selectPortfolio: selectPortfolio

In danielhorn/multicrit_result_test: Multi-objective selection of algorithm portfolios

Description

Method used to start the multi-objective algorithm portfolio selection

Usage

selectPortfolio(data, var.cols, algo.col, repl.col, indicator = "hv",
  ref.point = c(1.1, 1.1), lambda = 100, eta = 0.5, w = c(0.05, 0.95),
  cp = 0.1, normalize = TRUE)

Arguments

data	[data.frame] Data to be analysed. Must contain all var.cols, algo.col and repl.col. var.cols should form a Pareto front for every combination of repl.col and algo.col.
var.cols	[character(2)] Names of columns with containing the values of the Pareto fronts.
algo.col	[character(1)] Name of column containing the particular algorithm.
repl.col	[character(1)] Name of column containing the particular algorithm.
indicator	[character(1)] Which multi-objective indicator should be used? Possible values are: “hv”: Dominated Hypervolume (S-Metric), the default “epsilon”: epsilon-indicator “r2”: R2-indicator
ref.point	[numeric(2)] Reference point for the hypervolume indicator. Default is c(1.1, 1.1)
lambda	[numeric(1)] Number of vector used in the calculation of the R2-indicator. Default is 100.
eta	[numeric(1)] In how many replications each algorithm must have non-dominated points to be selected in the first selection step? Must be in [0, 1] - 0.5 requres non-dominated points more than in half of all replications. Default is 0.5
w	[numeric(2)] Weight vector for the augmented Tschebbyscheff Norm in the second selection stept. The first elements corresponds to the number of algorithms, the second to the optimality gap. If you want to have a low optimality gap, the second element of w should be set to an high value, if a low number of algorithms is more important the first element must be high. All positive real weights are allowed and will be normalized to sum(w) = 1. Default is c(0.05, 0.95).
cp	[numeric(1)] Complexity parameter of the decision tree used in step 3. Same parameter as in rpart. Default is 0.1.
normalize	[logical(1)] Should the data be normalized to [0, 1]? Default is TRUE.

Value

Object of class frontTestResult. Named list with the elements:

• non.dominated.algos [logical]Named vector, each element corresponds to one algorithm. TRUE if algorithm is selected in step 1.

• algos.domination.count [numeric]Named vector, each element corresponds to one algorithm. Number of replications with non-dominated points for each algorithm.

• relevant.algos [logical]Named vector, each element corresponds to one algorithm selected in step 1. TRUE if the algorithm is selected in step 2.

• algos.selection.vals [data.frame]Data.frame with 4 cols and 2^(number of algorithms after in step 2) - 1 rows. First cols active in this subset. Last 2 rows are numeric. First is the number of algos in this subset, second one the optimality gap..

• best.algo.order [Factor]Vector gives the order of the algorithms on the common Pareto front, algorithms low values of the first var.col first.

• split.vals [numeric]Split values between the algorithms given in best.algo.order.

• args [list] List containing all input arguments

Examples

## Not run: 
# Load data - for the data with subsampling enabled use apprSubsampleSVMParetoFronts
data(apprSVMParetoFronts)

# Avaible datasets: codrna, mnist, protein, vehicle
data = subset(apprSVMParetoFronts, apprSVMParetoFronts$dataset == "mnist")

# Start the front analysis with the main procedure
res = selectPortfolio(
  data = data,
  var.cols = c("error", "execTime"),
  algo.col = "solver",
  repl.col = "repl",
  indicator = "hv",
  ref.point = c(1.1, 1.1),
  eta = 0.5,
  w = c(0.05, 0.95),
  cp = 0.01,
  normalize = TRUE
) 
print(res)
plot(res, colors = c("turquoise", "green", "violet", "red", "black", "blue"))
## End(Not run)

danielhorn/multicrit_result_test documentation built on May 14, 2019, 4:05 p.m.