Session crashes / low-level error from irls_poisson_fast.cpp if no solution found by solve()

[matteodelucchi]

No description provided.

[matteodelucchi]

Given two poisson variables:

  mydists <- list(a="poisson",
                  b="poisson")
  a <- rpois(1000, lambda = 1)
  z <- exp(1+2*a)
  b <- rpois(1000, lambda = z)
  mydf <- data.frame("a" = a,
                     "b" = as.integer(b))

We can estimate the following model:

  glm(b~a, data = mydf, family = "poison")

Problem 1

A similar behaviour as from glm() should be expected from buildScoreCache

>   mycache.mle <- buildScoreCache(data.df = mydf,
+                                  data.dists = mydists,
+                                  method = "mle",
+                                  max.parents = 1,
+                                  verbose = TRUE)
evaluation # 1:
$row.num
[1] 1

result of evaluating expression:
[1] -1302.247  2606.494  2611.402  2612.095
got results for task 1
numValues: 1, numResults: 1, stopped: FALSE
returning status FALSE
evaluation # 2:
$row.num
[1] 2

result of evaluating expression:
[1] -1263.188  2530.377  2540.193  2541.579
got results for task 2
numValues: 2, numResults: 2, stopped: FALSE
returning status FALSE
evaluation # 3:
$row.num
[1] 3

result of evaluating expression:
[1] -Inf  Inf  Inf  Inf
got results for task 3
numValues: 3, numResults: 3, stopped: FALSE
returning status FALSE
evaluation # 4:
$row.num
[1] 4

result of evaluating expression:
<std::runtime_error: solve(): solution not found>
got results for task 4
accumulate got an error result
numValues: 4, numResults: 4, stopped: FALSE
returning status FALSE
numValues: 4, numResults: 4, stopped: TRUE
first call to combine function
evaluating call object to combine results:
  fun(result.1, result.2, result.3)
Error in { : task 4 failed - "solve(): solution not found"
In addition: Warning messages:
1: In check.valid.groups(group.var = group.var, data.df = data.df,  :
  No cor.vars specified. Using all but group.var instead.
2: In check.valid.groups(group.var = group.var, data.df = data.df,  :
  No cor.vars specified. Using all but group.var instead.
3: In check.valid.buildControls(control = control, method = method,  :
  Control parameters provided that are not used with method mle are ignored.

The error seems to come from irls_poisson_fast.cpp() which can not find solutions for all possible combinations.

• Why does glm() work but buildScoreCache doesn't?
• 1st priority: Have similar behaviour as glm.
• 2nd priority: Can we at least return a more informative error message?

Problem 2

If a is wrongly provided as

a <- runif(1000, min=1, max=2.5)

but is still specified in data.dists=list(a="poisson"...); this is not catched anywhere and results in a silent (nothing written to syslog, dmesg, debug, ... and also no special behaviour detected in htop) crash of the R session.

• Why does R crash silently?
• Let R not crash!!!