Disallowed Combinations

[StevenCHowell]

Does skpr support disallowed combinations?

[tylermorganwall]

Yes--simply remove them from the candidate set. skpr only uses the test points that you provide when searching for a design, so disallowed combinations are trivial to add.

[StevenCHowell]

That makes sense.

For the next guy (or future-me), here is one way to remove the points (6, 7, 8 in this example):

candidateset <- candidateset[-c(6, 7, 8), ]

[tylermorganwall]

If you load the package dplyr, you can also filter them according via a inequality or boolean expression.

For example, here I filter the allowed points down to a circle around the origin:

library(dplyr)

candset = expand.grid(x=seq(-1,1,0.1),y=c(-1,1,0.1))

filteredcandset = filter(candset, x^2+y^2 <= 1)

[StevenCHowell]

I am finding that even when I remove options from the candidate set, they still end up in the design. Here is some example code demonstrating this

#Setting random number generator seed:
set.seed(1)

# Generating candidate set:
candidateset = expand.grid(f1 = c(-1,1),
                           f2 = c(1,2,3,4,5))

# Remove disallowed combinations:
candidateset <- subset(candidateset,
                        !(f2 %in% c(5) & f1 %in% c(-1)) & 
                        !(f2 %in% c(1, 2, 3) & f1 %in% c(1)))

# Generating design for hard-to-change factors: 
design_htc = gen_design(candidateset = candidateset, 
                        model = ~f1,
                        trials = 4)

# Generating design:
design = gen_design(candidateset = candidateset, 
                    model = ~.,
                    trials = 12,
                    splitplotdesign = design_htc,
                    splitplotsizes = 3,
                    optimality = "Alias",
                    splitcolumns = TRUE)

with the resulting candidate set being

	f1	f2
1	-1	1
3	-1	2
5	-1	3
7	-1	4
8	1	4
10	1	5

and the resulting design

	Block1	f1	f2
1.1	1	-1	5
1.2	1	-1	1
1.3	1	-1	5
2.1	2	-1	1
2.2	2	-1	5
2.3	2	-1	1
3.1	3	1	4
3.2	3	1	1
3.3	3	1	4
4.1	4	1	4
4.2	4	1	1
4.3	4	1	1

Note that I the (f1, f2) combinations (-1, 5) and (1, 1) were both removed from the candidate set but still end up in the design.

Here is the output from sessionInfo() showing the R and skpr versions:

R version 3.4.3 (2017-11-30)
Platform: x86_64-w64-mingw32/x64 (64-bit)
Running under: Windows 10 x64 (build 14393)

Matrix products: default

locale:
[1] LC_COLLATE=English_United States.1252 
[2] LC_CTYPE=English_United States.1252   
[3] LC_MONETARY=English_United States.1252
[4] LC_NUMERIC=C                          
[5] LC_TIME=English_United States.1252    

attached base packages:
[1] stats     graphics  grDevices utils     datasets  methods   base     

other attached packages:
[1] skpr_0.49.1          shiny_1.0.5          RevoUtils_10.0.8    
[4] RevoUtilsMath_10.0.1

loaded via a namespace (and not attached):
 [1] pbdZMQ_0.2-6      repr_0.12.0       splines_3.4.3     lattice_0.20-35  
 [5] colorspace_1.3-2  htmltools_0.3.6   viridisLite_0.2.0 rintrojs_0.2.0   
 [9] survival_2.41-3   rlang_0.1.6       pillar_1.0.1      withr_2.1.1      
[13] registry_0.5      rngtools_1.3.1    doRNG_1.6.6       uuid_0.1-2       
[17] foreach_1.4.5     plyr_1.8.4        pkgmaker_0.27     stringr_1.2.0    
[21] munsell_0.4.3     rvest_0.3.2       codetools_0.2-15  kableExtra_0.9.0 
[25] evaluate_0.10.1   knitr_1.18        doParallel_1.0.12 httpuv_1.3.5     
[29] parallel_3.4.3    IRdisplay_0.4.4   Rcpp_0.12.14      xtable_1.8-2     
[33] readr_1.1.1       backports_1.1.2   scales_0.5.0      IRkernel_0.8.11  
[37] jsonlite_1.5      mime_0.5          hms_0.4.0         digest_0.6.13    
[41] stringi_1.1.6     grid_3.4.3        rprojroot_1.3-1   bibtex_0.4.2     
[45] tools_3.4.3       magrittr_1.5      shinythemes_1.1.1 tibble_1.4.1     
[49] crayon_1.3.4      pkgconfig_2.0.1   Matrix_1.2-12     xml2_1.1.1       
[53] rmarkdown_1.8     httr_1.3.1        rstudioapi_0.7    iterators_1.0.9  
[57] R6_2.2.2          nlme_3.1-131      compiler_3.4.3   

[tylermorganwall]

It looks like this is a bug--it shouldn't be hard to fix.

[StevenCHowell]

I expect that if f1 was easy-to-change this might not happen. I see how the hard-to-change values are assigned to the first level of randomization. When the selects points for the second level, it seems to not check the candidate set.

[StevenCHowell]

If I instead do

#Setting random number generator seed:
set.seed(1)

# Generating candidate set:
candidateset = expand.grid(f1 = c(-1,1),
                           f2 = c(1,2,3,4,5))

# Remove disallowed combinations:
candidateset <- subset(candidateset,
                        !(f2 %in% c(5) & f1 %in% c(-1)) & 
                        !(f2 %in% c(1, 2, 3) & f1 %in% c(1)))

# Generating design:
design = gen_design(candidateset = candidateset, 
                    model = ~.,
                    trials = 12,
                    optimality = "Alias",
                    splitcolumns = TRUE)

the resulting design is

f1	f2
-1	1
1	4
-1	4
1	4
-1	1
1	5
1	5
-1	4
-1	4
-1	1
-1	4
-1	1

which looks like it is working. This supports the idea that the issue is related to introducing a split-plot factor with disallowed combinations.

[tylermorganwall]

It was a simple normalization error--this is fixed in the version I just updated, v0.53.1.

[StevenCHowell]

@tylermorganwall, how often do you push a new version to cran? I do not have the ability to compile from source on the machine I am using this on, and it looks like the cran version is still v0.49.1.

[tylermorganwall]

I can only upload a new version every thirty days or so--any more frequently and the CRAN complains. I'll see if I can upload the most current version today.

[tylermorganwall]

The latest version is on CRAN--it usually takes the CRAN a few days to build all of the binaries, though.