Julien – c’est sur un serveur local (debian/squeeze wims-4.00-4 + patch pour
bug #8408), dans un répertoire pour « prof visiteur » [je trouve ça toujours
le plus simple à gérer], c.a.d.
log/classes/1234/
qui contient
ls -F
anstype/ content.html doc/ Exindex exostat livret/ score/ sheets/ supervisor vote/
cdt/ def/ exams/ Extitles forum/ noscore/ scoreavg src/ versionWims et apache ont le droit de lecture de anstype/, anstype/algexpmynewname et
anstype/algexpmynewname.input (et d’execution de anstype/).
Un problème de droits d’accès me semble envisageable, mais je ne vois
pas où exactement. Faut-il le droit d’écrire des fichiers temporaires qq part ?
Ou peut-être c’est une erreur de chemin ?
Joke – thanks for helping me read the source – in that case I think
I’ll use type=function explicitly. The evaluation method for functions
does not constitute a serious mathematical proof that two expressions
are equal, but for practical purposes in « ordinary » physics (no Dirac
delta functions, Cantor functions, etc.) it may well be enough.
In that case – this is just wondering aloud – what are the chances of
either incorrectly marking a wrong answer to be correct or a right answer
to be wrong?
Continuing with my question 1 bis, I can get a « wrong » answer (one or more
terms of Taylor expansion at t=0) to be marked nearly correct or fully
correct if the precision is not too strict and if the range is close
to zero:
\precision{100000}
\range{0.001..0.1}…
\answer{1. La reponse ?}{\theanswer}{type=function}
2/t bad precision, the correct answer is cos(t/8)/(4*sin(t/8)).
You have got a score of 7.2 on 10.
2/t - t/96 good answer.I didn’t manage to create an example where the right answer is marked
wrong, without any warning of an error in the script. In principle,
requiring a precision of 1e-16 for double precision calculations
should give different answers when calculating something in one step
versus many steps with uncorrelated least significant bit errors.
I had never thought of this sort of numerical checking to be anything
more than a hack before, but the chance of it failing either way are
probably low for « reasonable » practical cases, as long as the author(s)
of the exercise do not forget the nature of the algorithm. Good :).