predict.cubist unable to predict properly using sample (cubistControl)

[Laurae2]

I am trying to predict a linear relation in Cubist between three variables using Cubist (from a very large data set). However, it seems to always break down when trying to predict, when "it does not break down" when predicting manually using the output formulas.

Here are the vectors used:

v1 <- c(66787, 47194, 39871, 44990, 103933, 57240, 70013, 113002, 31145, 
47194, 64492, 36441, 197228, 202286, 14601, 25862, 120784, 84379, 
67224, 57301, 142191, 59581, 160405, 45648, 56873, 111957, 74430, 
84701, 0, 72055, 44301, 124789, 128377, 65629, 125613, 54780, 
78418, 36186, 59571, 148794, 17387, 79497, 47886, 160173, 100197, 
67793, 101231, 32230, 69549, 140863)

v2 <- c(187113, 244099, 142255, 116351, 179189, 127059, 174851, 233094, 
132003, 187143, 160828, 201573, 193093, 214011, 188806, 252668, 
173534, 355734, 160811, 215225, 204655, 221497, 175405, 126996, 
315174, 242112, 167534, 156679, 305221, 252339, 202403, 280700, 
206511, 257729, 184985, 291769, 108440, 259298, 252483, 213778, 
251058, 179890, 182320, 223046, 225751, 253243, 185440, 187539, 
169371, 254666)

v3 <- c(173569, 235079, 134319, 107179, 157561, 115288, 160548, 209768, 
128156, 177842, 147652, 193357, 151329, 171255, 186572, 248310, 
148230, 340352, 147043, 203874, 174868, 209683, 141529, 117734, 
304407, 219056, 152238, 139106, 306722, 237975, 191885, 256744, 
179724, 244784, 158690, 281340, 91991, 254696, 240824, 182605, 
248527, 164788, 172846, 189454, 205161, 239806, 164428, 181482, 
155141, 227106)

Expected answer: v1 = 3729.1 - 4.526_v3 + 4.54_v2

vdf <- data.frame(v2, v3)
set.seed(11111)
commiteeControl <- cubistControl(sample = 50, rules = 1)
commiteeModel <- cubist(x = vdf, y = v1, control = commiteeControl)
print(commiteeModel)
summary(commiteeModel)
predictions <- predict(commiteeModel, newdata = vdf)
sqrt(mean((predictions - v1)^2)) #RMSE
cor(predictions, v1)^2 #R^2

Output of print/summary:

> print(commiteeModel)

Call:
cubist.default(x = vdf, y = v1, control = commiteeControl)

Number of samples: 50 
Number of predictors: 2 

Number of committees: 1 
Number of rules: 1 
Other options: 50% sub-sampling
> summary(commiteeModel)

Call:
cubist.default(x = vdf, y = v1, control = commiteeControl)


Cubist [Release 2.07 GPL Edition]  Sat Apr 02 14:13:15 2016
---------------------------------

    Target attribute `outcome'

Read 25 cases (3 attributes) from undefined.data

Model:

  Rule 1: [25 cases, mean 88864.3, range 0 to 197228, est err 2566.0]

    outcome = 3729.1 - 4.526 v3 + 4.54 v2


Evaluation on training data (25 cases):

    Average  |error|             1965.2
    Relative |error|               0.05
    Correlation coefficient        1.00


    Attribute usage:
      Conds  Model

             100%    v2
             100%    v3


Evaluation on test data (25 cases):

    Average  |error|             3513.7
    Relative |error|               0.09
    Correlation coefficient        0.99


Time: 0.0 secs

Everything looks perfect... until:

> predictions <- predict(commiteeModel, newdata = vdf)
> sqrt(mean((predictions - v1)^2)) #RMSE
[1] 77990.2
> cor(predictions, v1)^2 #R^2
[1] 1.221792e-06

And when I compare to the following, it does not make sense as it is the right expected answer:

> predictions <- 3729.1 - 4.526*v3 + 4.54*v2
> sqrt(mean((predictions - v1)^2)) #RMSE
[1] 3297.196
> cor(predictions, v1)^2 #R^2
[1] 0.9952073

Tested under:
R version 3.2.4 Revised
R version 3.2.3
Rgui and RStudio under Windows 7, Windows 8.1, Windows 10
Virtual machine and non-virtual machine
Different computers

Removing from $model "sample="0.5" init="3965" cleared the issue in my case.

I put this in $Model of the model:

commiteeModel$model <- "id=\"Cubist 2.07 GPL Edition 2016-04-02\"\nprec=\"0\" globalmean=\"76794.88\" extrap=\"1\" insts=\"0\" ceiling=\"404572\" floor=\"0\"\natt=\"outcome\" mean=\"76794.8\" sd=\"49683.92\" min=\"0\" max=\"202286\"\natt=\"v2\" mean=\"209191\" sd=\"42482.83\" min=\"108440\" max=\"305221\"\natt=\"v3\" mean=\"193537.4\" sd=\"47135.79\" min=\"91991\" max=\"306722\"\nentries=\"1\"\nrules=\"1\"\nconds=\"0\" cover=\"25\" mean=\"76794.9\" loval=\"0\" hival=\"202286\" esterr=\"1456.4\"\ncoeff=\"144.4\" att=\"v2\" coeff=\"4.66\" att=\"v3\" coeff=\"-4.64\"\n"

instead of:

commiteeModel$model <- "id=\"Cubist 2.07 GPL Edition 2016-04-02\"\nprec=\"0\" globalmean=\"76794.88\" extrap=\"1\" insts=\"0\" ceiling=\"404572\" floor=\"0\"\natt=\"outcome\" mean=\"76794.8\" sd=\"49683.92\" min=\"0\" max=\"202286\"\natt=\"v2\" mean=\"209191\" sd=\"42482.83\" min=\"108440\" max=\"305221\"\natt=\"v3\" mean=\"193537.4\" sd=\"47135.79\" min=\"91991\" max=\"306722\"\nsample=\"0.5\" init=\"3965\"\nentries=\"1\"\nrules=\"1\"\nconds=\"0\" cover=\"25\" mean=\"76794.9\" loval=\"0\" hival=\"202286\" esterr=\"1456.4\"\ncoeff=\"144.4\" att=\"v2\" coeff=\"4.66\" att=\"v3\" coeff=\"-4.64\"\n"

And the prediction worked perfectly.