| ga_parsimony {GAparsimony} | R Documentation |
GA-PARSIMONY
Description
A GA-based optimization method for searching accurate parsimonious models by combining feature selection, model tuning, and parsimonious model selection (PMS). PMS procedure is based on separate cost and complexity evaluations. The best individuals are initially sorted by an error fitness function, and afterwards, models with similar costs are rearranged according to their model complexity so as to foster models of lesser complexity. The algorithm can be run sequentially or in parallel using an explicit master-slave parallelisation.
Usage
ga_parsimony(fitness, ...,
min_param, max_param, nFeatures,
names_param=NULL, names_features=NULL,
object=NULL, iter_ini=NULL,
type_ini_pop="improvedLHS",
popSize = 50, pcrossover = 0.8, maxiter = 40,
feat_thres=0.90, rerank_error = 0.0, iter_start_rerank = 0,
pmutation = 0.10, feat_mut_thres=0.10, not_muted=3,
elitism = base::max(1, round(popSize * 0.20)),
population = parsimony_population,
selection = parsimony_nlrSelection,
crossover = parsimony_crossover,
mutation = parsimony_mutation,
keep_history = FALSE,
path_name_to_save_iter = NULL,
early_stop = maxiter, maxFitness = Inf, suggestions = NULL,
parallel = FALSE,
monitor = if (interactive()) parsimony_monitor else FALSE,
seed_ini = NULL, verbose=FALSE)
Arguments
fitness |
the fitness function, any allowable R function which takes as input an individual Note: the chromosome is a concatenated real vector with the model parameters (parameters-chromosome) and the binary selection of the input features (features-chromosome). For example, a chromosome defined as c(10, 0.01, 0,1,1,0,1,0,0) could corresponds to a SVR model parameters C=10 & gamma=0.01, and a selection of three input features (second, third and fifth) from a dataset of 7 features (0110100). |
... |
additional arguments to be passed to the fitness function. This allows to write fitness functions that keep some variables fixed during the search. |
min_param |
a vector of length equal to the model parameters providing the minimum of the search space. |
max_param |
a vector of length equal to the model parameters providing the maximum of the search space. |
nFeatures |
a value specifying the number of maximum input features. |
names_param |
a vector with the name of the model parameters. |
names_features |
a vector with the name of the input features. |
object |
object of 'ga_parsimony' class to continue GA process. 'ga_parsimony@history' must be provided. Note: all GA settings are obtained from 'object' in order to continue the GA process. |
iter_ini |
Iteration/generation of 'object@history' to be used when 'object' is provided. If 'iter_ini==NULL' uses the last iteration of 'object'. |
type_ini_pop |
method to create the first population with 'parsimony_population' function. This function is called when iter_ini==0 and 'suggestions' are not provided. Methods='randomLHS','geneticLHS','improvedLHS','maximinLHS','optimumLHS','random'. First 5 methods correspond with several latine hypercube sampling. |
popSize |
the population size. |
pcrossover |
the probability of crossover between pairs of chromosomes. Typically this is a large value and by default is set to 0.8. |
maxiter |
the maximum number of iterations to run before the GA process is halted. |
feat_thres |
proportion of selected features in the initial population. It is recommended a high percentage of selected features for the first generations. By default is set to 0.90. |
rerank_error |
when a value is provided, a second reranking process according to the model complexities is called by parsimony_rerank function. Its primary objective is to select individuals with high validation cost while maintaining the robustness of a parsimonious model. This function switches the position of two models if the first one is more complex than the latter and no significant difference is found between their fitness values in terms of cost. Therefore, if the absolute difference between the validation costs are lower than 'rerank_error' they are considered similar. Default value=0.01 |
iter_start_rerank |
iteration when ReRanking process is actived. Default=0. Sometimes is useful not to use ReRanking process in the first generations. |
pmutation |
the probability of mutation in a parent chromosome. Usually mutation occurs with a small probability. By default is set to 0.10. |
feat_mut_thres |
probability of the muted 'features-chromosome' to be one. Default value is set to 0.10. |
not_muted |
number of the best elitists that are not muted in each generation. Default value is set to 3. |
elitism |
the number of best individuals to survive at each generation. By default the top 20% individuals will survive at each iteration. |
population |
an R function for randomly generating an initial population. See |
selection |
an R function performing selection, i.e. a function which generates a new population of individuals from the current population probabilistically according to individual fitness. See |
crossover |
an R function performing crossover, i.e. a function which forms offsprings by combining part of the genetic information from their parents. See |
mutation |
an R function performing mutation, i.e. a function which randomly alters the values of some genes in a parent chromosome. See |
keep_history |
If it is TRUE keeps in the list |
path_name_to_save_iter |
If it is not NULL save the 'ga_parsimony' object to the 'path_name_to_save_iter' file at the end of each iteration. Note: use extension '.RData', example 'object.RData' |
early_stop |
the number of consecutive generations without any improvement in the best fitness value before the GA is stopped. |
maxFitness |
the upper bound on the fitness function after that the GA search is interrupted. Default value is set to +Inf |
suggestions |
a matrix of solutions strings to be included in the initial population. If provided the number of columns must match (object@nParams+object@nFeatures). Can be used a previous population, for example: 'ga_parsimony@history[[2]]$population'. |
parallel |
a logical argument specifying if parallel computing should be used ( |
monitor |
a logical or an R function which takes as input the current state of the |
seed_ini |
an integer value containing the random number generator state. This argument can be used to replicate the results of a GA search. Note that if parallel computing is required, the doRNG package must be installed. |
verbose |
if it is TRUE shows additional information for debugging. |
Details
GAparsimony package is a new GA wrapper automatic procedure that efficiently generated prediction models with reduced complexity and adequate generalization capacity.
ga_parsimony function is primarily based on combining feature selection and model parameter tuning with a second novel GA selection process (parsimony_rerank function), in order to achieve better overall parsimonious models.
Unlike other GA methodologies that use a penalty parameter for combining loss and complexity measures into a unique fitness function, the main contribution of this package is that ga_parsimony selects the best models by considering cost and complexity separately. For this purpose, the ReRank algorithm rearranges individuals by their complexity when there is not a significant difference between their costs. Thus, less complex models with similar accuracy are promoted. Furthermore, because the penalty parameter is unnecessary, there is no consequent uncertainty associated with assigning a correct value beforehand. As a result, with GA-PARSIMONY, an automatic method for obtaining parsimonious models is finally made possible.
Value
Returns an object of class ga_parsimony-class. See ga_parsimony-class for a description of available slots information.
Author(s)
Francisco Javier Martinez de Pison. fjmartin@unirioja.es. EDMANS Group. https://edmans.webs.com/
References
Urraca R., Sodupe-Ortega E., Antonanzas E., Antonanzas-Torres F., Martinez-de-Pison, F.J. (2017). Evaluation of a novel GA-based methodology for model structure selection: The GA-PARSIMONY. Neurocomputing, Online July 2017. https://doi.org/10.1016/j.neucom.2016.08.154
Sanz-Garcia A., Fernandez-Ceniceros J., Antonanzas-Torres F., Pernia-Espinoza A.V., Martinez-de-Pison F.J. (2015). GA-PARSIMONY: A GA-SVR approach with feature selection and parameter optimization to obtain parsimonious solutions for predicting temperature settings in a continuous annealing furnace. Applied Soft Computing 35, 23-38.
Fernandez-Ceniceros J., Sanz-Garcia A., Antonanzas-Torres F., Martinez-de-Pison F.J. (2015). A numerical-informational approach for characterising the ductile behaviour of the T-stub component. Part 2: Parsimonious soft-computing-based metamodel. Engineering Structures 82, 249-260.
Antonanzas-Torres F., Urraca R., Antonanzas J., Fernandez-Ceniceros J., Martinez-de-Pison F.J. (2015). Generation of daily global solar irradiation with support vector machines for regression. Energy Conversion and Management 96, 277-286.
See Also
ga_parsimony-class,
summary.ga_parsimony,
plot.ga_parsimony,
parsimony_Population,
parsimony_Selection,
parsimony_Crossover,
parsimony_Mutation,
parsimony_importance,
parsimony_rerank.
Examples
## Not run:
#################################
### Example 1: Classification ###
#################################
# This a toy example that shows how to search, for the *iris* database,
# a parsimony classification NNET model with 'GAparsimony'
# and 'caret' packages. Validation errors and iterations have been
# reduced to speedup the process
library(GAparsimony)
# Training and testing Datasets
library(caret)
data(iris)
# Z-score of input features
iris_esc <- data.frame(scale(iris[,1:4]),Species=iris[,5])
# Define an 70
set.seed(1234)
inTraining <- createDataPartition(iris_esc$Species, p=.70, list=FALSE)
data_train <- iris_esc[ inTraining,]
data_test <- iris_esc[-inTraining,]
# Function to evaluate each SVM individual
# ----------------------------------------
fitness_SVM <- function(chromosome, ...)
{
# First two values in chromosome are 'C' & 'sigma' of 'svmRadial' method
tuneGrid <- data.frame(C=chromosome[1],sigma=chromosome[2])
# Next values of chromosome are the selected features (TRUE if > 0.50)
selec_feat <- chromosome[3:length(chromosome)]>0.50
# Return -Inf if there is not selected features
if (sum(selec_feat)<1) return(c(kappa_val=-Inf,kappa_test=-Inf,complexity=Inf))
# Extract features from the original DB plus response (last column)
data_train_model <- data_train[,c(selec_feat,TRUE)]
data_test_model <- data_test[,c(selec_feat,TRUE)]
# Validate each individual with only a 2-CV
# Yo obtain a robust validation measure
# use 'repeatedcv' with more folds and times
# (see 2nd and 3rd examples...)
train_control <- trainControl(method = "cv",number = 5)
# train the model
set.seed(1234)
model <- train(Species ~ ., data=data_train_model,
trControl=train_control,
method="svmRadial", metric="Kappa",
tuneGrid=tuneGrid, verbose=FALSE)
# Extract validation and test accuracy
accuracy_val <- model$results$Accuracy
accuracy_test <- postResample(pred=predict(model, data_test_model),
obs=data_test_model[,ncol(data_test_model)])[2]
# Obtain Complexity = Num_Features*1E6+Number of support vectors
complexity <- sum(selec_feat)*1E6+model$finalModel@nSV
# Return(validation accuracy, testing accuracy, model_complexity)
vect_errors <- c(accuracy_val=accuracy_val,
accuracy_test=accuracy_test,complexity=complexity)
return(vect_errors)
}
# ---------------------------------------------------------------------------------
# Search the best parsimonious model with GA-PARSIMONY by using Feature Selection,
# Parameter Tuning and Parsimonious Model Selection
# ---------------------------------------------------------------------------------
library(GAparsimony)
# Ranges of size and decay
min_param <- c(0.0001, 0.00001)
max_param <- c(0.9999, 0.99999)
names_param <- c("C","sigma")
# ga_parsimony can be executed with a different set of 'rerank_error' values
rerank_error <- 0.001
GAparsimony_model <- ga_parsimony(fitness=fitness_SVM,
min_param=min_param,
max_param=max_param,
names_param=names_param,
nFeatures=ncol(data_train)-1,
names_features=colnames(data_train)[-ncol(data_train)],
keep_history = TRUE,
rerank_error = rerank_error,
popSize = 20,
maxiter = 20,
early_stop=7,
feat_thres=0.90,# Perc selec features in first iter
feat_mut_thres=0.10,# Prob. feature to be 1 in mutation
not_muted=1,
parallel = FALSE, # speedup with 'n' cores or all with TRUE
seed_ini = 1234)
print(paste0("Best Parsimonious SVM with C=",
GAparsimony_model@bestsolution['C'],
" sigma=",
GAparsimony_model@bestsolution['sigma'],
" -> ",
" AccuracyVal=",
round(GAparsimony_model@bestsolution['fitnessVal'],6),
" AccuracyTest=",
round(GAparsimony_model@bestsolution['fitnessTst'],6),
" Num Features=",
round(GAparsimony_model@bestsolution['complexity']/1E6,0),
" Complexity=",
round(GAparsimony_model@bestsolution['complexity'],2)))
print(summary(GAparsimony_model))
print(parsimony_importance(GAparsimony_model))
#################################
### Example 2: Classification ###
#################################
#This example shows how to search, for the *Sonar* database,
#a parsimony classification SVM model with 'GAparsimony' and 'caret' packages.
# Training and testing Datasets
library(caret)
library(GAparsimony)
library(mlbench)
data(Sonar)
set.seed(1234)
inTraining <- createDataPartition(Sonar$Class, p=.80, list=FALSE)
data_train <- Sonar[ inTraining,]
data_test <- Sonar[-inTraining,]
# Function to evaluate each SVM individual
# ----------------------------------------
fitness_SVM <- function(chromosome, ...)
{
# First two values in chromosome are 'C' & 'sigma' of 'svmRadial' method
tuneGrid <- data.frame(C=chromosome[1],sigma=chromosome[2])
# Next values of chromosome are the selected features (TRUE if > 0.50)
selec_feat <- chromosome[3:length(chromosome)]>0.50
# Return -Inf if there is not selected features
if (sum(selec_feat)<1) return(c(kappa_val=-Inf,kappa_test=-Inf,complexity=Inf))
# Extract features from the original DB plus response (last column)
data_train_model <- data_train[,c(selec_feat,TRUE)]
data_test_model <- data_test[,c(selec_feat,TRUE)]
# How to validate each individual
# 'repeats' could be increased to obtain a more robust validation metric. Also,
# 'number' of folds could be adjusted to improve the measure.
train_control <- trainControl(method = "repeatedcv",number = 10,repeats = 10)
# train the model
set.seed(1234)
model <- train(Class ~ ., data=data_train_model, trControl=train_control,
method="svmRadial", metric="Kappa",
tuneGrid=tuneGrid, verbose=FALSE)
# Extract kappa statistics (repeated k-fold CV and testing kappa)
kappa_val <- model$results$Kappa
kappa_test <- postResample(pred=predict(model, data_test_model),
obs=data_test_model[,ncol(data_test_model)])[2]
# Obtain Complexity = Num_Features*1E6+Number of support vectors
complexity <- sum(selec_feat)*1E6+model$finalModel@nSV
# Return(validation error, testing error, model_complexity)
vect_errors <- c(kappa_val=kappa_val,kappa_test=kappa_test,complexity=complexity)
return(vect_errors)
}
# ---------------------------------------------------------------------------------
# Search the best parsimonious model with GA-PARSIMONY by using Feature Selection,
# Parameter Tuning and Parsimonious Model Selection
# ---------------------------------------------------------------------------------
library(GAparsimony)
# Ranges of size and decay
min_param <- c(00.0001, 0.00001)
max_param <- c(99.9999, 0.99999)
names_param <- c("C","sigma")
# ga_parsimony can be executed with a different set of 'rerank_error' values
rerank_error <- 0.001
# 40 individuals per population, 100 max generations with an early stopping
# of 10 generations (CAUTION! 7.34 minutes with 8 cores)!!!!!
GAparsimony_model <- ga_parsimony(fitness=fitness_SVM,
min_param=min_param,
max_param=max_param,
names_param=names_param,
nFeatures=ncol(data_train)-1,
names_features=colnames(data_train)[-ncol(data_train)],
keep_history = TRUE,
rerank_error = rerank_error,
popSize = 40,
maxiter = 100,
early_stop=10,
feat_thres=0.90,# Perc selec features in first iter
feat_mut_thres=0.10,# Prob. feature to be 1 in mutation
parallel = TRUE, seed_ini = 1234)
print(paste0("Best Parsimonious SVM with C=",
GAparsimony_model@bestsolution['C'],
" sigma=",
GAparsimony_model@bestsolution['sigma'],
" -> ",
" KappaVal=",
round(GAparsimony_model@bestsolution['fitnessVal'],6),
" KappaTst=",
round(GAparsimony_model@bestsolution['fitnessTst'],6),
" Num Features=",
round(GAparsimony_model@bestsolution['complexity']/1E6,0),
" Complexity=",
round(GAparsimony_model@bestsolution['complexity'],2)))
print(summary(GAparsimony_model))
# Plot GA evolution ('keep_history' must be TRUE)
elitists <- plot(GAparsimony_model)
# Percentage of appearance of each feature in elitists
print(parsimony_importance(GAparsimony_model))
#############################
### Example 3: Regression ###
#############################
# This example shows how to search, for the *Boston* database, a parsimony regressor ANN
# model with 'GAparsimony' and 'caret' packages.
# Load Boston database and scale it
library(MASS)
data(Boston)
Boston_scaled <- data.frame(scale(Boston))
# Define an 80
set.seed(1234)
trainIndex <- createDataPartition(Boston[,"medv"], p=0.80, list=FALSE)
data_train <- Boston_scaled[trainIndex,]
data_test <- Boston_scaled[-trainIndex,]
# Restore 'Response' to original values
data_train[,ncol(data_train)] <- Boston$medv[trainIndex]
data_test[,ncol(data_test)] <- Boston$medv[-trainIndex]
print(dim(data_train))
print(dim(data_test))
# Function to evaluate each ANN individual
# ----------------------------------------
fitness_NNET <- function(chromosome, ...)
{
# First two values in chromosome are 'size' & 'decay' of 'nnet' method
tuneGrid <- data.frame(size=round(chromosome[1]),decay=chromosome[2])
# Next values of chromosome are the selected features (TRUE if > 0.50)
selec_feat <- chromosome[3:length(chromosome)]>0.50
if (sum(selec_feat)<1) return(c(rmse_val=-Inf,rmse_test=-Inf,complexity=Inf))
# Extract features from the original DB plus response (last column)
data_train_model <- data_train[,c(selec_feat,TRUE)]
data_test_model <- data_test[,c(selec_feat,TRUE)]
# How to validate each individual
# 'repeats' could be increased to obtain a more robust validation metric. Also,
# 'number' of folds could be adjusted to improve the measure.
train_control <- trainControl(method = "repeatedcv",number = 10,repeats = 5)
# train the model
set.seed(1234)
model <- train(medv ~ ., data=data_train_model, trControl=train_control,
method="nnet", tuneGrid=tuneGrid, trace=FALSE, linout = 1)
# Extract errors
rmse_val <- model$results$RMSE
rmse_test <- sqrt(mean((unlist(predict(model, newdata = data_test_model)) -
data_test_model$medv)^2))
# Obtain Complexity = Num_Features*1E6+sum(neural_weights^2)
complexity <- sum(selec_feat)*1E6+sum(model$finalModel$wts*model$finalModel$wts)
# Return(-validation error, -testing error, model_complexity)
# errors are negative because GA-PARSIMONY tries to maximize values
vect_errors <- c(rmse_val=-rmse_val,rmse_test=-rmse_test,complexity=complexity)
return(vect_errors)
}
# ---------------------------------------------------------------------------------
# Search the best parsimonious model with GA-PARSIMONY by using Feature Selection,
# Parameter Tuning and Parsimonious Model Selection
# ---------------------------------------------------------------------------------
library(GAparsimony)
# Ranges of size and decay
min_param <- c(1, 0.0001)
max_param <- c(25 , 0.9999)
names_param <- c("size","decay")
# ga_parsimony can be executed with a different set of 'rerank_error' values
rerank_error <- 0.01
# 40 individuals per population, 100 max generations with an early stopping
# of 10 generations (CAUTION! 33.89 minutes with 8 cores)!!!!!
GAparsimony_model <- ga_parsimony(fitness=fitness_NNET,
min_param=min_param,
max_param=max_param,
names_param=names_param,
nFeatures=ncol(data_train)-1,
names_features=colnames(data_train)[-ncol(data_train)],
keep_history = TRUE,
rerank_error = rerank_error,
popSize = 40,
maxiter = 100, # Change to 100
early_stop=10,
feat_thres=0.90,# Perc selec features in first iter
feat_mut_thres=0.10,# Prob. feature to be 1 in mutation
not_muted=2,
parallel = TRUE, seed_ini = 1234)
print(paste0("Best Parsimonious ANN with ",
round(GAparsimony_model@bestsolution['size']),
" hidden neurons and decay=",
GAparsimony_model@bestsolution['decay'],
" -> ",
" RMSEVal=",
round(-GAparsimony_model@bestsolution['fitnessVal'],6),
" RMSETst=",
round(-GAparsimony_model@bestsolution['fitnessTst'],6)))
print(summary(GAparsimony_model))
# Plot GA evolution ('keep_history' must be TRUE)
elitists <- plot(GAparsimony_model)
# Percentage of appearance of each feature in elitists
print(parsimony_importance(GAparsimony_model))
## End(Not run)