In ChaddFrasier/planetLearn: Planetary Learning Package

library(knitr)
library(planetLearn)
library(neuralnet)
library(ggplot2)
library(dplyr)

# document init
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)
# configuration for all code blocks
knitr::opts_chunk$set(out.width = "100%",
                      text = element_text(size=5),
                      axis.text.x = element_text(angle=90, hjust=1),
                      memory.limit(20000000),
                      setwd("D:/cfrasier/work/R"))

Neural Network

Neural Networks are great at finding hidden relationships between the data in matrices as a whole but it is suseptable to overfitting when too many features are used. Yet we though this would be a viable means of training models for the fact that we are unsure of the exact linear realtionships between all the features used.

Read the Data in R

# load the data from the data folder
load(file = "../../data/lunarData.rda")

# filter out incidence angles greater than some number of degrees
data <- lunarData[ lunarData$Incidence <= 70, ]

# check for 0 variance
sc <- scale(data[1:200000,-1])

# Splitting the data into test and train
set.seed(150)

# train with only a 80 % of the total data
train = sample( 1:nrow(sc), size = (nrow(sc) * 0.60) )
test = (-train)

trainNN <- cbind(DN=data[train,1], sc[train,])
testNN <- cbind(DN=data[test,1], sc[test,])

Train Model

# fit neural network
NN <- neuralnet(DN ~ Phase + Emission + Incidence + LEmission + LIncidence + Lat + Long + SunAz + CraftAz, data=trainNN, hidden = c(12,6) , linear.output = TRUE )

# plot neural network
plot(NN)