Skip to content

Latest commit

 

History

History
131 lines (103 loc) · 5.15 KB

README_(English).md

File metadata and controls

131 lines (103 loc) · 5.15 KB

Introduction_of__ML_by_R

Practice to construct model of Mechanical Learning (ML) by R and the memo. I try to use below packages and to introduce ML to my reserach that is by Molecular Dynamics (MD) simulation. To keep accuracy of this README, I write on Japanese, README (Japanese).md, too just in case.

Overview

My study investigates the characteristics that metal cations selectivity adsorb to micro porous carbon with applied voltage by MD simulation. In this trial, I construct the ML model that predicts the probability pred_P that metal cations adsorb to a pore with 7 parameters, mass, valent, the first/second hydration radius r1/ r2, the maximum value of RDF gr_max, voltage vol and pore diameter pore_d, by R. The number of dataset is 157.

Description

Packages

packages about ML
caret package

packages about MPI
doParallel package

packages about NN
nnet packeage
neuralnet package

Installation

install.packeges('[package name]')
(When you use packages after installation, don't forget library([package name]).)

How to use each packages

doParallel package

This package can calculate with MPI on your PC.

cl_n <- detectCores()
cl <- makePSOCKcluster(cl_n)
registerDoParallel(cl)

nnet package

This package can construct 3 layers NN.
nn <- nnet(train_data, train_labels, size = 100, rang = 0.5, decay = 0, maxit = 1000)
Optimizer: BFGS method
act. func.: softmax or linear conbination(?)
err. func.: MSE or closs enthoropy(?)
About this detail, please shows here.

neuralnet package

This package can construct large scare NN model.

 nn <- neuralnet(formula = formula.nn, 
                  data = train,
                  hidden = c(5), #the number of nodes in hidden layers: c(1st layer, 2nd layer, 3rd layer,,,)
                  act.fct="logistic", #select activated func. logistic/tanh
                  learningrate = 0.01, 
                  threshold = 0.01,
                  stepmax = 1e+07,
                  err.fct = "sse", #select error func. ce(logistics)/sse()
                  startweights = NULL,
                  learningrate.factor = list(minus = 10.0, plus=10.0),
                  algorithm = "backprop",
                  linear.output=TRUE #If this model is logistic regression, this term sets "FALSE"
  )

Optimizer: Back propagation, RPROP+ or RPROP-.
act. func.: logistic sigmoid, tanh or linear conbination
err. func.: MSE or closs enthoropy
About this detail, please shows here.

caret package

This package can deal with many ML. In this trial, I use it for data processing mainly.
About this detail, please shows here.

Results and Discussion

normalization

In this trial, I normalize with below equation because the dataset is uniform distribution.

norm <- function(x){
  return((x-min(x)) / (max(x)-min(x)))
}

multiple regression model

I tried to construct multiple regression model by below codes.

#sinple multiple regression
sol.m <- lm(s_data$Ptotal~.,data=s_data)
#multiple regression considered exchange interaction 
sol.m <- lm(s_data$Ptotal~.^2,data=s_data)

Here, s_data expresses the above 7 parameters.

This model optimized by step(sol.m) function that carries out variable reduction method.

NN model by neuralnet package

I construced NN model by above codes.

good points
- it can build DL model.
- it can choose some act. func. and optimizer.
- it can plot NN model.

bad points
- neuralnet function can't deal with unknown dataset because this function specifies both parameters and labels. So, I don't find how to introduce unknown dataset to themodel.
- it can't use ReLU function that is high performance act. function.
- it can't process complex DL model with personal PC although it can construct DL model.

NN model by nnet package

I construced NN model by above codes.

good points
- easy.
- output leraning process on console.
- speedy than the neuralnet package.

bad points
- restrictive. it make less than 3 layers NN.
- it can't plot.
- The convergence by this optimizer is not global minimum.
- I've not understood how to use activated function yet.

Conclusion

In conclusion, I could construct NN model. In the future, I try to intoduce more packages because my study give only little datasets. The package of R has many good points and bad points. So, I want to introdce too tensorflow of python that is genelic.

Especially, I want to more study about caret package because I feel it is very convinient.

My Problems

I show my problems below.
- Is it possible to use the values that are over the range of training dataset as parameters for prediction?
- How to normalize unknown dataset?
- How to use neuralnet package for prediction with unknown dataset?
- How to define reLU function on R?