frbs.learn: The frbs model building function

Description

This is one of the central functions of the package. This function is used to generate/learn the model from numerical data using fuzzy rule-based systems.

Usage

frbs.learn(data.train, range.data = NULL, method.type = c("WM"),
  control = list())

Arguments

data.train

a data frame or matrix (\(m \times n\)) of data for the training process, where \(m\) is the number of instances and \(n\) is the number of variables; the last column is the output variable. It should be noted that the training data must be expressed in numbers (numerical data). And, especially for classification tasks, the last column representing class names/symbols isn't allowed to have values 0 (zero). In the other words, the categorical values 0 should be replaced with other values.

range.data

a matrix (\(2 \times n\)) containing the range of the data, where \(n\) is the number of variables, and first and second rows are the minimum and maximum values, respectively. It should be noted that for "FRBCS.W", "FRBCS.CHI", "GFS.GCCL", "FH.GBML", and "SLAVE", \(n\) represents the number of input variables only (without the output variable). It will be assigned as min/max of training data if it is omitted.

method.type

this parameter determines the learning algorithm to be used. The following methods are implemented:

"WM": Wang and Mendel's technique to handle regression tasks. See WM;
"SBC": subtractive clustering method to handle regression tasks. See SBC;
"HYFIS": hybrid neural fuzzy inference systems to handle regression tasks. See HyFIS;
"ANFIS": adaptive neuro-fuzzy inference systems to handle regression tasks. See ANFIS;
"FRBCS.W": fuzzy rule-based classification systems with weight factor based on Ishibuchi's method to handle classification tasks. See FRBCS.W;
"FRBCS.CHI": fuzzy rule-based classification systems based on Chi's method to handle classification tasks. See FRBCS.CHI;
"DENFIS": dynamic evolving neuro-fuzzy inference systems to handle regression tasks. See DENFIS;
"FS.HGD": fuzzy system using heuristic and gradient descent method to handle regression tasks. See FS.HGD;
"FIR.DM": fuzzy inference rules by descent method to handle regression tasks. See FIR.DM;
"GFS.FR.MOGUL": genetic fuzzy systems for fuzzy rule learning based on the MOGUL methodology to handle regression tasks. See GFS.FR.MOGUL;
"GFS.THRIFT": Thrift's technique based on genetic algorithms to handle regression tasks. See GFS.Thrift;
"GFS.GCCL": Ishibuchi's method based on genetic cooperative-competitive learning to handle classification tasks. See GFS.GCCL;
"FH.GBML": Ishibuchi's method based on hybridization of genetic cooperative-competitive learning and Pittsburgh to handle classification tasks. See FH.GBML;
"SLAVE": structural learning algorithm on vague environment to handle classification tasks. See SLAVE;
"GFS.LT.RS": genetic algorithm for lateral tuning and rule selection. See GFS.LT.RS

control

a list containing all arguments, depending on the learning algorithm to use. The following list are parameters required for each methods, whereas their descriptions will be explained later on.

WM:

list(num.labels, type.mf, type.tnorm, type.defuz,

type.implication.func, name)
HYFIS:

list(num.labels, max.iter, step.size, type.tnorm,

type.defuz, type.implication.func, name)
ANFIS and FIR.DM:

list(num.labels, max.iter, step.size,

type.tnorm, type.implication.func , name)
SBC:

list(r.a, eps.high, eps.low, name)
FS.HGD:

list(num.labels, max.iter, step.size, alpha.heuristic,

type.tnorm, type.implication.func, name)
FRBCS.W and FRBCS.CHI:

list(num.labels, type.mf, type.tnorm,

type.implication.func, name)
DENFIS method:

list(Dthr, max.iter, step.size, d, name)
GFS.FR.MOGUL:

list(persen_cross, max.iter, max.gen, max.tune,

persen_mutant, epsilon, name)
GFS.THRIFT method:

list(popu.size, num.labels, persen_cross,

max.gen, persen_mutant, type.tnorm, type.defuz,

type.implication.func, name)
GFS.GCCL:

list(popu.size, num.class, num.labels, persen_cross,

max.gen, persen_mutant, name)
FH.GBML:

list(popu.size, max.num.rule, num.class, persen_cross,

max.gen, persen_mutant, p.dcare, p.gccl, name)
SLAVE:

list(num.class, num.labels, persen_cross, max.iter,

max.gen, persen_mutant, k.lower, k.upper, epsilon, name)
GFS.LT.RS:

list(popu.size, num.labels, persen_mutant, max.gen,

mode.tuning, type.tnorm, type.implication.func,

type.defuz, rule.selection, name)

Description of the control Parameters

num.labels: a positive integer to determine the number of labels (linguistic terms). The default value is 7.
type.mf: the following type of the membership function. The default value is GAUSSIAN. For more detail, see fuzzifier.
- TRIANGLE: it refers triangular shape.
- TRAPEZOID: it refers trapezoid shape.
- GAUSSIAN: it refers gaussian shape.
- SIGMOID: it refers sigmoid.
- BELL: it refers generalized bell.
type.defuz: the type of the defuzzification method as follows. The default value is WAM. For more detail, see defuzzifier.
- WAM: the weighted average method.
- FIRST.MAX: the first maxima.
- LAST.MAX: the last maxima.
- MEAN.MAX: the mean maxima.
- COG: the modified center of gravity (COG).
type.tnorm: the type of conjunction operator (t-norm). The following are options of t-norm available. For more detail, please have a look at inference. The default value is MIN.
- MIN means standard type (minimum).
- HAMACHER means Hamacher product.
- YAGER means Yager class (with tao = 1).
- PRODUCT means product.
- BOUNDED mean bounded product.
type.snorm: the type of disjunction operator (s-norm). The following are options of s-norm available. For more detail, please have a look at inference. The default value is MAX.
- MAX means standard type (maximum).
- HAMACHER means Hamacher sum.
- YAGER means Yager class (with tao = 1).
- SUM means sum.
- BOUNDED mean bounded sum.
type.implication.func: the type of implication function. The following are options of implication function available: DIENES_RESHER, LUKASIEWICZ, ZADEH, GOGUEN, GODEL, SHARP, MIZUMOTO, DUBOIS_PRADE, and MIN. For more detail, please have a look at WM. The default value is ZADEH.
name: a name for the model. The default value is "sim-0".
max.iter: a positive integer to determine the maximal number of iterations. The default value is 10.
step.size: the step size of the gradient descent, a real number between 0 and 1. The default value is 0.01.
r.a: a positive constant which is effectively the radius defining a neighborhood. The default value is 0.5.
eps.high: an upper threshold value. The default value is 0.5.
eps.low: a lower threshold value. The default value is 0.15.
alpha.heuristic: a positive real number representing a heuristic value. The default value is 1.
Dthr: the threshold value for the envolving clustering method (ECM), between 0 and 1. The default value is 0.1.
d: a parameter for the width of the triangular membership function. The default value is 2.
persen_cross: a probability of crossover. The default value is 0.6.
max.gen: a positive integer to determine the maximal number of generations of the genetic algorithm. The default value is 10.
max.tune: a positive integer to determine the maximal number of tuning iterations. The default value is 10.
persen_mutant: a probability of mutation. The default value is 0.3.
epsilon: a real number between 0 and 1 representing the level of generalization. A high epsilon can lead to overfitting. The default value is 0.9.
popu.size: the size of the population which is generated in each generation. The default value is 10.
max.num.rule: the maximum size of the rules. The default value is 5.
num.class: the number of classes.
p.dcare: a probability of "don't care" attributes. The default value is 0.5.
p.gccl: a probability of the GCCL process. The default value is 0.5.
k.lower: a lower bound of the noise threshold with interval between 0 and 1. The default value is 0.
k.upper: an upper bound of the noise threshold with interval between 0 and 1. The default value is 1.
mode.tuning: a type of lateral tuning which are "LOCAL" or "GLOBAL". The default value is "GLOBAL".
rule.selection:a boolean value representing whether performs rule selection or not. The default value is "TRUE".

Value

The frbs-object.

Details

This function makes accessible all learning methods that are implemented in this package. All of the methods use this function as interface for the learning stage, so users do not need to call other functions in the learning phase. In order to obtain good results, users need to adjust some parameters such as the number of labels, the type of the shape of the membership function, the maximal number of iterations, the step size of the gradient descent, or other method-dependent parameters which are collected in the control parameter. After creating the model using this function, it can be used to predict new data with predict.

Examples

Run this code

# NOT RUN {
##################################
## I. Regression Problem
## Suppose data have two input variables and one output variable.  
## We separate them into training, fitting, and testing data.
## data.train, data.fit, data.test, and range.data are inputs 
## for all regression methods.
###################################
## Take into account that the simulation might take a long time 
## depending on the hardware you are using. The chosen parameters 
## may not be optimal.
## Data must be in data.frame or matrix form and the last column 
## is the output variable/attribute.
## The training data must be expressed in numbers (numerical data).
data.train <- matrix(c(5.2, -8.1, 4.8, 8.8, -16.1, 4.1, 10.6, -7.8, 5.5, 10.4, -29.0, 
                      5.0, 1.8, -19.2, 3.4, 12.7, -18.9, 3.4, 15.6, -10.6, 4.9, 1.9, 
                      -25.0, 3.7, 2.2, -3.1, 3.9, 4.8, -7.8, 4.5, 7.9, -13.9, 4.8, 
                      5.2, -4.5, 4.9, 0.9, -11.6, 3.0, 11.8, -2.1, 4.6, 7.9, -2.0, 
                      4.8, 11.5, -9.0, 5.5, 10.6, -11.2, 4.5, 11.1, -6.1, 4.7, 12.8, 
                      -1.0, 6.6, 11.3, -3.6, 5.1, 1.0, -8.2, 3.9, 14.5, -0.5, 5.7, 
                      11.9, -2.0, 5.1, 8.1, -1.6, 5.2, 15.5, -0.7, 4.9, 12.4, -0.8, 
                      5.2, 11.1, -16.8, 5.1, 5.1, -5.1, 4.6, 4.8, -9.5, 3.9, 13.2, 
                      -0.7, 6.0, 9.9, -3.3, 4.9, 12.5, -13.6, 4.1, 8.9, -10.0, 
                      4.9, 10.8, -13.5, 5.1), ncol = 3, byrow = TRUE)
colnames(data.train) <- c("inp.1", "inp.2", "out.1") 

data.fit <- data.train[, -ncol(data.train)]

data.test <- matrix(c(10.5, -0.9, 5.8, -2.8, 8.5, -0.6, 13.8, -11.9, 9.8, -1.2, 11.0,
                     -14.3, 4.2, -17.0, 6.9, -3.3, 13.2, -1.9), ncol = 2, byrow = TRUE)

range.data <- matrix(apply(data.train, 2, range), nrow = 2)

#############################################################
## I.1 Example: Constructing an FRBS model using Wang & Mendel
#############################################################
method.type <- "WM" 

## collect control parameters into a list
## num.labels = 3 means we define 3 as the number of linguistic terms
control.WM <- list(num.labels = 3, type.mf = "GAUSSIAN", type.tnorm = "MIN",  
type.defuz = "WAM", type.implication.func = "ZADEH", name = "Sim-0") 

## generate the model and save it as object.WM
object.WM <- frbs.learn(data.train, range.data, method.type, control.WM)

#############################################################
## I.2 Example: Constructing an FRBS model using SBC
#############################################################
# }
# NOT RUN {
method.type <- "SBC" 
control.SBC <- list(r.a = 0.5, eps.high = 0.5, eps.low = 0.15, name = "Sim-0")

object.SBC <- frbs.learn(data.train, range.data, method.type, control.SBC)
# }
# NOT RUN {
#############################################################
## I.3 Example: Constructing an FRBS model using HYFIS
#############################################################
# }
# NOT RUN {
method.type <- "HYFIS"

control.HYFIS <- list(num.labels = 5, max.iter = 50, step.size = 0.01, type.tnorm = "MIN", 
                      type.defuz = "COG", type.implication.func = "ZADEH", name = "Sim-0")

object.HYFIS <- frbs.learn(data.train, range.data, method.type, control.HYFIS)
# }
# NOT RUN {
#############################################################
## I.4 Example: Constructing an FRBS model using ANFIS
#############################################################
# }
# NOT RUN {
method.type <- "ANFIS" 

control.ANFIS <- list(num.labels = 5, max.iter = 10, step.size = 0.01, type.tnorm = "MIN", 
                      type.implication.func = "ZADEH", name = "Sim-0") 

object.ANFIS <- frbs.learn(data.train, range.data, method.type, control.ANFIS)
# }
# NOT RUN {
#############################################################
## I.5 Example: Constructing an FRBS model using DENFIS
#############################################################

# }
# NOT RUN {
control.DENFIS <- list(Dthr = 0.1, max.iter = 10, step.size = 0.001, d = 2, 
                       name = "Sim-0")
method.type <- "DENFIS"

object.DENFIS <- frbs.learn(data.train, range.data, method.type, control.DENFIS)
# }
# NOT RUN {
#############################################################
## I.6 Example: Constructing an FRBS model using FIR.DM
#############################################################
# }
# NOT RUN {
method.type <- "FIR.DM"
 
control.DM <- list(num.labels = 5, max.iter = 10, step.size = 0.01, type.tnorm = "MIN", 
                     type.implication.func = "ZADEH", name = "Sim-0") 
object.DM <- frbs.learn(data.train, range.data, method.type, control.DM)
# }
# NOT RUN {
#############################################################
## I.7 Example: Constructing an FRBS model using FS.HGD
#############################################################
# }
# NOT RUN {
method.type <- "FS.HGD" 
 
control.HGD <- list(num.labels = 5, max.iter = 10, step.size = 0.01, 
               alpha.heuristic = 1, type.tnorm = "MIN",  
               type.implication.func = "ZADEH", name = "Sim-0") 
object.HGD <- frbs.learn(data.train, range.data, method.type, control.HGD)
# }
# NOT RUN {
#############################################################
## I.8 Example: Constructing an FRBS model using GFS.FR.MOGUL
#############################################################
# }
# NOT RUN {
method.type <- "GFS.FR.MOGUL" 
 
control.GFS.FR.MOGUL <- list(persen_cross = 0.6, 
                    max.iter = 5, max.gen = 2, max.tune = 2, persen_mutant = 0.3, 
                    epsilon = 0.8, name="sim-0") 
object.GFS.FR.MOGUL <- frbs.learn(data.train, range.data, 
                       method.type, control.GFS.FR.MOGUL)
# }
# NOT RUN {
#############################################################
## I.9 Example: Constructing an FRBS model using Thrift's method (GFS.THRIFT)
#############################################################
# }
# NOT RUN {
method.type <- "GFS.THRIFT" 
 
control.Thrift <- list(popu.size = 6, num.labels = 3, persen_cross = 1, 
                      max.gen = 5, persen_mutant = 1, type.tnorm = "MIN", 
                      type.defuz = "COG", type.implication.func = "ZADEH", 
                      name="sim-0") 
object.Thrift <- frbs.learn(data.train, range.data, method.type, control.Thrift)
# }
# NOT RUN {
##############################################################
## I.10 Example: Constructing an FRBS model using
##      genetic for lateral tuning and rule selection (GFS.LT.RS)
#############################################################
## Set the method and its parameters
# }
# NOT RUN {
method.type <- "GFS.LT.RS" 
  
control.lt.rs <- list(popu.size = 5, num.labels = 5, persen_mutant = 0.3,
               max.gen = 10, mode.tuning = "LOCAL", type.tnorm = "MIN",  
                type.implication.func = "ZADEH", type.defuz = "WAM", 
                rule.selection = TRUE, name="sim-0")

## Generate fuzzy model
object.lt.rs <- frbs.learn(data.train, range.data, method.type, control.lt.rs)
# }
# NOT RUN {
#############################################################
## II. Classification Problems 
#############################################################
## The iris dataset is shuffled and divided into training and 
## testing data. Bad results in the predicted values may result
## from casual imbalanced classes in the training data.
## Take into account that the simulation may take a long time 
## depending on the hardware you use. 
## One may get better results with other parameters. 
## Data are in data.frame or matrix form and the last column is 
## the output variable/attribute
## The data must be expressed in numbers (numerical data).

data(iris)
irisShuffled <- iris[sample(nrow(iris)),]
irisShuffled[,5] <- unclass(irisShuffled[,5])
tra.iris <- irisShuffled[1:105,]
tst.iris <- irisShuffled[106:nrow(irisShuffled),1:4]
real.iris <- matrix(irisShuffled[106:nrow(irisShuffled),5], ncol = 1)

## Please take into account that the interval needed is the range of input data only.
range.data.input <- matrix(apply(iris[, -ncol(iris)], 2, range), nrow = 2)

######################################################### 
## II.1 Example: Constructing an FRBS model using 
##      FRBCS with weighted factor based on Ishibuchi's method
###############################################################
## generate the model
# }
# NOT RUN {
method.type <- "FRBCS.W"
control <- list(num.labels = 3, type.mf = "TRIANGLE", type.tnorm = "MIN", 
               type.implication.func = "ZADEH", name = "sim-0") 

object <- frbs.learn(tra.iris, range.data.input, method.type, control)

## conduct the prediction process
res.test <- predict(object, tst.iris)
# }
# NOT RUN {
######################################################### 
## II.2 Example: Constructing an FRBS model using 
##      FRBCS based on Chi's method
###############################################################
## generate the model
# }
# NOT RUN {
method.type <- "FRBCS.CHI"
control <- list(num.labels = 7, type.mf = "TRIANGLE", type.tnorm = "MIN", 
               type.implication.func = "ZADEH", name = "sim-0") 

object <- frbs.learn(tra.iris, range.data.input, method.type, control)

## conduct the prediction process
res.test <- predict(object, tst.iris)
# }
# NOT RUN {
######################################################### 
## II.3 The example: Constructing an FRBS model using GFS.GCCL
###############################################################
# }
# NOT RUN {
method.type <- "GFS.GCCL" 

control <- list(popu.size = 5, num.class = 3, num.labels = 5, persen_cross = 0.9, 
                    max.gen = 2, persen_mutant = 0.3,
                    name="sim-0") 
## Training process
## The main result of the training is a rule database which is used later for prediction.
object <- frbs.learn(tra.iris, range.data.input, method.type, control)

## Prediction process
res.test <- predict(object, tst.iris)
# }
# NOT RUN {
######################################################### 
## II.4 Example: Constructing an FRBS model using FH.GBML
###############################################################
# }
# NOT RUN {
method.type <- "FH.GBML" 
 
control <- list(popu.size = 5, max.num.rule = 5, num.class = 3, 
			persen_cross = 0.9, max.gen = 2, persen_mutant = 0.3, p.dcare = 0.5, 
             p.gccl = 1, name="sim-0") 
 
## Training process
## The main result of the training is a rule database which is used later for prediction.
object <- frbs.learn(tra.iris, range.data.input, method.type, control)

## Prediction process
res.test <- predict(object, tst.iris)
# }
# NOT RUN {
######################################################### 
## II.5 The example: Constructing an FRBS model using SLAVE
###############################################################
# }
# NOT RUN {
method.type <- "SLAVE" 
 
control <- list(num.class = 3, num.labels = 5,
			persen_cross = 0.9, max.iter = 5, max.gen = 3, persen_mutant = 0.3, 
             k.lower = 0.25, k.upper = 0.75, epsilon = 0.1, name="sim-0") 
 
## Training process
## The main result of the training is a rule database which is used later for prediction.
object <- frbs.learn(tra.iris, range.data.input, method.type, control)

## Prediction process
res.test <- predict(object, tst.iris)
# }
# NOT RUN {
# }

Run the code above in your browser using DataLab