numpy.linalg.linalg.LinAlgError on a bbob-biobj problem about pycma

[brockho]

I now rerun the experiment which showed a crash on another problem instance. Most importantly, I could record the random seed for reproducing the error (the error message is the same as above and thus cut):

bbob-biobj_f48_i04_d40 2-objective problem (bbob-biobj(bbob_f015_i09_d40__bbob_f020_i10_d40)(7608))
  --> new weight: 1.020408e-01
(7_w,15)-aCMA-ES (mu_w=4.5,w_1=34%) in dimension 40 (seed=361690, Sat May 20 02:19:30 2017)
(7_w,15)-aCMA-ES (mu_w=4.5,w_1=34%) in dimension 40 (seed=402495, Sat May 20 02:19:35 2017)
Traceback (most recent call last):
  File "example_experiment-WS-normalized.py", line 488, in <module>
    main(budget, max_runs, current_batch, number_of_batches)
[...]
  File "/usr/lib/python2.7/dist-packages/numpy/linalg/linalg.py", line 1187, in eigh
    raise LinAlgError, 'Eigenvalues did not converge'
numpy.linalg.linalg.LinAlgError: Eigenvalues did not converge

[nikohansen]

Excellent, now to reproduce the error it would be useful to also have the single-objective function definition and the precise arguments with which cma.fmin was called.

[brockho]

After quite some trials to find again the right random seed that reproduces the error (I had to run about 50 runs with different random seeds to find one that actually gives an error). The attached zip file below gives the corresponding example experiment that failed.

example_experiment-WS-normalized-reconstructerror.zip

Note that I realized later on that the error is not reproducible on all our compute servers! Here are some details on the machines where it worked and where it failed:

working on porte-d-auteuil with
python 2.7.3, numpy 1.6.1, GCC 4.6.3 (linux 2)

not working on porte-de-la-chapelle with
python 2.6.6, numpy 1.10.4, GCC 4.4.7 20120313 (Red Hat 4.4.7-3)

In both cases, I use the same CMA-ES version, imported from a zip file, obtained from the pycma github page:

>>> import cma
>>> cma.__version__
'2.1.0  $Revision: 4394 $ $Date: 2017-05-13 15:04:41 +0200 (Sat, 13 May 2017) $'

Hope this helps. @nikohansen: if you need more details about the machines, please let me know. I can also tell you offline where the code is located to actually reproduce the error yourself.

[nikohansen]

@brockho I unzipped the above file, but I am not sure whether I tested the right thing, I namely called the script with the parameter bbob-biobj 1e5, which doesn't reproduce the error.

[nikohansen]

The above referenced commit, de30cca, should convert the error into a warning. I am undecided though whether this is a good or a bad thing, as this may obscure the observation that something weird is going on.

[brockho]

Did you try on your machine or on porte-de-la-chapelle in the lab? On porte-d-auteuil, for example, I could also not reproduce the error.