[Debian] t_MetaModelAlgorithm_std.py selects the wrong distribution for Uniform[-pi, pi] on Arm v8, IBM power 8 and IBM s390x

[pgrt]

Hello,

I am in the progress of packaging version 1.20 for Debian. I met an issue with the t_MetaModelAlgorithm_std.py Python test on arm64, ppc64el and s390x: when testing
ot.ComposedDistribution([ot.Uniform(-pi, pi)] * dim)

the distribution
ComposedDistribution(TruncatedNormal(mu = 1.94e+05, sigma = 4.74e+03, a = -3.14, b = 3.13), Uniform(a = -3.14, b = 3.14), IndependentCopula(dimension = 2))
is selected instead of the targeted one: I get the output

---

543/581 Test #1047: pyinstallcheck_MetaModelAlgorithm_std ..................................***Failed 20.09 sec
E20221122 17:23:24.506155 2236764 trust_region_minimizer.cc:95] Terminating: Number of consecutive invalid steps more than Solver::Options::max_num_consecutive_invalid_steps: 5
E20221122 17:23:25.012102 2236764 trust_region_minimizer.cc:95] Terminating: Number of consecutive invalid steps more than Solver::Options::max_num_consecutive_invalid_steps: 5
E20221122 17:23:33.809854 2236764 trust_region_minimizer.cc:95] Terminating: Number of consecutive invalid steps more than Solver::Options::max_num_consecutive_invalid_steps: 5
--- /<>/python/test/t_MetaModelAlgorithm_std.expout 2022-11-08 10:49:19.000000000 +0000
+++ /<>/builddir/python/test/t_MetaModelAlgorithm_std.out 2022-11-22 17:23:33.814440566 +0000
@@ -1,4 +1,4 @@
-ComposedDistribution(Uniform(a = -3.14, b = 3.14), Uniform(a = -3.14, b = 3.14), IndependentCopula(dimension = 2))
+ComposedDistribution(TruncatedNormal(mu = 1.94e+05, sigma = 4.74e+03, a = -3.14, b = 3.13), Uniform(a = -3.14, b = 3.14), IndependentCopula(dimension = 2))
ComposedDistribution(Normal(mu = 3.98, sigma = 2.02), Normal(mu = 3.92, sigma = 2.02), IndependentCopula(dimension = 2))
ComposedDistribution(Exponential(lambda = 0.967, gamma = 0.00106), FisherSnedecor(d1 = 1.99, d2 = 157), IndependentCopula(dimension = 2))
ComposedDistribution(WeibullMin(beta = 0.626, alpha = 1.25, gamma = -0.496), Gamma(k = 1.41, lambda = 2.34, gamma = -0.497), IndependentCopula(dimension = 2))

---

I looked at the verbose output (ot.Log.Show(1)) for this targeted distribution, enclosed. Lines 30 and 31 show that the p-values of the Kolmogorov tests are greater than 0.05 only for the TruncatedNormal and Uniform distributions. I did not dig to find the reasons the TruncatedNormal gets selected (higher p-value? Alphabetical order? ...?) but this is not the desired output anyway.
outputMetaModelAlgorithm_uniform.txt

On other architectures I don't meet this issue.

Cheers,
Pierre

P.S.: thanks for the fixes of the other open issues you addressed today ! :) I had other new issues when packaging 1.20 but I trust they will be solved by your solutions to #2046 and #2047 as they are the same kind of problems.

[jschueller]

I dont have this one, which version of debian is this ? which version of ceres ?

[pgrt]

Hi,

I got it using Debian experimental (but I have the same ones on Debian unstable), with ceres 2.1.0.

Today I ran the test again, and surprisingly the estimated parameters of TruncatedNormal are different, as I got

---

DBG - class=TestResult name=Unnamed type=Kolmogorov TruncatedNormal binaryQualityMeasure=true p-value threshold=0.001 p-value=0.853 statistic=0.0135 description=[TruncatedNormal(mu = 6.68, sigma = 16.7, a = -3.14, b = 3.14) vs sample Unnamed]
DBG - class=TestResult name=Unnamed type=Kolmogorov Uniform binaryQualityMeasure=true p-value threshold=0.001 p-value=0.197 statistic=0.024 description=[Uniform(a = -3.14, b = 3.14) vs sample Unnamed]

---

on ppc64el (IBM Power8) but also on amd64 (exactly the same estimated parameters, the same version of ceres, and also on Debian experimental).
Still, the distribution which is selected is

---

ComposedDistribution(TruncatedNormal(mu = 1.94e+05, sigma = 4.74e+03, a = -3.14, b = 3.13), Uniform(a = -3.14, b = 3.14), IndependentCopula(dimension = 2))

---

on ppc64el and

---

ComposedDistribution(Uniform(a = -3.14, b = 3.14), Uniform(a = -3.14, b = 3.14), IndependentCopula(dimension = 2))

---

on amd64 although the estimated distributions and p-values of the Kolmogorov tests on both architectures are exactly the same: only the selection of the winning distribution seems to differ.

Are you able to get a p-value higher than 0.05 for TruncatedNormal on your side?

Thanks a lot,

Pierre

[jschueller]

same as #2046