WINNER: University of Oklahoma’s Advanced Regional Prediction System
INNOVATOR: Kelvin Droegemeier
Seeing as he shares his first name with a temperature scale, it was evidently Kelvin Droegemeier’s destiny to make his mark in the field of weather. Droegemeier, a computer scientist at the Center for Analysis and Prediction of Storms at the University of Oklahoma in Norman, got his chance in the late 1980s when the U.S. government asked him to analyze an unprecedented trove of meteorological data coming in from a network of 160 radar-based weather sensors that were just being set up around the country. His task was to see if he could use the data to predict the progress of storms. Unfortunately, the stations monitored only two variables--the speed of the wind that was parallel to the radar beam, and rain intensity. That may have been more information than meteorologists had ever had before, but it still wasn’t enough to make predictions a few hours ahead of time. Droegemeier needed to solve equations that require the perpendicular components of wind speed, as well as temperature, air pressure, and ten or so other variables. How could he possibly come up with a way to fill in the missing data? It’s a computational problem that’s just a nightmare, he sighs.
Droegemeier’s group got around the problem by programming the center’s supercomputers to guess at a prediction and then run the equations backward to see how consistent the resulting initial weather conditions were with the known data. Then they adjusted the initial conditions and ran the equations forward to get a new prediction. By repeating this process 50 times or so, continually fine-tuning the predictions and the values of the missing data, the team was able to predict storms fairly accurately up to six hours ahead of time. In tests in June 1996, in fact, the program accurately predicted the occurrence of storms six or seven hours in advance about 80 percent of the time. That level of accuracy could save U.S. industry more than $14 billion a year in avoidable storm-related damage, as well as saving lives.
Is that what Droegemeier’s parents had in mind when they named him? Not likely. I’m pretty sure they picked ‘Kelvin’ out of a book, he says.
Bad Drivers of Beantown
MIT’s Microscopic Traffic Simulator
INNOVATOR: Qi Yang
Qi Yang hates getting stuck in traffic as much as the next guy, but at least he can say he did something about it. The mit researcher developed software that simulates traffic jams right down to the nasty habits of individual drivers. The program is already helping at least one city, Boston, plan its way out of clogged roadways.
Yang’s is not to be mistaken for the garden variety traffic simulators that for years have helped city planners predict the effects of a new road, say, or more traffic lights. Planners of Boston’s multibillion- dollar central artery project discovered that existing simulators simply didn’t provide accurate predictions, because they assume all drivers behave the same average way. As anyone who has ever driven a car in Boston knows, that is a dangerous assumption. Not everyone respects traffic rules, Yang deadpans.
To correct that flaw, Yang and other students got together to create a 100,000-line computer program--complete with millions of digital drivers who display every combination of several dozen different idiosyncrasies, and who drive 15 different kinds of motor vehicles, from compacts to big rigs. There are slowpokes and people who tailgate; young drivers who can react to an event in less than a second, and old drivers who may require nearly three; lane changers and lane huggers. As it turns out, the difference between driving styles has a big effect on the behavior of traffic, which is obvious to anyone who has seen a quick-lane-changer encounter a slowpoke in the middle lane.
Yang is now adapting the program to simulate the flow of traffic around outer beltways, such as the one in chaotic Washington, D.C. These road networks have a lot more path choices in them, explains Yang, for whom understatement appears to be a policy. It’s going to take a lot of debugging.
Sailing a Sea of Inflection
Motorola’s Chinese Voice Recognition Software
INNOVATOR: Steve Austin
Having spent most of his career trying to get computers to understand spoken English, Steve Austin thought that, at least where speech recognition was concerned, he had pretty much heard it all. But when he picked up his new assignment a few years ago, he suddenly felt as if he were back in high school. Austin was now supposed to get computers to understand spoken Chinese.
Although Austin, a computer scientist, doesn’t speak much Chinese, the experts he brought onto his team made him appreciate the magnitude of the task. For starters, a Chinese word can have more than half a dozen entirely different meanings, depending on inflection. The word ma, for example, means mother if uttered in a flat tone, numb if in a rising tone, horse if your voice dips as you say it, and curse if the word is barked out. What’s more, Chinese words in a sentence can be combined in far more ways than English ones, offering no standard way of breaking the sentence apart to analyze individual words. It makes the task of transcribing a sentence combinatorially large, says Austin.
To solve the problem, Austin and his team created software that looks at the frequency of the voice and whether it is rising or falling and distinguishes its inflection. It looks not just at individual words but at whole sentences, to help pull out word meanings by their relationship to one another in the sentence. The resulting prototype software, finished at the end of last year, can transcribe Mandarin Chinese that is spoken at a natural rhythm and pace--which is more than most other programs can do for the English language.
The benefits of the software, expected to be available for pcs later this year or early next, could be enormous. To reproduce the thousands of written characters, Chinese keyboards employ various arcane schemes that can take years to master. Speech recognition may be the best hope for bringing computing to the masses. It’s an input scheme that everyone knows how to use, says Austin.
Stanford University and Extempo Systems’ Imp Software Characters
INNOVATOR: Barbara Hayes-Roth
The free-flowing banter of on-line computer chat rooms can be intimidating to the neophyte. Even when you manage to get a word in edgewise, there is a good chance you’ll be either ignored or flamed. And chat rooms are proliferating so fast that there aren’t nearly enough good moderators to go around.
Barbara Hayes-Roth has tapped a new labor pool of chat-room hosts, all of them expert at making people feel at home on-line. She calls them imps. They greet people, get them talking, and introduce them to others in the chat room. They are, of course, computer programs, but sometimes they take on cartoon-human form, moving around the room to mingle, serving virtual drinks and even smoothing over those awkward moments in virtual conversations. They’re there to acquaint people to the environment and engage in social activity, just as a greeter at a party would do, she explains.
Two years ago Hayes-Roth, a professor of computer science at Stanford, formed her own company--Extempo Systems, in Santa Clara, California--to develop the chat-line imps. They were finished last December. They function by evaluating a menu of hundreds of possible actions and comments against an ever-changing list of conditions in the chat room. Imps can be custom-tailored to the needs of a particular application or be imported as fully formed characters. One favorite is Erin, a slightly punkish bartender. If you wander into a virtual bar and ask Erin for a drink, she might suggest Coke or Pepsi, flirt with you, or ignore you if she thinks you’re being rude.
It’s a different sort of problem from traditional artificial intelligence, in which there’s an emphasis on deep expertise, and where there’s a right thing to do, she says. Here the goal is to create characters that are endearing and engaging. It opens up different challenges. And needless to say, it’s a lot of fun.
W$%t Doe5 This S@y?
The State University of New York at Buffalo and the U.S. Postal Service’s Automated Postal System
INNOVATOR: Sargur Srihari
Anyone who’s ever tried to decode bad handwriting can understand why postal workers have a reputation for being tightly wrapped: they read some 50 million hand-addressed envelopes a day. Sargur Srihari, however, is trying to eliminate that frustrating and time-consuming task once and for all.
Srihari, director of the Center of Excellence for Document Analysis and Recognition at the State University of New York at Buffalo, has come up with a handwriting-recognition computer program capable of pulling useful information from chicken-scratched envelopes. Similar programs, particularly the ones on tiny handheld computers known as personal digital assistants, have been around for a while, but they have the advantages of being able to feel how each letter is formed on an electronic pad and to learn an individual’s handwriting style. Such software would be woefully inadequate at untangling the typical envelope address, in which individual characters are often obscure, if not dropped entirely. When I started on the problem a decade ago, says Srihari, nobody had even thought of dealing with this.
Rather than having his program try to decipher one character at a time, Srihari taught it to consider the overall shape of the word, which can yield important clues in cursive writing. The program also uses whatever pieces of information it can pull out of the address to narrow down the possible meanings of otherwise illegible parts. For instance, if the program knows from the zip code that a letter is addressed to Camden, New Jersey, and it can read only the first four letters of the street name- -say, camd--then it can look up all possible street names that match and come up with Camden Court.
Despite these tricks, Srihari’s program still can’t read 40 percent of handwritten addresses. Nevertheless, in a 34-site, one-month test in December, the U.S. Postal Service saved $5 million. Once the technology is installed in every processing center in the country in September, it should start saving taxpayers at least $50 million a year, the postal service estimates. And who could ever put a price on the aggravation saved the men and women in gray?