The problem is that technology moves so fast that vendors do not want to fix up their outdated offerings because they want to move forward and keep up with other providers. Furthermore, if the error can be traced back to a higher level source and more than one vendor utilizes this outdated source, the innocent may think all is OK because 2 independent sources give the same results, which can be wrong. Please remember it was not that long ago when computations were made by rooms full of equipment and people thought they were doing OK if they got reasonable results under restricted conditions using approximations with limited validity. My example is the modified Bessel function of integer order. It can be used to calculate the probability density function for the difference of 2 Poisson distributed random variables. If two Poisson distributed random variables each has expected value of 100.0, then the probability of observing j for the difference equals exp(-200) I sub(j)(200.0) With Isub(j) the modified Bessel function of integer order j. Substituting numbers and utilizing Excel 2013 for the value of the modified Bessel function, one finds that the probability of observing 3 is less likely than observing 4. This is obviously wrong. This error has existed since at least 2002. Mathcad had the same error and agreed their algorithm was in error.
Implementing state-of-the-art algorithms gets harder and harder for a particular problem because the algorithms found will be more complicated over time to be able to achieve speedup. It also means that developers should be able to follow publications. It is too costly to update all of the algorithms you are using. However an erroneous implementation is not acceptable of course. It should be corrected no matter what.
Only possible solution in the current software market infrastructure seems to be a dedicated implementer company which provides up-to-date implementations of the algorithms to other software developers. Even such a company exists and financially feasible, it will take some time to settle in the market.
Expecting vendors to change voluntarily is too optimistic. They will only do something when they see profit. Or when their reputation is damaged. And that is exactly what we can accomplish.
When you want to buy something like a stereo, there are many sites where you can read about other people's experiences. Something like that could be done for computational software as well.
A problem is that people may buy software for other applications and unknowinly use it for some application where it can give incorrect answers. Your idea is good for purely computational software. Also sometimes computational softwar is purchased in conjunction with a course; so the instructor can steer students away from poor choices and this can be important because the software may be utilized for years.
Dear William,
You have pointed out a very important discussion, since I have experienced many deficiencies in some simulation projects with out dated algorithms. As the best examples are two phase flash algorithms in thermodynamics which have been changed a lot during years but still out dated algorithms are available in open literature. Unfortunately many up to date algorithms (especially in thermodynamics) are mostly patented and licensed by huge corporations and therefore are totally out of reach of researchers.
A good comprimise may be to remove the erroneous software until an improved version is offerred.
- Badges
- Science topic
- Similar topics
- Computer Science and Engineering
- Algorithms