[Tutor] unittest with random population data

[Sydney Shall]

On 31/05/2015 03:00, Cameron Simpson wrote:
> On 30May2015 12:16, Sydney Shall <s.shall at virginmedia.com> wrote:
>> Following advice from you generous people, I have chosen a project
>> >that interests me, to develop some knowledge of python.
>> My projest is a simulation of a biological population.
>> I have a base class and a simulation function, which uses instances of
>> the class.
>> This, after many months of work and lots of advice, now seems to work
>> well. It generates sensible data and when I write a small test program
>> it gives sensible output.
>> Now I want to learn to use unittest.
>> I have written a unittest class which works OK.
>> But the problem I have is that because I use the random module to
>> populate my initial arrays, my data is not strictly predictable even
>> though I am using seed(0). So the tests return many *fails* because
>> the numbers are not exactly correct, although they are all rather
>> close, consistent with the sigma value I have chosen for the spread of
>> my population. I do of course use *almostEqual* and not *Equal*.
>
> First of all, several people have posted suggestions for getting
> identical results on every run.
>
> However, there is another approach, which you might consider. (And use
> in addition, not inseadt of, the reproducable suggestions).
>
> It is all very well to have a unit test that runs exactly the same with
> a test set of data - it lets you have confidence that algorithm changes
> do not change the outcome. But on for that data set.
>
> You say that your results are "all rather close, consistent with the sigma
> value I have chosen for the spread of my population". I would advocate
> making some "contraint" tests that verify this property for _any_ input
> data set.
>
> Then you can run with random and _changing_ input data sets to verify
> that your code produces the expected _kind_ of results with many data sets.
>
> So you would have one test which ran with a fixed data set which
> confirms preidctable unchanging results. And you have other tests with
> run with randomly chosen data and confirms that outcomes fall within the
> parameters you expect. You can apply those checks ("outcome in range")
> to both sets of tests.
>
> As an exmaple, I have a few classes which maintain data structures which
> are sensitive to boundary conditions. The glaring example is a numeric
> range class which stores contiguous ranges efficiently (a sequence of
> (low,high) pairs). It has a few add/remove operations which are meant to
> maintain that sequence on ordered minimal form. cutting and merging
> adjacent ranges is very easy to get wrong, very sensitive to off-by-one
> logic errors.
>
> So my tests for this class include some random tests which do random
> unpredictable add/remove operations, and run a consistency check on the
> object after each operation. This gives me good odds of exercising some
> tricky sequence which I have not considered explicitly myself.
>
> You can see the test suite here:
>
>   https://bitbucket.org/cameron_simpson/css/src/tip/lib/python/cs/range_tests.py
>
> It has a bunch of explicit specific tests up the top, and then the
> random consistency test down the bottom as "test30random_set_equivalence".
>
> Cheers,
> Cameron Simpson <cs at zip.com.au>
>
> MS-Word is Not a document exchange format - Jeff Goldberg
> http://www.goldmark.org/netrants/no-word/attach.html
>
Cameron,
Thanks for your most helpful reply.
I have studied the material you indicated and it has been most helpful. 
I think that I have understood the principle involved, but I have had 
some problem implementing it.

The range tests are mostly clear to me but there is one aspect I cannot 
follow.
You use in this suite imports from Range, including Range, overlap, 
spans and Span.
Are these programs that you have written? If so, are they specific to 
your set up or are they generic? If so, is it possible to obtain these 
programs?
I have established a very primitive test suite based on your cs.range 
notions and it works fine, but it would be better, I am sure, to do it 
properly.

Finally, I have one comment for the respected Moderator, if he is not 
out on a walk in the highlands in this cold  and wet weather.

I have taken the liberty of raising my problem here rather than 
elsewhere, because I have observed that the biological and bio-medical 
community, who always come late to new notions, is now rapidly 
discovering  python. A great deal of work in these fields involve either 
stochastic simulations or statistical problems of analysis. The latter 
are more or less straight-forward, but the simulations are not.

Thanks for all the help. You people are a model of how we could perhaps 
civilize humanity.



-- 
Sydney