I used to live in Lubbock, Texas, site of one of the nation’s worst tornado disasters in the mid-70s. During the late ‘90s, my student and I did a research project on this, and reported on it in: “Tornado Prediction: A Markov Decision Process," Zhanyu Ge, Donald Wunsch, Loren Philips, The Conference on Severe Weather, Lubbock, TX, February 9-11, 1999.

It turns out that we can address the fundamental scientific problems in tornado detection, which would probably lead to some increased warnings. The more vexing problem is in software.

I’m fond of saying that we’re still in the stone ages of software. One of the most serious problems is one that we encountered in this project. Basically, the National Severe Storms Laboratory in Oklahoma collects high-quality data on such storms. However, this data is not easily accessible for training neural networks or other machine learning algorithms. The problem is that software development costs are high, simply for developing the interface. In addition to the costs of software development increasing over time, because it is so labor intensive, there are also hurdles that a technically unnecessary – namely, that different systems don’t talk to each other very well. This means that having someone trained in statistics and machine learning techniques isn’t sufficient – you might also need a full-time software engineer.

We will get past these barriers and create many new solutions, mostly software based, that address problems like tornado detection. The techniques are improving all the time. But, if and when these barriers are finally lowered, get ready for a blistering increase in the rate of progress.