--- title: "ADAS Utils vignette" output: rmarkdown::html_vignette vignette: > %\VignetteIndexEntry{ADAS Utils vignette} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r setup, include = FALSE} knitr::opts_chunk$set( collapse = TRUE, comment = "#>", fig.dim=c(8, 4), out.width="100%" ) library(adas.utils) library(tidyverse) ``` # Factorial plans The package provides tools for dealing with factorial plan according to the *Design of Experiments* (DoE) protocols. The functions for dealing with DoE have names starting with `fp_`. As much as possible, we are aiming at a tidyverse-like syntax, so that the functions can be used in a pipe. We are following conventions and techniques illustrated in the book *Design and Analysis of Experiments* by Douglas C. Montgomery. ## Full factorial plan You can create a full factorial plan with the `fp_design_matrix` function, passing the number of factors: ```{r} (dm <- fp_design_matrix(2, rep=2) %>% mutate(Y=rnorm(n()))) ``` In this case, the factors are the first `n` capital letters. If you want different factor names, use a right-side-only formula combining all the named factors with `*`: ```{r} fp_design_matrix(~Speed*Weight) ``` **NOTE**, though, that using custom factor names is discouraged, and won't work as expected if you are using the functions for dealing with **fractional factorial plans**, especially for the analysis of alias structures among factors. The yield column `Y` must then be completed according to the randomized `RunOrder` column. It is possible to add custom scales to the factors, and also add names to the factors: ```{r} fp_design_matrix(2) %>% fp_add_names(A="Temperature", B="Pressure") %>% fp_add_scale(A=c(20, 25), B=c(75, 125), suffix=".scaled") ``` ## Custom levels If you want a $k^n$ factorial plan with custom levels, pass the `levels` argument. In this case, though, the `.treat` column with Yates' treatment codes would be `NA`: ```{r} fp_design_matrix(2, levels=-1:1) ``` ## Augment a plan You can augment a plan by adding a central point, typically repeated: ```{r} fp_design_matrix(3) %>% fp_augment_center(rep=4) ``` Then if needed (because the analysis show low p-value for the quadratic term) you can add axial points to get a central composite design: ```{r} fp_design_matrix(3) %>% fp_augment_center(rep=3) %>% fp_augment_axial(rep=2) ``` ### Full example Let's see a full example using the `ccd_experiment_yield` dataset, which contains a list of the yield data for three sequential experiments in a central composite design. First, we design a $3\cdot 2^2$ factorial plan, with two factors and two levels each: ```{r} fp <- fp_design_matrix(2, rep=3) ``` Ideally, we would then sort the table according to the `RunOrder` column, and complete the `Y` column with the yield data from the the real experiments. For the sake of documenting the package, we can directly add the yield data from the `base` field of the `ccd_experiment_yield` dataset (which holds values **in standard Yates' order**): ```{r} fp$Y <- ccd_experiment_yield$base ``` Now we can fit a linear model to the data, and check the p-values of the ANOVA: ```{r} fp %>% lm(Y ~ A*B, data=.) %>% anova() ``` All factors and their interactions are significant. But is the two-level model enough? Let's check for the quadratic terms, by augmenting the plan with a central point repeated 4 times. We also load the `center` field from the `ccd_experiment_yield` dataset: ```{r} fpc <- fp %>% fp_augment_center(rep=4) fpc$Y[fpc$.treat == "center"] <- ccd_experiment_yield$center ``` Now we can fit a model with the quadratic term, using either $A$ or $B$: since we only have a central point, we cannot discriminate which factor is contributing to the curvature in the response surface. We get: ```{r} fpc %>% lm(Y ~ A*B+I(A^2), data=.) %>% anova() ``` So the contribution of the quadratic term **is significant**. This means that we have to further augment the plan with axial points and investigate a *Central Composite Design* (CCD). **Note**: if the quadratic term contribution were not significant, we would have to remove the quadratic term from the model and accept the two-level model. So let's load the axial points from the `axial` field of the `ccd_experiment_yield` dataset, and fit a model with the quadratic terms and their interactions: ```{r} fpccd <- fpc %>% fp_augment_axial(rep=2) fpccd$Y[fpccd$.treat == "axial"] <- ccd_experiment_yield$axial fpccd %>% lm(Y ~ A*B*I(A^2)*I(B^2), data=.) %>% anova() ``` So we can finally state that a proper model would be `Y ~ A*B+I(A^2)`: ```{r} fpccd %>% lm(Y ~ A*B+I(A^2), data=.) %>% summary() ``` ## Save to/load from a file Once the design matrix is prepared, you typically want to save it to a file and use it for collecting data form experiments. You can use the `write.csv` function, but it is recommended to use the `fp_write_csv` function, which will also save the design matrix properties as comments: ```{r eval=FALSE} dm <- fp_design_matrix(2) %>% fp_add_names(A="Temperature", B="Pressure") %>% fp_add_scale(A=c(2, 12), B=c(40, 60), suffix="_s") %>% fp_write_csv("design_matrix.csv") ``` Note that the `fp_write_csv` function invisibly returns the same design matrix, so you can use it in a pipe chain. Also, the CVS files has the rows arranged in the same order as the `RunOrder` column (i.e. randomized). Once the CSV file has been completed, you can load it back into R using the `fp_read_csv` function: ```{r eval=FALSE} dm <- dm %>% fp_read_csv("design_matrix.csv") ``` Note that `fp_read_csv` returns the design matrix reordered according to Yates' standard order. ## Fractional factorial plan It is possible to divide a design matrix into a fractional factorial plan using the `fp_fraction` function. The fraction uses a **defining relationship** (*dr*) as $I=ABCD$, which is mapped in R as a one side formula `~A*B*C*D`. Any fraction is added to the `factorial.plan` object in the `fraction` attribute. A full $2^n$ factorial plan can be reduced to a fractional factorial plan $2^{n-p}$ by applying the `fp_fraction` function $p$ times. For example, to get a $2^{5-2}$ plan with the defining relationships $I=ABCD$ and $I=BCDE$: ```{r} fp_design_matrix(5) %>% fp_fraction(~A*B*C*D) %>% fp_fraction(~B*C*D*E) ``` Note that with the `remove` option you can control if you want to keep both fractions, and later on `filter(ABC==1)` them out. ```{r} fp_design_matrix(3) %>% fp_fraction(~A*B*C, remove=FALSE) ``` Also, note that the `remove` option is sticky, so that when you can apply the `fp_fraction` function multiple times and the first time has the option set to `remove=FALSE`, then all the following `fp_fraction` calls will have the same option set to `FALSE`. Setting `remove=FALSE` to any of the following calls **can have unexpected behavior**. ## Alias structure Any fraction of a factorial plan results in a set of *aliases* among effects. The package provides the following functions to deal with alias structures: - `fp_alias_matrix`: returns a matrix with the alias structure of the factors in the design matrix. The alias matrix has a plot method. - `fp_all_drs`: given a set of defining relationships, returns the dependent one. - `fp_merge_drs`: given a set of defining relationships, returns the merged one, i.e. the one having all the factors. - `fp_gen2alias`: given a *generator* (i.e. the right side of a DR) and an effect name as strings, calculates the resulting alias. For example: ```{r} (am <- fp_alias_matrix(~A*B*C, ~B*C*D)) ``` ```{r} am %>% plot() ``` The design matrix can be converted to a tibble thanks to the proper `as_tibble.design.matrix` S3 method: ```{r} am %>% as_tibble() ``` # Statistics The package also provides some useful functions for statistical analysis of data. ## Plotting ### Normal probability plot The normal probability plot is provided as an alternative to the quantile-quantile plot: ```{r warning=FALSE} df <- tibble( xn = rnorm(100, mean=20, sd=5), xu = runif(100, min=0, max=40) ) df %>% normplot(xn) df %>% normplot(xu) ``` ### Pareto chart The Pareto chart is a bar chart that displays the relative importance of problems in a format that is very easy to interpret. The bars are sorted in descending order, and the cumulative percentage of the total is shown by the line. It can prove useful in the context of factorial plans, to identify the most important factors, or in sensitivity analysis, to identify the most important parameters. The package provides a *generic function*, `pareto_chart`, that can be used with a tibble (or a data frame), or with a linear model (an `lm` object). In the latter case, the function produces the Pareto chart of the model effects. For the general case, when you have a tibble with values and names: ```{r} set.seed(1) tibble( val=rnorm(10, sd=5), cat=LETTERS[1:length(val)] ) %>% pareto_chart(labels=cat, values=val) ``` For the case of a linear model: ```{r} filtration %>% lm(Y~A*B*C*D, data=.) %>% pareto_chart() ``` ### Daniel's plot In case of non-replicated factorial plans, the Daniel's plot can be used to identify the most important factors: a quantile-quantile plot of the factors effects shows the significant factors and interactions off the diagonal. ```{r} daniel_plot_qq(lm(Y~A*B*C*D, data=filtration)) ``` If you prefer, you can rather use a half-normal plot: ```{r} filtration %>% lm(Y~A*B*C*D, data=.) %>% daniel_plot_hn(nlab=6, repel=TRUE) ``` It shows that none of the effects containing the `B` factor are significant, so we can reduce the model to `Y~A*C*D`: ```{r} filtration %>% lm(Y~A*C*D, data=.) %>% anova() ``` Even better, the model can be further reduced to `Y~A*C+A*D`. Compare this conclusion with the last Pareto chart above. # Utilities This package also provides a function for easily loading data files made available on the accompanying course documentation on : ```{r} examples_url("battery.dat") %>% read.table(header=TRUE) ```