HW 13 Stat 705  Fall 2015

Assigned Monday 11/2/15 DUE Monday 11/16/15

Access the "Concrete" dataset and Readme file in the data directory
http://www.math.umd.edu/~slud/s705/Data/  on the course web-page.

Find the best predictive model you can for the "Concrete Compressive Strength"
variable in terms of the other variables. Note that this dataset was placed 
in  a "machine learning data repository" (UC Irvine) indicating that the best 
predictors are likely to be fairly complicated, so you shoul expect that 
nonlinear recodes of the variables, including possibly high-order interactions, 
are likely to be very important. But confince yourself to regression 
techniques, using whatever recodes and nonlinear terms and interactions you 
can think of. There is a lot of data here, so you may want to try splitting the
data, fitting a model on part and testing it on the rest.

NOTE: please read the descriptions of variables in the "README" file. Note 
for example that the first 7 variables, measured in kg per cubic meter 
mixture, should probably be viewed as proportions, so maybe recoding the 
variables X1 - X7 to   X1/(X1+...+X7) etc. will help.

Again, the prupose is not an "optimal" data analysis, but rather a use of R 
tools to use graphics and linear (and nonlinear if you like -- e.g., nlme) 
modeling techniques to achieve the best prediction in terms of mean-squared 
error that you can.

Given that the data are compicated and that you are looking for patterns, 
you may also want to recode some of the variables into new categorical 
variables, either to search for ways to subset the data meaningfully, or 
to give yourself access to other tools such as glm or nlme techniques. 

But please do not veer off into "machine learning" packages. Do the best you 
can with regression techniques.