Page 27 - 2016Fall

Richard Hamming,
Professor of Electrical Engineering, Naval Postgraduate School, Monterey, CA (1915-1998)
The business of computing is insight, not numbers (Hamming, 1973).
George E. P. Box,
Professor of Statistics, University of Wisconsin (1919-2013)
Essentially, all models are wrong, but some are useful. Remember that all models are wrong. The practical question is how wrong do they have to be to not be useful (Draper, 1987).
Clark’s Law of Heroic Efforts
If you “bust your chops” in an heroic effort,
you will usually be repaid with compound ingratitude (Clark, 2016).
It is tempting to use the latest cool fad algorithm that is making its way through the literature. However, many times people do this while putting insufficient priority on the fun- damental physics and signal processing of the problem.
“Data chasing” is a term coined by my friend, colleague, and mentor Dr. James V. Candy of the Lawrence Livermore Na- tional Laboratory (LLNL). Data chasing is what one does when (1) one knows little or nothing about the physical processes that created the data being processed (no mod- els or prior knowledge), (2) one has no access to controlled experiments in which the signal-processing “answer” is known, and/or (3) one applies various filters, DFTs, and ad hoc signal-processing algorithms to the data, yet one cannot explain the meaning of the results.
Data chasing and the use of ad hoc algorithms can have very serious consequences, including the following. (1) One does not understand the meaning of the processing results. The results are often inconclusive and/or not useful. (2) Results are often not repeatable by other researchers. (3) If the algo- rithms work, one does not know why they worked. If the al- gorithms do not work, one does not know why they did not work. (4) One does not know what to do to make things bet- ter, yet by this time, one has probably exhausted one’s time and money for the project. (5) The results are usually not ex- tensible. Model-based signal processing can often mitigate these shortcomings of data chasing (Candy, 2006).
Data chasing leads to enormous waste and stress. Countless times, people have come to me near the end of a project with a project review scheduled in a few days or weeks. They typi- cally ask me, “What signal-processing magic can you do to save the project?” Usually, the measurements are inadequate for processing.
First, I tell them that I'll apply my special "SESP (sow's ear- to-silk purse) Algorithm." After we chuckle, I tell them that the experiment is flawed, the data are inadequate, this is a Phase III or IV project, and they really should have includ- ed me in the experiment planning at the beginning of the project. I tell them that I do not work on Phase III projects. Of course, that gets me nowhere, and I end up “busting my chops” during nights and weekends in a vain attempt to put a last-minute Band-Aid on a gaping wound. Finally, when the project cannot be saved, I get blamed.
When a manager asks you to do something beyond the call of duty to save him/her from looking bad or to do something absurd or wrong in some way, he/she usually says something like, “I’ll remember this at raise time.” Don’t believe it. My colleagues and I have many horror stories to the contrary. Later, the manager usually says something like, “I never said that!” Also, the ingratitude is not simple. It is exponential (compound ingratitude).
Horror Story About an Heroic Effort
Nikola Tesla came to the United States in 1884 from Croatia and was hired by Thomas Edison. After about a year, Edison was impressed by Tesla’s abilities and offered to give him a $50,000 bonus if he could create an improved design for Edi- son’s direct current (DC) dynamos. After months of work, Tesla delivered the desired solution and requested his bonus. Edison replied, “Tesla, you don’t understand our American humor.” Tesla resigned soon after and went on to create a vast legacy of important inventions (History.com Staff, 2009).
The ways to avoid data chasing (aside from avoiding Phase III projects) include the following. (1) Study, really study, the physics/science of the problem and use all possible prior knowledge. (2) Build and validate models of the physical process and the measurement system that produce the mea- sured signals. Do simulation studies. Use first-principles models, nonparametric models, or parametric models that give your insight into the measurements. (3) Involve a statis- tical signal-processing specialist in the very beginning parts of the project, including basic system design and especially experiment design. The performance of signal-processing
Fall 2016 | Acoustics Today | 25