Implementierung von Tabu Suche in R

Question

Ich versuche Tabu Search auf einem Klassifikationsdatensatz zu implementieren, nämlich Indische Patienten Lebererkrankung im UCI-Repository unter https://archive.ics.uci.edu/ml/datasets/ILPD+(Indian+Liver+Patient+Dataset), aber mit Problemen. Es folgt der Code, den ich verwendet habe

NF <- 10 
NTR <- 193 
NTE <- 193 
library(class) 
library(e1071) 
library(caret) 
library(party) 
library(nnet) 
ILPD <- read.csv("C:/Users/Dell/Desktop/Codes and Datasets/ILPD.csv") 
nrow(ILPD) 
set.seed(9850) 
gp<-runif(nrow(ILPD)) 
ILPD<-ILPD[order(gp),] 
idx <- createDataPartition(y = ILPD$Class, p = 0.7, list = FALSE) 
train<-ILPD[idx,] 
test<-ILPD[-idx,] 
ver<-test[,11] 
evaluate <- function(th){ 
if (sum(th) == 0)return(0)    
model <- svm(train[ ,th==1], train[,11] , gamma = 0.1, kernel ="sigmoid", na.action = na.omit) 
pred <- predict(model, test[ ,th==1]) 
csRate <- sum(pred == ver)/NTE 
penalty <- (NF - sum(th))/NF 
return(csRate + penalty) 
} 
library(tabuSearch) 
res <- tabuSearch(size = NF, iters = 2, objFunc = evaluate, config =  matrix(1,1,NF), listSize = 5, nRestarts = 4) 
plot(res) 
plot(res, "tracePlot") 
summary(res, verbose = TRUE) 

Fehler:

Error in if (any(co)) { : missing value where TRUE/FALSE needed 
In addition: Warning message: 
In FUN(newX[, i], ...) : NAs introduced by coercion 
Called from: svm.default(train[, th == 1], train[, 11], gamma = 0.1, kernel = "sigmoid", na.action = na.omit) 

Ein Teil der Daten

structure(list(age = c(55L, 48L, 14L, 17L, 40L, 37L), gender = c(0L, 
0L, 0L, 0L, 1L, 0L), TB = c(0.9, 2.4, 0.9, 0.9, 0.9, 0.7), DB = c(0.2, 
1.1, 0.3, 0.2, 0.3, 0.2), Alkphos = c(116L, 554L, 310L, 224L, 
293L, 235L), SGPT = c(36L, 141L, 21L, 36L, 232L, 96L), sgot = c(16L, 
73L, 16L, 45L, 245L, 54L), TP = c(6.2, 7.5, 8.1, 6.9, 6.8, 9.5 
), ALB = c(3.2, 3.6, 4.2, 4.2, 3.1, 4.9), AG = c(1, 0.9, 1, 1.55, 
0.8, 1), Class = structure(c(2L, 1L, 2L, 1L, 1L, 1L), .Label = c("One", 
"Two"), class = "factor")), .Names = c("age", "gender", "TB", 
"DB", "Alkphos", "SGPT", "sgot", "TP", "ALB", "AG", "Class"), row.names = c(216L, 
405L, 316L, 103L, 20L, 268L), class = "data.frame") 

Wenn jemand könnte mir helfen, damit

Answer 1

Ich wollte um zu sehen, wie Tabu überhaupt funktionierte, schien es ein guter Anfang zu sein.

Grundsätzlich müssen Sie Ihren Code besser testen, evaluate hat einfach nicht funktioniert. Es ist einfach, von Hand zu testen, indem Sie Werte von th erstellen und dann evaluate auf ihnen aufrufen.

Verwenden Sie auch Kommentare auf hoher Ebene, um Ihren Code zu organisieren und zu verfolgen, was Sie gerade tun, vor allem, wenn Sie in SO schreiben, um uns Zeit zu sparen, herauszufinden, was Sie vorhaben.

Nicht sicher, ob diese Ergebnisse gut sind, die Datenmenge ist so minimal, dass es schwer zu sagen ist.

Auf jeden Fall ist hier der geänderte Code:

NF <- 10 
NTR <- 193 
NTE <- 193 
library(class) 
library(e1071) 
library(caret) 
library(party) 
library(nnet) 

ILPD1 <- structure(
list(
    age = c(55L,48L,14L,17L,40L,37L), 
    gender = c(0L,0L,0L,0L,1L,0L), 
    TB = c(0.9,2.4,0.9,0.9,0.9,0.7), 
    DB = c(0.2,1.1,0.3,0.2,0.3,0.2), 
    Alkphos = c(116L,554L,310L,224L,293L,235L), 
    SGPT = c(36L,141L,21L,36L,232L,96L), 
    sgot = c(16L,73L,16L,45L,245L,54L), 
    TP = c(6.2,7.5,8.1,6.9,6.8,9.5), 
    ALB = c(3.2,3.6,4.2,4.2,3.1,4.9), 
    AG = c(1,0.9,1,1.55,0.8,1), 
    Class = structure(c(2L,1L,2L,1L,1L,1L), 
    .Label = c("One","Two"), 
    class = "factor") 
    ), 
    .Names = c("age","gender","TB","DB","Alkphos", 
       "SGPT","sgot","TP","ALB","AG","Class"), 
    row.names = c(216L,405L,316L,103L,20L,268L), 
    class = "data.frame" 
) 

ILPD <- ILPD1 
#ILPD <- read.csv("ILPD.csv") 
nrow(ILPD) 

set.seed(9850) 

# setup test and training data 
gp <- runif(nrow(ILPD)) 
ILPD <- ILPD[order(gp),] 
idx <- createDataPartition(y = ILPD$Class,p = 0.7,list = FALSE) 
train <- ILPD[idx,] 
test <- ILPD[ - idx,] 
ver <- test[,11] 

evaluate <- function(th) { 
    # evaluate the tabu for a value of th 
    # tabuSearch will use this function to evaluate points in its search space 
    # 

    # if everything is turned off just return zero as we are not interested 
    if(sum(th) == 0) return(0) 

    # we just want to train our svm on the columns for which th==1 
    svmtrn <- train[,th==1] 

    # but we need to have the Class varible as our label 
    if (is.null(trn$Class)) return(0) 

    # Train up an svm now 
    # Note that the first argument is the forumula we are training 
    model <- svm(Class~.,svmtrn,gamma = 0.1,kernel = "sigmoid",na.action = na.omit) 

    pred <- predict(model,test) 

    # now evaluate how well our prediction worked 
    csRate <- sum(pred == ver)/NTE 
    penalty <- (NF - sum(th))/NF 
    return(csRate + penalty) 
} 

library(tabuSearch) 
evaluate(matrix(1,1,NF)) 
res <- tabuSearch(size = NF,iters = 2,objFunc = evaluate, 
        config = matrix(1,1,NF),listSize = 5,nRestarts = 4) 
plot(res) 
plot(res,"tracePlot") 
summary(res,verbose = TRUE) 

Hier sind die Ausgangsergebnisse:

[1] 6 
[1] 0.005181347 
Tabu Settings 
    Type          = binary configuration 
    No of algorithm repeats     = 1 
    No of iterations at each prelim search  = 2 
    Total no of iterations      = 12 
    No of unique best configurations   = 8 
    Tabu list size        = 5 
    Configuration length      = 10 
    No of neighbours visited at each iteration = 10 
Results: 
    Highest value of objective fn = 0.70518 
    Occurs # of times    = 1 
    Optimum number of variables  = 3 
Optimum configuration: 
[1] 1 0 0 0 0 1 0 0 0 1 

Und hier ist Ihr Grundstück: