pml.Rmd

---
title: "pml"
author: "Graydyn Young"
date: "April 15, 2015"
output: html_document
---

The objective of this document is to determine how well a person is doing a bicep curl based off of accelerometer data.  We will be predicting 5 classes of curls, with class A representing good form, while classes B-E represent various common problems.

Starting by loading our data, removing all those pesky NAs, and then splitting the training set into train and cross validation sets.
I'm also removing any variables that don't look like predictors, such as timestamps.
```{r}
library(caret)
set.seed(1234)
trainFull <- read.csv("pml-training.csv")
test <- read.csv("pml-testing.csv")

#remove NA
trainFull <- trainFull[, colSums(is.na(trainFull)) == 0] 
test <- test[, colSums(is.na(test)) == 0] 

#remove the index (X), timestamps, and whatever 'window' is
classe <- trainFull$classe
remove <- grepl("^X|timestamp|window", names(trainFull))
trainFull <- trainFull[, !remove]
trainFull <- trainFull[, sapply(trainFull, is.numeric)]
trainFull$classe <- classe #need to put the classe back, since it isn't numeric
remove <- grepl("^X|timestamp|window", names(test))
test <- test[, !remove]
test <- test[, sapply(test, is.numeric)]

partition <- createDataPartition(trainFull$classe, p = .7,list = FALSE,times = 1)
train <- trainFull[partition, ]
cv <- trainFull[-partition,]
```

Lets train a model.  Starting with random forest.
```{r}
rfModel <- train(classe ~ ., data=train, method = "rf")
```

```{r}
rfPrediction <- predict(rfModel, cv)
print(paste("random forest - accuracy", toString(sum(cv$classe == rfPrediction) / nrow(cv)), " -expected out of sample error", toString(1 - sum(cv$classe == rfPrediction) / nrow(cv))))
```

I was going to try some other models as well, but I tried "rf" first and it seems highly unlikely that I'll find a better cross validation error rate than that.

Since the random forest performed so well, I'll use that to make my test predictions.

```{r}
testPredict <- predict(rfModel, test)
```