CHAOS CONTROLLED Professor Christos Cassandras, Harnessing Technology's Power   By Bari Walsh
On just about every street corner, machines loaded with cash stand on the alert, ready to dispense their loot to all comers. By now, getting cash on demand seems almost as unremarkable as picking up the phone and hearing a dial tone. But the systems that control both of those modern conveniences are far from mundane. They're laden with sophisticated computer-based technologies that help them juggle a multitude of competing requests simultaneously, instantaneously. And we expect them to work perfectly every time.	Professor Christos Cassandras and Graduate Teaching Fellow Molavine (Alvin) Widitora in the Automated Design and Manufacturing Systems Laboratory.
It's hard to remember what life was like before ATMs; soon, perhaps, life without microchip-enhanced sunglasses offering instant connection to our e-mail will be just as unthinkable. The growing complexity of the products and services that collectively constitute modern life-even those as seemingly low tech as traffic lights and elevator systems-has given rise to a whole host of engineering challenges. How can we build and maintain computerized systems that respond not only to competing demands from so many quarters, but also to the uncertainty caused by ever-changing conditions and the fallibility of us humans, the creators and users of all this new technology? This is precisely the kind of question that Christos Cassandras is grappling with in the Control of Discrete Event Systems (CODES) Laboratory at the College of Engineering. Cassandras, a professor in the Departments of Manufacturing Engineering and Electrical and Computer Engineering, was one of the first researchers to recognize the need for new tools to analyze and control technological systems whose unprecedented capabilities are matched by their unprecedented complexity. "Complexity is a double-edged sword," Cassandras says. "The more complex we make a system, the more it can deliver, presumably, in terms of services and improving our lives. But we don't know how to deal with complexity when we design systems." With his CODES lab colleagues, Cassandras is working on a project with the Electric Power Research Institute and the Department of Defense to improve the security, performance, and robustness of the interconnected, complex networks that control America's energy, telecommunications, transportation, and financial infrastructures. In a fundamental sense, the computer systems on which these infrastructures rest, and on which our economy and society depend, are out of control: they are not overseen by any single entity, and the level of their complexity threatens to outstrip our ability to keep them running smoothly. In a similar vein, Cassandras and Manufacturing Engineering colleagues Michael Caramanis and Ioannis Paschalidis are working on the National Science Foundation's Knowledge and Distributed Intelligence initiative, funded by a $1.2 million grant. The issues raised by both projects demonstrate, as Cassandras says, that "right now, we're catching up with ourselves. It's perhaps the first time in the history of humanity that technology is ahead of science." Though This Be Madness Cassandras was at Harvard in the late 1970s, doing his doctoral work with Professor Y. C. Ho, when he got his first taste of discrete event systems, though that term didn't exist yet. The two researchers were coming up against certain manufacturing problems that they didn't have the mathematical tools to solve, and they realized that "essentially the entire history of science had been based on models and premises driven by the realities of nature," Cassandras says, the laws of gravity, mechanics, and physical chemistry. What these laws measure-displacement, velocity, acceleration, temperature, pressure, and the like-are all continuous variables, meaning that they can take on any real value as time moves. Cassandras and Ho saw that a shift had taken place. In human-made technological systems, the important variables are "not quantities that change with time, but quantities that change with events." Cassandras and Ho realized that human-made entities like automated manufacturing systems, intelligent transportation systems, and communication networks are driven by events, such as the pushing of a button, the execution of a keyboard command, or a traffic light turning green. They're governed by discrete quantities: the number of parts in an inventory, the number of planes on a runway, the number of telephone calls that are active. The systems, therefore, must act as event coordinators, responding to occurrences both foreseen and unforeseen. Think of a network of traffic lights. Everyone knows the frustration of watching each light along a stretch of road turn red upon approach, or sitting in gridlocked traffic at a busy intersection as some lights change too fast, others too slowly. In exploring ways to better control such networks, Cassandras says, one first has to understand that the variables in a traffic light system are "green, yellow, and red. There isn't a natural way of measuring it-it's just green, yellow, or red. We just don't have the mathematics to control that, or even a way of thinking about it." In a system like that, he says, it's not time that causes changes; it's the asynchronous events that happen throughout the day-the ebb and flow of traffic caused by weather, accidents, and breakdowns, for example. "We realized we needed a new way of looking at technological systems where things change because events occur randomly-some controllable, some not controllable. "On his Web site, Cassandras quotes Hamlet: "Though this be madness, yet there is method in't." That is his hope. His job is to find the method. Teaching Machines to Learn At the CODES lab, Cassandras and colleagues develop not only basic mathematical and computer simulation models that help them analyze how machines respond to discrete events, they also develop software based on something called Rapid Learning Technology (RLT) to control and optimize the machines.With RLT, the researchers are teaching machines to learn, to respond to unpredictability, and to correct for the errors that we, their creators, build into them. "The way people learn," Cassandras says, "is through trial and error. They try something one way, another way, a third way, and they pick what works. But what if there are a billion different options?" With RLT, machines can be programmed to learn through a sort of abbreviated trial and error. They can try one option, and while they are weighing the results of that attempt, they can infer what might have happened if different but related solutions had been attempted. "We're teaching the systems to do thought experiments, something Einstein, for instance, has shown can be very effective," says Cassandras. "While you do one trial, you learn about a million different options. It's philosophically exciting, and practically, too, because it means that you have to come up with the right algorithms, the right software, to implement these thought experiments." Although Cassandras's work has always had a theoretical bent, he is constantly motivated by "real problems, real juicy problems," and he maintains a healthy interest in both theory and application. "As engineers, it is our obligation to keep our feet on the ground and look for applications for theoretical work," he says. "But for me, it's also the reverse: the real world is a wonderful source of great problems that eventually give rise to theories." A great real-world problem for a discrete events theoretician, he says, is an elevator system. Elevators move within the constraints of natural laws, but discrete events govern their operation. Cassandras has worked with Otis Elevator to program elevator systems to find efficient ways to respond to these events. "If you go to a large office building, there are usually between six and eight elevators and sixteen to thirty floors. When you press a button on the fifteenth floor and want to go down, which elevator should stop for you? Does it make more sense for the first elevator that approaches to ignore you, if it's just going to have to stop on every floor below you until you get to the third, or wherever you might want to go? How do you coordinate the elevators so that no one is waiting too long on any floor? The problem is on the same level of complexity as a game of chess. There are so many possible scenarios." Cassandras developed schemes to improve elevator scheduling through software that enabled the elevator system to learn in five working days what the optimal schedule for individual cars should be in specific situations. "Of course, that won't be without error, due either to the people designing the software or the inherent uncertainty of the system itself," he says. "All of these systems have the same element of uncertainty. The trick is to anticipate it."

 

[Table of Contents] [Current Issue][Archives] [ENG Home]

Copyright Trustees of Boston University
Last updated on: .