not be possible for all sizes of the input variable set.

bordercolor.special.selected

border color for another kind of special nodes in plot 2. The “no branching” L.v or L.f node, if different from the corresponding node in the solution where branching is allowed, is marked with this border color. If pruning.keep.best is FALSE, these alternative L.v and L.f nodes may not be defined, in which case the special color will not be used. If color.by.error is TRUE, this border color is also used to mark nodes that would be marked with fillcolor.selected in the case where color.by.error is FALSE.

color.by.error

a logical flag. If TRUE nodes in plot 2 are colored using a color gradient between fillcolor.small and fillcolor.large according to the validation error in the node. If FALSE, the nodes are colored by category (normal, pruned, selected, levelbest).

ramp.space

color space to be used in plots number 2 and 3 if color.by.error is TRUE. Either "Lab" (the default) or "rgb". See colorRamp.

ramp.size

the number of colors to be used in the color gradient of plot number 3 if color.by.error is TRUE. See colorRampPalette.

error.limits

a numeric vector giving the minimum (first value) and maximum (second value) validation error. These are used as the endpoints of the color gradient used in plots number 2 and 3 if color.by.error is TRUE.

category.labels

text labels to be used in plot number 3 if color.by.error is FALSE. A character vector with elements named "normal", "pruned", "levelbest" and "selected". See the corresponding arguments with the name prefix "fillcolor". The vector must also have elements named "special.levelbest" and "special.selected". See the corresponding arguments with the name prefix "bordercolor". The final required elements are "shape.normal" and "shape.highlighted", which correspond to rectangular and circular nodes, respectively. Circular shape highlights nodes that have the lowest validation error considering the number of inputs used. Also highlighted is each node with the lowest validation error per number of variables but without using branches, if available and different from the unrestricted best node.

integrate.colorkey

a logical flag. If TRUE, plots 2 (graph) and 3 (color and shape key for the graph) will be integrated if possible. This involves a version requirement on the "Rgraphviz" package. If FALSE or the version requirement is not met, the plots will be drawn separately.

colorkey.gap

a numeric value giving the space (in inches) between the graph and the color key when plot 2 and 3 are integrated (integrate.colorkey).

colorkey.space

location of the color and shape key (plot 3) relative to the graph (plot 2). One of "bottom", "right", "top" and "left".

colorkey.title.gp

graphical parameters for the titles in plot 3. See gpar.

nodesep

a Graphviz attribute giving the minimum space in inches between adjacent nodes representing the same number of input variables. This numeric value applies to plot number 2.

ranksep

a Graphviz attribute giving the minimum space in inches between adjacent rows or columns of nodes, where a row or column consists of nodes representing the same number of input variables. This numeric value applies to plot number 2.

graph.attributes

a named character vector of extra Graphviz graph attributes. Applies to plot number 2.

node.attributes

a named character vector of extra Graphviz node attributes. Applies to plot number 2.

edge.attributes

a named character vector of extra Graphviz edge attributes. Applies to plot number 2.

Details

In argument plotArgs, plotArgs[[1]] is passed to matplot, plotArgs[[2]] to the plot method for class "Ragraph", and plotArgs[[3]] to draw.colorkey$key.

For possible color values, see col2rgb.

Value

When 2 %in% which, the function invisibly returns a graph of class "graphNEL" representing the search graph of a run of sisal. Otherwise NULL.

Author(s)

Mikko Korpela

References

For information about graph, node and edge attributes for plot number 2, see the Graphviz web site: https://www.graphviz.org/.

See Also

sisal

Examples

library(graphics)
foo <- testSisal(dataset="toy", Mtimes=10)
## Plotting the search graph requires "Rgraphviz" and "graph"
if (requireNamespace("Rgraphviz", quietly=TRUE) &&
    requireNamespace("graph", quietly=TRUE)) {
    plot(foo, which=2)
}
## Default output is a mean squared error plot
plot(foo)

Plotting Sets of Inputs Produced by Sequential Input Selection

Description

Draws a table depicting the inputs selected by a number of sisal runs, one row for each run.

Usage

## S3 method for class 'sisal'
plotSelected(x, useAllNames = TRUE,
             pickIntPart = FALSE, intTransform = function(x) x,
             formatCArgs = list(), xLabels = 1, yLabels = NULL,
             L.f.color = "black", L.v.color = "grey50",
             other.color = "white", naFill = other.color,
             naStripes = L.v.color, selectedLabels = TRUE,
             otherLabels = FALSE,
             labelPar = gpar(fontface = 1, fontsize = 20, cex = 0.35),
             nestedPar = gpar(fontface = 3),
             ranking = c("pairwise", "nested"), tableArgs = list(),
             ...)

## S3 method for class 'list'
plotSelected(x, ...)

Arguments

x

an object of class "sisal" or a list of such objects giving the results of input selection.

useAllNames

a logical flag. If TRUE, collects the names of input variables from all elements of a list x or from the single "sisal" object. Each unique name is represented by one column in the table. If FALSE, all elements of x are assumed to have the same set of input variables in the same order.

pickIntPart

a logical vector. If pickIntPart[k] is TRUE, the input names collected from x[[k]] (x is a list) or from x (x is a single "sisal" object and k == 1) are filtered so that any name containing an integer part is converted to that integer (the remaining part is dropped). If the length of the vector and the number of rows in the table differ, the values of the vector are recycled.

intTransform

a function that transforms integral valued input names to another integer. Used if and only if the relevant element of pickIntPart is TRUE. The function must accept a numeric vector argument and return a numeric vector. The default value is an identity function.

formatCArgs

a named list of arguments to formatC. If the relevant element of pickIntPart is TRUE, the integral valued column names are formatted with formatC using these arguments. For example, it is possible to add a sign with list(flag = "+").

xLabels

a numeric value, character vector or list affecting the column labels in the table. If useAllNames is TRUE, a named list or character vector can be used to rename inputs. In this case, the names in the vector must contain all the input names gathered from x. The new names (display names) are taken from the values in the vector, indexed with the names from x. If useAllNames is TRUE, a numeric value has no effect. If useAllNames is FALSE, a numeric value is an index to x indicating the object to be used when collecting input names. An unnamed list or character vector of column names can also be used when useAllNames is FALSE.

yLabels

a character vector or list giving the row labels in the table. NULL (the default) means no labels.

L.f.color

fill color for table cells representing an input variable in the L.f set.

L.v.color

fill color for table cells representing an input variable in the L.v set.

other.color

fill color for table cells representing an input variable outside both L.f and L.v.

naFill

background color for table cells representing a missing input variable.

naStripes

stripe color for table cells representing a missing input variable.

selectedLabels

a logical flag. If TRUE (the default), draw labels on table cells representing input variables in the L.f or L.v sets. The label shows the importance rank of the variable. See ‘Details’.

otherLabels

a logical flag. If TRUE, draw labels on table cells representing input variables not included the L.f or L.v sets. The label shows the importance rank of the variable. The default value is FALSE. See ‘Details’.

labelPar

graphical parameters for labels of table cells.

nestedPar

graphical parameters for labels on rows that represent input selection runs where the best nodes of each size are all nested. See ‘Details’. Only used if ranking includes "nested". These take precedence over values set in labelPar.

ranking

which input ranking method(s) to use. A character vector containing one or both of "pairwise" and "nested". Abbreviated versions can be used. See ‘Details’ for a description of the ranking methods. If both rankings are requested by the user and exist, they are both written on the label, but only where the ranks differ. The first element indicates the preferred primary ranking method, and any differing ranks produced by a possible secondary ranking method are presented in parentheses after the rank indicated by the primary method. The default is to use both methods when possible, preferring the always available "pairwise" method.

tableArgs

a named list of arguments passed to sisalTable. This can also be used when arguments of sisalTable and the "sisal" method of plotSelected have the same name.

...

In the "sisal" method, arguments passed to sisalTable. In the "list" method, arguments passed to the next method, determined by the class of the first element in the list.

Details

Currently the "sisal" and "list" methods are the only methods for the generic function plotSelected defined by the sisal package.

Mathematical annotation can be used in text. See plotmath. If the same input is in both the L.f and the L.v sets, L.f.color and L.v.color are mixed in alternating stripes. See col2rgb for a description of possible color values.

The importance rank of input variables is determined using one or both of the following two methods (see ranking):

"nested"

This method requires that all the nodes with the smallest validation error among the nodes with the same number of input variables are nested. Let's imagine a path through the incrementally smaller best nodes (not necessarily a path in the search graph) where the edges are labeled with the ID of the input removed in order to create the smaller model. In this ranking method, the remaining input variable gets rank 1. Traversing the path in the reverse direction and printing the edge labels produces the rest of the input variables from smaller rank to larger. If hbranches = 1 in sisal, the models are always nested and the method agrees with "pairwise".

"pairwise"

This is Copeland's pairwise aggregation method. It can be used in all cases, unlike "nested". The score of an input variable is the number of pairwise victories minus the number of pairwise defeats when compared with other inputs. The inputs are ranked by their score. The method may result in ties. Tied nodes are ranked according to ties.method = "min" in rank.

The pairwise comparisons are performed in the following way: In sisal, at each stage of the search, input variables are ordered and inputs are removed starting from one or more (when hbranches > 1) of the worst ones according to that order. A record, let's say C[A, B], is kept of each pair of inputs (A, B) in order to keep track of how many times A was better than B. Let L be the set of inputs to remove at the current stage of the search in one of the branches and M the set of remaining inputs. Then, C[A, B] is incremented by one for all A in M and B in L, but also for all A in L and B in L such that A is better than B according to the order used for picking the inputs to remove. A gets a pairwise victory over B if C[A, B] > C[B, A].

For information on setting graphical parameters (labelPar, nestedPar), see gpar.

Value

The function is usually called for the side effect (a plot is drawn), but it also returns a grob representation of the plot.

Author(s)

Mikko Korpela

References

Pomerol, J.-C. and Barba-Romero, S. (2000) Multicriterion decision in management: principles and practice. Springer. p. 122. ISBN: 0-7923-7756-7.

See Also

sisal, sisalTable, plotmath, gpar

Examples

library(grDevices)
library(grid)
toy1.2 <- list(testSisal(Mtimes=10, stepsAhead=1, dataset="tsToy"),
               testSisal(Mtimes=10, stepsAhead=2, dataset="tsToy"))
## Resizing enabled:
## - mathematical expressions in titles
## - extracting the integer part of input variable names
grid.newpage()
plotSelected(toy1.2, yLabels = c("+1", "+2"),
             main = "Toy time series",
             xlab = expression(paste("input variables ",
                                     italic(y[t+l]))),
             ylab = expression(paste("output ", italic(y[t+k]))),
             pickIntPart = TRUE, intTransform = function(x) -x)
## Fixed size plot:
## - some graphical parameters adjusted
## - cex in labelPar adjusts the space around the text in table cells
## - new device the same size as the plot
grb <- plotSelected(toy1.2, resizeText = FALSE, resizeTable = FALSE,
                    axesPar = gpar(fontsize = 11, col = "red"),
                    labelPar = gpar(fontsize = 14/0.25, cex = 0.25),
                    fg = "wheat", outerRect = FALSE,
                    linePar = gpar(lty = "dashed"),
                    xAxisRot = 45, just = c("left", "top"),
                    tableArgs = list(x = 0, y = 1), draw = FALSE)
devWidth <- convertWidth(grobWidth(grb), unitTo = "inches",
                         valueOnly = TRUE)
devHeight <- convertHeight(grobHeight(grb), unitTo = "inches",
                           valueOnly = TRUE)
dev.new(width = devWidth, height = devHeight, units = "in", res = 72)
grid.draw(grb)
if (interactive()) {
    dev.set(dev.prev())
} else {
    dev.off()
}

Printing Sequential Input Selection Objects

Description

Prints information contained in a sequential input selection object.

Usage

## S3 method for class 'sisal'
print(x, max.warn = 10, ...)

Arguments

x

an object of class "sisal".

max.warn

a numeric value giving the maximum number of warnings to show. See max.warn in sisal.

...

additional arguments passed to other print methods.

Details

The following information is printed:

  • Parameter values used in the sisal call

  • Data dimensions

  • Names of the input variables, if available

  • Selected inputs, L.v (smallest validation error)

  • Selected inputs, L.f (result within error margin)

  • Whether L.f is a subset of L.v (nested model) or not

  • The removal order and / or rank of the input variables (see plotSelected.sisal)

  • The stages of search (if any) at which branching reduced validation error compared to a hbranches = 1 solution. Not printed if branching was not used or if it is possible that the search did not proceed through every set of variables on the hbranches = 1 path, i.e. if pruning.keep.best was FALSE. One must note that these results, like many others, are subject to randomness. Thus the results may differ between successive runs of sisal.

  • Any warnings produced by the sisal run (see max.warn)

Value

Invisibly returns x.

Author(s)

Mikko Korpela

See Also

More information can be obtained with summary.sisal.

Examples

foo <- testSisal(dataset="toy", nData = 200, Mtimes = 10,
                 noiseSd = 0.5, verbose = 0)
print(foo)

Sequential Input Selection Algorithm (SISAL)

Description

Identifies relevant inputs using a backward selection type algorithm with optional branching. Choices are made by assessing linear models estimated with ordinary least squares or ridge regression in a cross-validation setting.

Usage

sisal(X, y, Mtimes = 100, kfold = 10, hbranches = 1,
      max.width = hbranches^2, q = 0.165, standardize = TRUE,
      pruning.criterion = c("round robin", "random nodes",
                            "random edges", "greedy"),
      pruning.keep.best = TRUE, pruning.reverse = FALSE,
      verbose = 1, use.ridge = FALSE,
      max.warn = getOption("nwarnings"), sp = -1, ...)

Arguments

X

a numeric matrix where each column is a predictor (independent variable) and each row is an observation (data point)

y

a numeric vector containing a sample of the response (dependent) variable, in the same order as the rows of X

Mtimes

the number of times the cross-validation is repeated, i.e. the number of predictions made for each data point. An integral value (numeric or integer).

kfold

the number of approximately equally sized parts used for partitioning the data on each cross-validation round. An integral value (numeric or integer).

hbranches

the number of branches to take when removing a variable from the model. In Tikka and Hollmén (2008), the algorithm always removes the “weakest” variable (hbranches equals 1, also the default here). By using a value larger than 1, the algorithm creates branches in the search graph by removing each of the hbranches “weakest” variables, one at a time. The number of branches created is naturally limited by the number of variables remaining in the model at that point. See also max.width.

max.width

the maximum number of nodes with a given number of variables allowed in the search graph. The same limit is used for all search levels. An integral value (numeric or integer). See pruning.criterion and pruning.keep.best.

q

a numeric value between 0 and 0.5 (endpoints excluded) defining the quantiles 1-q and q. The difference of these sample quantiles is used as the width of the sampling distribution (a measure of uncertainty) of each coefficient in a linear model. The default value 0.165 is the same as used by Tikka and Hollmén (2008). In case of a normally distributed parameter, the width is approximately twice the standard deviation (one standard deviation on both sides of the mean).

standardize

a logical flag. If TRUE, standardizes the data to zero mean and unit variance. If FALSE, uses original data. This affects the scale of the results. If use.ridge is TRUE, this should be set to TRUE or the search graph and the sets of selected variables could be affected.

pruning.criterion

a character string. Options are "round robin", "random nodes", "random edges" and "greedy". Abbreviations are allowed. This affects how the search tree is pruned if the number of nodes to explore is about to exceed max.width. One of the following methods is used to select max.width nodes for the next level of search.

If "round robin", the nodes of the current level (i variables) take turns selecting nodes for the next level (i-1 variables). The turns are taken in order of increasing validation error. Each parent node chooses children according to the order described in ‘Details’. If a duplicate choice would be made, the turn is skipped.

If "random nodes", random nodes are selected with uniform probability.

If "random edges", random nodes are selected, with the probability of a node directly proportional to the number of edges leading to it.

If "greedy", a method similar to "round robin" is used, but with the (virtual) looping order of parents and children swapped. Whereas the outer loop in "round robin" operates over children and the inner loop over parents, the outer loop in "greedy" operates over parents and the inner loop over children. That is, a "greedy" parent node selects all its children before passing on the turn to the next parent.

pruning.keep.best

a logical flag. If TRUE, the nodes that would also be present in the hbranches = 1 case are immune to pruning. If FALSE, the result may underperform the original Tikka and Hollmén (2008) solution in terms of (the lowest) validation error as function of the number of inputs.

pruning.reverse

a logical flag. If TRUE, all the methods described in pruning.criterion except "random nodes" use reverse orders or inverse probabilities. The default is FALSE.

verbose

a numeric or integer verbosity level from 0 (no output) to 5 (all possible diagnostics).

use.ridge

a logical flag. If TRUE, the function uses ridge regression with automatic selection of the regularization (smoothing) parameter.

max.warn

a numeric value giving the maximum number of warnings to store in the returned object. If more warnings are given, their total number is still recorded in the object.

sp

a numeric value passed to magic if use.ridge is TRUE. Initial value of the regularization parameter. If negative (the default), initialization is automatic.

...

additional arguments passed to magic if use.ridge is TRUE. It is an error to supply arguments named "S" or "off".

Details

When choosing which variable to drop from the model, the importance of a variable is measured by looking at two variables derived from the sampling distribution of its coefficient in the linear models of the repeated cross-validation runs:

  1. absolute value of the median and

  2. width of the distribution (see q).

The importance of an input variable is the ratio of the median to the width: hbranches variables with the smallest ratios are dropped, one variable in each branch. See max.width and pruning.criterion.

The main results of the function are described here. More details are available in ‘Value’.

The function returns two sets of inputs variables:

L.v

set corresponding to the smallest validation error.

L.f

smallest set where validation error is close to the smallest error. The margin is the standard deviation of the training error measured in the node of the smallest validation error.

The mean of mean squared errors in the training and validation sets are also returned (E.tr, E.v). For the training set, the standard deviation of MSEs (s.tr) is also returned. The length of these vectors is the number of variables in X. The i:th element in each of the vectors corresponds to the best model with i input variables, where goodness is measured by the mean MSE in the validation set.

Linear models fitted to the whole data set are also returned. Both ordinary least square regression (lm.L.f, lm.L.v, lm.full) and ridge regression models (magic.L.f, magic.L.v, magic.full) are computed, irrespective of the use.ridge setting. Both fitting methods are used for the L.f set of variables, the L.v set and the full set (all variables).

Value

A list with class "sisal". The items are:

L.f

a numeric vector containing indices to columns of X. See ‘Details’.

L.v

a numeric index vector like L.f. See ‘Details’.

E.tr

a numeric vector of length d + 1. See ‘Details’.

s.tr

a numeric vector of length d + 1. See ‘Details’.

E.v

a numeric vector of length d + 1. See ‘Details’.

L.f.nobranch

a numeric vector or NULL. Like L.f but for the “no branching” solution. NULL if branching is not used or if some elements of branching.useful are missing.

L.v.nobranch

like L.f.nobranch but related to L.v.

E.tr.nobranch

a numeric vector or NULL. Like E.tr but for the “no branching” solution. NULL when branching.useful is NULL. An element is missing when the corresponding element of branching.useful is missing.

s.tr.nobranch

like E.tr.nobranch but related to s.tr.

E.v.nobranch

like E.tr.nobranch but related to E.v.

n.evaluated

a numeric vector of length d + 1. The number of nodes evaluated for each model size, indexed by the number of variables used plus one.

edges

a list of directed edges between nodes in the search graph. There is an edge from node A to node B if and only if B was a candidate for a new node to be evaluated, resulting from removing one variable in A. The i:th element of the list contains edges directed away from the node represented by the i:th element of vertices. Each element is a list with one element, "edges", which is a numeric vector of indices to vertices, pointing to the nodes towards which the edges are directed. There are no edges directed away from pruned nodes or nodes representing a single variable.

vertices

a character vector the same size as edges. Contains the names of the nodes in the search graph. Each name contains the indices of the variables included in the set in question, separated by dots.

vertices.logical

a logical matrix containing an alternative representation of vertices. Number of rows is the length of vertices and number of columns is d. The i:th column indicates whether the i:th input variable is present in a given node. The row index and the index to vertices are equivalent.

vertex.data

A data.frame with information about each node in the search graph (missing information means pruned node). The rows correspond to items in vertices. The columns are:

E.tr

mean of MSEs, training.

s.tr

standard deviation (n-1) of MSEs, training.

E.v

mean of MSEs, validation.

E.v.level.rank

rank of the node among all the evaluated (non-pruned) nodes with the same number of variables, in terms of validation error. Smallest error is rank 1.

n.rank.deficient

number of rank deficient linear models. This problem arises when the number of input variables is large compared to the number of observations and use.ridge is FALSE.

n.NA.models

number of models that could not be estimated due to lack of any samples

n.inputs

number of input variables used in the model represented by the node.

min.branches

the smallest branching factor large enough for producing the node. This is a number k between 1 and hbranches. The value for the root node (all input variables) is 1. The value for other nodes is the minimum of the set of values suggested by its parents. The value suggested by an individual parent is the min.branches value of the parent itself or the ranking of the child in terms of increasing importance of the removed variable (see ‘Details’), whichever is larger. For example, when pruning.keep.best is TRUE, the hbranches = 1 search path can be followed by looking for nodes where min.branches is 1.

var.names

names of the variables (column names of X).

n

number of observations in the (X, y) data.

d

number of variables (columns) in X.

n.missing

number of samples where either y or all variables of X are missing.

n.clean

number of complete samples in the data set X, y.

lm.L.f

lm model fitted to L.f variables.

lm.L.v

lm model fitted to L.v variables.

lm.full

lm model fitted to all variables.

magic.L.f

magic model fitted to L.f variables.

magic.L.v

magic model fitted to L.v variables.

magic.full

magic model fitted to all variables.

mean.y

mean of y.

sd.y

standard deviation (denominator n - 1) of y.

zeroRange.y

a logical value indicating whether all non-missing elements of y are equal, with some numeric tolerance.

mean.X

column means of X.

sd.X

standard deviation (denominator n - 1) of each column in X.

zeroRange.X

a logical vector. Like zeroRange.y but for each column of X.

constant.X

a logical vector where the i:th value indicates whether the i:th column of X has a (nearly) constant, non-zero value (NA values allowed).

params

a named list containing the values used for most of the parameter-like formal arguments of the function, and also anything in .... The names are the names of the parameters.

pairwise.points

a numeric square matrix with d rows and columns. The count in row i, column j indicates the number of times that variable i was better than variable j. See ‘Details’ in plotSelected.sisal.

pairwise.wins

a logical square matrix with d rows and columns. A TRUE value in row i, column j indicates that i is more important than variable j. Derived from pairwise.points.

pairwise.preferences

a numeric vector with d elements. Number of wins minus number of losses (when another variable wins) per variable. Derived from pairwise.wins.

pairwise.rank

an integer vector of ranks according to Copeland's pairwise aggregation method. Element number i is the rank of variable (column) number i in X. Derived from pairwise.preferences. See ‘Details’ in plotSelected.sisal.

path.length

a numeric vector of path lengths. Consider a path starting from the full model and continuing through incrementally smaller models, each with the smallest validation error among the nodes with that number of variables. However, the path is broken at each point where the model with one less variable cannot be constructed by removing one variable from the bigger model (is not nested). The vector contains the lengths of the pieces. Its length is the number of breaks plus one.

nested.path

a numeric vector containing the indices (column numbers) of the input variables in their removal order on the “nested path”. The first element is the index of the variable that was removed first. The remaining variable is the last element. If the path does not exist, this is NULL. See ‘Details’ in plotSelected.sisal.

nested.rank

an integer vector of ranks determined by nested.path. Element number i is the rank of variable (column) number i in X. NULL if nested.path is NULL. See ‘Details’ in plotSelected.sisal.

branching.useful

If branching is enabled (hbranches > 1), this is a logical vector of length d. If the i:th element is TRUE, branching improved the best model with i variables in terms of validation error. The result is NA if a comparison is not possible (may happen if pruning.keep.best is FALSE). If branching is not used, this is NULL.

warnings

warnings stored. A list of objects that evaluate to a character string.

n.warn

number of warnings produced. May be higher than number of warnings stored.

Author(s)

Mikko Korpela

References

Tikka, J. and Hollmén, J. (2008) Sequential input selection algorithm for long-term prediction of time series. Neurocomputing, 71(13–15):2604–2615.

See Also

See magic for information about the algorithm used for estimating the regularization parameter and the corresponding linear model when use.magic is TRUE.

See summary.sisal for how to extract information from the returned object.

Examples

library(stats)
set.seed(123)
X <- cbind(sine=sin((1:100)/5),
           linear=seq(from=-1, to=1, length.out=100),
           matrix(rnorm(800), 100, 8,
                  dimnames=list(NULL, paste("random", 1:8, sep="."))))
y <- drop(X %*% c(3, 10, 1, rep(0, 7)) + rnorm(100))
foo <- sisal(X, y, Mtimes=10, kfold=5)
print(foo)           # selected inputs "L.v" are same as
summary(foo$lm.full) # significant coefficients of full model

Download External Datasets for SISAL

Description

Loads external datasets for testing with SISAL. Choices are laser generated data and Poland electricity load data.

Usage

sisalData(dataset = c("poland", "laser", "laser.cont"), verify = TRUE)

Arguments

dataset

A character string: "poland" (default), "laser" or "laser.cont" (see ‘Note’).

verify

A logical flag. If TRUE, verifies the integrity of the downloaded data by computing a checksum and comparing it to a pre-computed value.

Details

The laser generated data come in two parts, "laser" and "laser.cont". The Poland electricity load data is also divided in two parts, but they are both returned with dataset="poland".

This function requires an Internet connection. The download may fail due to a problem such as the remote server being unavailable.

Value

With option dataset="laser", returns an integer vector of length 1000.

With option dataset="laser.cont", returns an integer vector of length 9093.

With option dataset="poland", returns a list with two numeric vectors:

learn

1400 values

test

201 values

Note

Checked on 2020-02-14, the Santa Fe datasets are no longer available at their previous location. Attempting to download them with this function will result in an error.

Author(s)

Mikko Korpela

References

The Santa Fe Time Series Competition Data / Data Set A: Laser generated data. Availability unknown (2020-02-14).

Environmental and Industrial Machine Learning Group / Datasets / Poland Electricity Load. https://research.cs.aalto.fi/aml/datasets.shtml. URL accessed on 2024-10-25.

See Also

testSisal

Examples

## Not run: 
foo <- sisalData("poland")
length(foo$learn) # 1400
length(foo$test)  # 201
## End(Not run)

Draw Table with Equally Sized Cells

Description

Draws a resizable or fixed-size table with equally sized cells. Main title, axis (tick) labels and axis titles (left, bottom) are optional. Cells can have individual background and text colors and stripes.

Usage

sisalTable(labels = matrix(seq_len(12), 3, 4),
           nRows = NROW(labels), nCols = NCOL(labels),
           bg = sample(colors(), nRows * nCols, replace = TRUE),
           stripeCol = NULL, fg = NULL, naFill = "white",
           naStripes = "grey50", main = NULL, xlab = NULL,
           ylab = NULL, xAxisLabels = NULL, yAxisLabels = NULL,
           draw = TRUE, outerRect = TRUE, innerLines = TRUE,
           nStripes = 7, stripeRot = 45, stripeWidth = 0.2,
           stripeScale = 0.95, resizeText = TRUE,
           resizeTable = TRUE, resizeMain = resizeText,
           resizeLab = resizeText, resizeAxes = resizeText,
           resizeLabels = resizeTable && resizeText,
           x = unit(0.5, "npc"), y = unit(0.5, "npc"),
           width = unit(0.97, "npc"), height = unit(0.97, "npc"),
           default.units = "npc", just = "center",
           clip = "inherit", xAxisRot = 0, yAxisRot = 0,
           xAxisJust = c(0.5, 1), xAxisX = 0.5, xAxisY = 1,
           yAxisJust = c(1, 0.5), yAxisX = 1, yAxisY = 0.5,
           mainMargin = if (resizeMain) 0.15 else unit(8, "points"),
           xlabMargin = if (resizeLab) 0.1 else unit(5, "points"),
           ylabMargin = if (resizeLab) 0.1 else unit(5, "points"),
           axesMargin = if (resizeAxes) 0.1 else unit(5, "points"),
           axesSize = 0.8, forceAxesSize = FALSE,
           mainSize = 1, xlabSize = 1, ylabSize = 1,
           mainPar = gpar(fontface = "bold", fontsize = 14),
           labPar = gpar(fontface = "plain", fontsize = 14),
           labelPars = gpar(fontsize = 20, cex = 0.6),
           axesPar = gpar(fontsize = 10),
           rectPar = gpar(), linePar = gpar(),
           name = NULL, gp = NULL, vp = NULL)

Arguments

labels

the labels to use in the table cells. A list or an atomic vector containing something that can be displayed as text, e.g. character values. One element is used for each cell. If the object has a "dim" attribute (matrix, array), it is used for determining the number of rows and columns in the table.

NA means no text.

nRows

the number of rows in the table. A positive integral number.

nCols

the number of columns in the table. A positive integral number.

bg

the background colors of the table cells. One element is used for each cell.

stripeCol

an optional vector of colors. If used, indicates the color of stripes to be painted on top of the background color in each table cell. One element is used for each table cell. NA means no stripes.

fg

the text colors of the table cells. One element is used for each cell. If NULL (the default), black or white text is used so that the contrast between foreground and background is maximized.

naFill

background color to use when the label of a table cell is NA. This is a single color value.

naStripes

table cells with an NA label are indicated with stripes. This is the color of the stripes, a single color value. The stripes can be hidden by using a value identical with that of naFill.

main

the main title of the plot.

xlab

a title for the x axis.

ylab

a title for the y axis.

xAxisLabels

a label for each column of the table.

yAxisLabels

a label for each row of the table.

draw

a logical flag indicating whether to draw the table. If FALSE, no drawing is done.

outerRect

a logical flag indicating whether a rectangle will be drawn around the table.

innerLines

a logical flag indicating whether line segments will be drawn between the table cells.

nStripes

a positive integral number giving the number of stripes to be drawn in table cells. Only applies to those cells where stripes are used, i.e. when the relevant element of label is NA or stripeCol is not NA. The stripes are spaced evenly. Defaults to 7.

stripeRot

an integral number giving the rotation angle (degrees, counterclockwise) of the stripes used in table cells. Defaults to 45 which means diagonal stripes parallel to a line segment between the lower left corner and the upper right corner of the cell. Value 0 means horizontal and 90 vertical stripes.

stripeWidth

a numerical value giving the width of the stripes used in cells as a proportion of the available width. Values between 0 and 1 are allowed, excluding the endpoints. Defaults to 0.2.

stripeScale

a numerical value indicating the proportion of the area of a table cell to be used for the stripe pattern. The pattern is always centered, and the possible empty space is left on the borders of the cell. Values between 0 and 1 are allowed, including the endpoints. Defaults to 0.95.

resizeText

a logical flag indicating whether to use dynamic text size. This is only used as the default value of resizeMain, resizeLab, resizeLabels and resizeAxes. Defaults to TRUE.

resizeTable

a logical flag indicating whether the size of the table will depend on the size of the main viewport, which itself may be static or depend on the size of the graphical device. Defaults to TRUE. See ‘Details’.

resizeMain

a logical flag indicating whether the main title will be resizable.

resizeLab

a logical flag indicating whether the the x axis and y axis titles will be resizable.

resizeLabels

a logical flag indicating whether the labels used in the table cells will be resizable.

resizeAxes

a logical flag indicating whether the row and column labels will be resizable.

x

a numeric vector or unit object of length one specifying the x location of the graphical object.

y

a numeric vector or unit object of length one specifying the y location of the graphical object.

width

a numeric vector or unit object of length one specifying the width of the graphical object. See ‘Details’.

height

a numeric vector or unit object of length one specifying the height of the graphical object. See ‘Details’.

default.units

a character string indicating the unit to use for numeric values of x, y, width and height.

just

a character or numeric vector of one or two values specifying the justification of the graphical object relative to its (x, y) location. See viewport.

clip

a character string specifying what to do if the graphical object overflows the viewport reserved for it. See ‘Details’.

xAxisRot

a numeric value giving the rotation angle of the column labels in degrees.

yAxisRot

a numeric value giving the rotation angle of the row labels in degrees.

xAxisJust

justification setting for column labels. A numeric or character vector. Rotation (if any) will be done before justification. See just in textGrob for possible values.

xAxisX

x location of column labels relative to the space allocated for them. A numeric value where 0 means left and 1 right.

xAxisY

y location of column labels relative to the space allocated for them. A numeric value where 0 means bottom and 1 top.

yAxisJust

justification setting for row labels. A numeric or character vector. See xAxisJust.

yAxisX

x location of row labels relative to the space allocated for them. A numeric value where 0 means left and 1 right.

yAxisY

y location of row labels relative to the space allocated for them. A numeric value where 0 means bottom and 1 top.

mainMargin

size of the margin between the main title and the table.

xlabMargin

size of the margin between the x axis title and the next graphical object towards the table.

ylabMargin

size of the margin between the y axis title and the next graphical object towards the table.

axesMargin

size of the margin between the row or column labels and the table.

axesSize

a positive numeric value specifying the desired ratio of fontsize in row and column labels to fontsize in table cells.

forceAxesSize

a logical flag. If TRUE, the function will reduce the size of text in table cells if it is necessary in order to achieve the desired axesSize.

mainSize

scale factor for fontsize of main title. A positive numeric value. Only effective when resizeMain is TRUE.

xlabSize

scale factor for fontsize of x axis title. A positive numeric value. Only effective when resizeLab is TRUE.

ylabSize

scale factor for fontsize of y axis title. A positive numeric value. Only effective when resizeLab is TRUE.

mainPar

graphical parameters for the main title.

labPar

graphical parameters for x and y axis titles.

labelPars

graphical parameters for labels used in table cells. Can also be a list, one element for each table cell, recycled if necessary.

axesPar

graphical parameters for row and column labels.

rectPar

graphical parameters for the rectangle around the table.

linePar

graphical parameters for the line segments between table cells.

name

a character string identifier for the graphical object returned by the function. If NULL (the default), a name will be assigned automatically.

gp

graphical parameters for the whole object.

vp

a "viewport" object, the name of a viewport object, a vpPath object pointing to a viewport or NULL (the default). If not NULL, this graphical object will be drawn in the given viewport. The name or the path must point to a descendant of the current viewport. See current.vpPath, current.vpTree, downViewport and grid.draw.

Details

This function was written to be used with plotSelected but it should be generic enough to be useful for other purposes, too.

The color and text vectors (including matrices and arrays) pointing to table cells (labels, bg, stripeCol, fg) are interpreted in column-major order, like linear indexing of a matrix. Each data.frame argument is collapsed to a list by combining its columns. Finally, values are recycled if needed, also in xAxisLabels and yAxisLabels.

For possible color values, see col2rgb.

In the various text objects, mathematical annotation (see plotmath) is supported in addition to character values.

For information on setting graphical parameters (gp, mainPar, labPar, ...), see gpar.

The graphical object returned is a gTree which contains a gList of graphical objects and a vpTree of viewports. The child viewports are placed inside the parent using a grid.layout. The size of the whole object is the size of the parent viewport. It will be fixed or depend on the space available to it:

  • If all graphical elements are non-resizable (but resizeLabels can be TRUE), a suitable fixed size will be computed.

  • Otherwise, the size is determined by width and height. However, if there are non-resizable elements, the graphical object may be larger than that.

The graphical object will not use any excess space. In other words, the width and height reported by grobWidth and grobHeight are tight. It is possible that some parts of the plot may overflow their assigned space and the bounds computed for the whole graphical object. Examples include using large fixed-size text elements or large values of the gpar graphical parameter "cex". Clipping can be adjusted through clip.

If resizeAxes is TRUE, axesMargin must be a non-negative numeric value giving the size of the margin as a proportion of the side length of a table cell. If resizeAxes is FALSE, axesMargin can also be a unit object. The arguments mainMargin and labMargin are analogous to axesMargin.

Value

The function is usually called for the side effect (a plot is drawn), but it also returns a grob representation of the plot. The returned object is a custom gTree of class "sisalTable".

Author(s)

Mikko Korpela

Examples

library(grDevices)
library(grid)
## Default: 3 by 4 table with labels 1:12 and random background colors
grid.newpage()
sisalTable()

## Four examples in a grid layout
rowCol <- c(1, 18, 2, 18, 1)
lo <- grid.layout(nrow = 5, ncol = 5,
                  widths = rowCol, heights = rowCol)
grid.newpage()
pushViewport(viewport(layout = lo, name = "bgLayout"))
grid.rect(gp=gpar(fill="grey75", col="grey75"))

rNames <- c("topmargin", "top", "hspace", "bottom", "bottommargin")
cNames <- c("leftmargin", "left", "vspace", "right", "rightmargin")
for (Row in c(2, 4)) {
    for (Col in c(2, 4)) {
        pushViewport(viewport(layout.pos.row = Row,
                              layout.pos.col = Col,
                              name = paste(rNames[Row],
                                           cNames[Col], sep="")))
        grid.rect(gp=gpar(fill="cadetblue"))
        upViewport(1)
    }
}

colors1Vec <- terrain.colors(12)
colors1Mat <- matrix(colors1Vec, 3, 4)
labels1Vec <- sample(c(letters, LETTERS), 12)
labels1Mat <- matrix(labels1Vec, 3, 4)

## Column vector, aligned with the right side of the viewport
longText <- rep("", 12)
longText[3] <- "a longish piece of text"
longText[9] <- "and some more"
sisalTable(labels1Vec, bg = colors1Vec, vp = "topleft",
           x = 1, just = "right",
           yAxisLabels = longText, xAxisLabels = "Boo")

## Matrix, zero margin
downViewport("topright")
sisalTable(labels1Mat, bg = colors1Mat,
           width = 1, height = 1, name = "trPlot",
           xAxisLabels = 1:4, yAxisLabels = LETTERS[1:3])
grid.rect(width = grobWidth("trPlot"), height = grobHeight("trPlot"),
          gp = gpar(lty="dashed", col = "white", lwd = 2))
upViewport(1)

## Transpose of matrix, width and height 0.75 "npc" units
downViewport("bottomleft")
sisalTable(t(labels1Mat), bg = t(colors1Mat),
           width = 0.75, height = 0.75, name = "blPlot",
           yAxisLabels = 1:4, xAxisLabels = LETTERS[1:3])
grid.rect(width = grobWidth("blPlot"), height = grobHeight("blPlot"),
          gp = gpar(lty="dashed", col = "white", lwd = 2))
upViewport(1)

## ?plotmath, some cells with no background color
labels2 <- expression(x^{y+x}, sqrt(x), bolditalic(x), NA)
bgCol <- c(rep("white", 3), NA)
sisalTable(labels2, nRows=3, nCols=5, bg = bgCol, naFill = NA,
           naStripes = "darkmagenta", vp="bottomright",
           main = "plotmath text")

Summarizing Sequential Input Selection Results

Description

summary method for class "sisal"

Usage

## S3 method for class 'sisal'
summary(object, ...)
## S3 method for class 'summary.sisal'
print(x, ...)

Arguments

object

an object of class "sisal".

x

an object of class "summary.sisal".

...

arguments passed to/from other methods.

Details

The functions compute and print summaries (summary.lm) of the ordinary least squares regression models stored in the object and some additional information.

Value

The function summary.sisal returns a list with class "summary.sisal", currently containing:

summ.full

summary of the full model. An object of class "summary.lm".

summ.L.v

summary of the L.v model. An object of class "summary.lm".

summ.L.f

summary of the L.f model. An object of class "summary.lm".

error.df

a data.frame containing information on the best variable sets with a given number of variables, with the following columns (copied from object):

n.inputs

number of inputs (row label).

E.tr

mean training MSE.

s.tr

standard deviation of training MSE.

E.v

mean validation MSE.

L.f.flag

logical vector where the location of TRUE points the smallest variable set with thr.flag TRUE.

L.v.flag

logical vector where the location of TRUE points the variable set with the smallest validation error.

thr.flag

logical vector where TRUE means that error is at most E.v[L.v.flag] + s.tr[L.v.flag].

The function print.summary.sisal invisibly returns x.

Author(s)

Mikko Korpela

See Also

sisal, print.sisal

Examples

foo <- testSisal(dataset="toy", Mtimes=10, hbranches=2)
summary(foo)

Testing the Sequential Input Selection Algorithm

Description

Tests sisal with example datasets or time series data. The function uses the training part of an example dataset or user-supplied numeric data interpreted as a time series.

Usage

testSisal(dataset = c("tsToy", "laser", "poland", "toy"), nData = Inf,
          FUN = "sisal", lags = NULL, stepsAhead = 1,
          noiseSd = 0.2, verbose = 1, ...)

Arguments

dataset

the dataset to use. A numeric vector containing time series data or one of "tsToy" (the default), "laser", "poland" and "toy".

nData

a numeric value containing the number of observations to use. If larger than the number of observations in the dataset, all of the data will be used (the default).

FUN

which function to call. By default, acts as a front end to sisal. This can be any function that accepts arguments named "X", "y" and "verbose". See match.fun for legal values.

lags

a numeric or integer vector. When using time series data (dataset is numeric, "laser", "poland" or "tsToy"), the function creates lagged versions of the time series to be used as input variables in sisal. The lags are specified here. These are non-negative integral values where 0 means the latest observation, 1 is the previous observation etc. The default values for "laser", "poland" and "tsToy" are 0:19, 0:14 and 0:9, respectively.

stepsAhead

an integral value specifying how many steps ahead to predict in a time series setting. The default is 1.

noiseSd

standard deviation of noise to be used with the "toy" dataset. The base noise is always the same (stored with the dataset) and only scaled to match this setting.

verbose

a numeric or integer verbosity level. This function only has two verbosity levels (0 and larger than 0), but the value is also propagated to FUN.

...

arguments passed to FUN.

Details

The function recognizes if a numeric dataset is the "laser" or "poland" dataset. In case repeated experiments will be performed on those datasets, it is best to explicitly fetch them with sisalData before using this function. Doing so reduces the amount of network traffic and makes offline work possible.

Value

The value returned by function FUN, when called with the given dataset (processed by this function) and parameters. See the help page of the relevant function, e.g. sisal.

Author(s)

Mikko Korpela

See Also

See sisalData, toy.learn and tsToy.learn for documentation on the datasets.

The performance of the models returned by this functions can be evaluated using bootMSE, which uses a separate test part of the dataset.

Examples

foo <- testSisal(dataset="toy", hbranches=2, max.width=2, Mtimes=5,
                 use.ridge=TRUE)
print(foo)
names(foo)

Toy Data for SISAL (Learning Set)

Description

Numeric matrix with independent and dependent variables and noise

Usage

toy.learn

Format

The format is:

 num [1:1000, 1:12] -0.62067 1.36985 0.00122 0.75527 -1.82271 ...
 - attr(*, "dimnames")=List of 2
  ..$ : NULL
  ..$ : chr [1:12] "y" "noise" "X1" "X2" ...

Details

This is the learning set of the toy data, i.e. 1000 rows of the whole 1500 row dataset.

Columns "X1", "X2", ..., "X10" were generated with rnorm to follow a standard normal distribution.

Column "y" is a linear combination of "X1", "X2", "X3", coefficients (1:3)/sqrt(sum((1:3)^2)), yielding a theoretical standard normal distribution.

Column "noise" was also generated from the standard normal distribution.

Use file.show(system.file("toyDataSrc", "sisalToy.R", package="sisal")) to view the script that generated the data.

See Also

toy.test, testSisal

Examples

library(graphics)
plot(as.data.frame(toy.learn))

Toy Data for SISAL (Test Set)

Description

Numeric matrix with independent and dependent variables and noise

Usage

toy.test

Format

The format is:

 num [1:500, 1:12] -0.543 -0.881 0.115 0.461 -0.173 ...
 - attr(*, "dimnames")=List of 2
  ..$ : NULL
  ..$ : chr [1:12] "y" "noise" "X1" "X2" ...

Details

This is the test set of the toy data, i.e. 500 rows of the whole 1500 row dataset.

For other details, see toy.learn.

See Also

toy.learn, bootMSE

Examples

library(graphics)
plot(as.data.frame(toy.test))

Toy Time Series Data for SISAL (Learning Set)

Description

Numeric vector with autoregressive (AR) time series data

Usage

tsToy.learn

Format

The format is:

 num [1:1000] 0.7529 -0.2576 0.441 0.8473 0.0164 ...

Details

This is the learning set of the toy time series data, i.e. the first 1000 of the total 3000 observations.

The data follow a second order AR model. The first order coefficient is -0.5 and the second order coefficient 0.3. The autocovariances for lags 0 to 4 are c(1.0, -0.71, 0.66, -0.54, 0.47) (theoretical values, two significant digits).

Use file.show(system.file("toyDataSrc", "sisalToyTs.R", package="sisal")) to view the script that generated the data.

See Also

tsToy.test, testSisal

Examples

library(graphics)
library(stats)
plot(tsToy.learn)
acf(tsToy.learn)

Toy Time Series Data for SISAL (Test Set)

Description

Numeric vector with autoregressive (AR) time series data

Usage

tsToy.test

Format

The format is:

 num [1:2000] 0.583 -0.71 -1.172 1.067 -0.719 ...

Details

This is the test set of the toy time series data, i.e. the last 2000 of the total 3000 observations.

The data follow a second order AR model. The first order coefficient is -0.5 and the second order coefficient 0.3.

Use file.show(system.file("toyDataSrc", "sisalToyTs.R", package="sisal")) to view the script that generated the data.

See Also

tsToy.learn, bootMSE

Examples

library(graphics)
library(stats)
plot(tsToy.test)
acf(tsToy.test, type="partial")