This is the title of a blog post by Jake Vanderplas, researcher in Astronomy & Machine Learning at University of Washington. He points out that conducting successful research requires more and more data manipulation skills, going along programming skills. However, in academia, ability to write good software is not promoted, if not discouraged !
"academia has been singularly successful at discouraging these very practices that would contribute to its success"
"any time spent building and documenting software tools is time spent not writing research papers, which are the primary currency of the academic reward structure"
On the other hand, software skills are very important and thus well rewarded in the industry, thus the idea of "Big Data Brain Drain" which pumps talented young graduates out of academic research.
After the diagnosis
Jake's post is the "medical diagnosis", and each disease calls for a treatment ! Since the problem is sociological/organizational, the treatment must be sociological/organizational. Jake lays 4 propositions, in particular the evolution of research evaluation criteria. Of course the "implementation details" of evaluation are always a tough issue, not only for research (thinking of learning and teaching evaluation here).
But in general, I hope that the recognition of good software will change positively, along with the general issue of reproducibility. In fact, I think that many academics are aware of the issue, but they just don't see the practical track to recover from the current "dead end" (and also senior researcher don't have much time to thoroughly work on the issue) :
"Making an openly available program for electrical machine sizing would be immensely useful for our research community! It would summarize 20 years of research of our group. I just don't take/find the time for it."
This is an (approximate & very shortened) transcript of the reaction of Hamid Ben Ahmed, one of my PhD advisor when discussing the topic this week. This means that in the field Electrical Engineering (which is has been tied for decades with closed source softwares like Simulink or 3D finite elements models) the feeling that "something is not working" is already there, and that's a good start!
Pushing the change
Now, it is all about academics pushing "le Système" (i.e. French academia), and not waiting for the change to come "from the top". Indeed, I feel that top-level research directors have too many other things to deal with, like managing huge research consortium, writing huge evaluation reports, ... no time for "far away issues" such as reproducibility 😉
Let's just push !
PS : not all of the electrical engineering research runs on closed software. See for example the open source work of Prof. Bernard Uguen and his team on radio wave propagation : www.pylayers.org (from IETR, a neighbor lab of Rennes University 1)
I found also two blog posts by Matthew Brett on the NiPy blog ("NeuroImaging with Python" community) about Unscientific Programming and the Ubiquity of Error in computing. The latter asserts that computing tools in science lead easily to results with many mistakes (not to mention the recent discussion about Excel spreadsheets mistakes). From my experience in computing for science, I very much agree with this fact. Often those mistakes are small though (i.e. the order of magnitude of the result is preserved), but not always...
From my research...
Matthew Brett blog reminded of a pretty bad example of error in computing that I encounter when working on my last conference paper dealing with the modeling of a sodium-sulfur battery (cf. PowerTech article on my publications page).
Just in time for the deadline, I submitted a "long abstract" in October 2012. I had just finished implementing the model of the battery and the simulation was up and running. One of the main figure presenting the results is copied here :
and now the interesting thing is to compare the October 2012 version to the February 2013 version (submission of the full paper). Beyond surface changes in the annotations, I've highlighted two big differences in the results. The most striking change is in the lower pane : the "cycle counting" drops from about 400 down to 25 cycles/month. One order of magnitude less !
What happened in the meantime...
Without entering the details, there were several really tiny errors in the implementation of the model. I would not even call these erros "bugs", those were just tiny mistakes. One was somehow like writing Q += i instead of Q += i*dt (omitting the time step when counting electric charges). And when the time step dt is 0.1, that's makes an easy method to miss exactly one order of magnitude ! Spotting those errors in fact takes quite some time and probably one or two weeks were devoted to debugging in November (just after the submission of the abstract).
Of course, reviewers have almost no way to spot this error. First the code is not accessible (the battery model is confidential) and second, the value that was wrong (cycle counting) cannot be easily checked with a qualitative reasoning.
Since I don't know how to solve the problem from the reviewer point of view, let's get back to my position : the man who writes the wrong code. And let's try to see how to make this code a little better.
Testing a physical simulation code
There is one thing which I feel makes model simulation code different from some other codes: it is the high number of tests required to check that it works well. Even with only 3 state variables and one input variable like the sodium-sulfur battery, there are quite a lot of combinations.
I ended up asking the code to report the evolution the mode on one single time step in this ASCII art fashion :
------ State at t0 ------
SoE| DoD_c| T_mod| N_cyc
0.500| 282.5| 305.0| 0.0
: ------------ Power flows [W] --------- | - State evolution -
and to cover enough situations, I have in fact 6 text files containing each 5 blocks like this one. That's 30 tables to check manually, so that's still doable, but there is no easy way to tell "oh, that -15,107 over there doesn't looks right"... it just take time and a critical eye (the latter is the most difficult to get).
Freeze the test for the future
If I now want to upgrade the battery simulation code, I want to ensure that the time spent by my thesis advisor and I on the result checking is not lost.I want to enforce that the code always generates the exact same numerical result.
This is the (somehow dirty) method I've used : run the same test, generate the ASCII string, and compare the result with a previous run stored in a text file. I pasted the code here so that the word "dirty" gets an appropriate meaning :
function that generates an ASCII test report and compare with a previous test result