A decision-
support system developed by IBM Research provides
decision-makers with multiple options for reaching complex business
objectives. Here's how it's helping a paper
company
become more
efficient.
In Brief:
A CEO's search for a way to improve the
efficiency of his paper company in a continually changing business environment led him to IBM Research. After a thorough analysis of its operations, a team of researchers applied an artificial intelligence technique known as
A-teams to the problem. The decision-support system is already in operation and promises to produce significant savings.
How can computers help managers make the most effective business decisions? Experience has shown that people want more than just one suggested option for attacking any specific problem. Moreover, most problems are too complex to respond to a single solution.
Enter the concept of decision support, a notion that is rapidly gaining prominence in IBM. By means of various techniques, including artificial intelligence and operations research, it allows one to balance complex and often conflicting business objectives and thereby develop, for example, efficient production schedules.
A way of generating such schedules was recently developed by IBM Research for Madison Paper Industries, a company in Madison, Maine. The scheduling system offers several favorable choices, leaving the final decision to a person. So, instead of replacing human intelligence, the system works like an intelligence amplifier.
Madison, a long-time customer of IBM, crossed paths with Research in 1993, when Jack Chinn, then Madison's chairman and CEO (now chairman), read an article about how certain artificial intelligence programs had successfully solved business problems. Because the paper industry is cyclical and prices fluctuate widely with changing supply and demand, Chinn wanted a way to increase production without making the large investment needed to buy another paper-making machine. So he contacted IBM marketing representative Kate Kfoury, who steered him to the Thomas J. Watson Research Center.
"After several meetings, we derived a problem statement," says James Yeh, a department group manager at Watson. "Even though we knew nothing about the paper industry, we believed we could improve their scheduling because we had done it before in the steel and airline industries." What's more, the team realized that, unlike those industries, a scheduling system for paper companies would have hundreds of potential customers worldwide. That opportunity was exactly the kind that Research is seeking in its new focus on services, applications and solutions (see "Seeking New Opportunities").
The scheduler faces a daunting number of decisions when planning the operations of the company. For example, he or she must ultimately decide when particular orders should be produced, how the individual reels rolling off the paper machine should be "trimmed" so as to minimize waste while satisfying customers' needs, and how the final products should be delivered. The IBM system recognizes that these individual problems are tightly coupled. For example, deciding when to produce particular orders affects trim efficiency while placing significant constraints on what modes of shipment may be applicable. In
delaying an order so as to increase trim efficiency, the company might be forced to ship the final product
by a faster but more expensive mode of transportation (see "Pulp Fact: The Paper Trail").
Because a scheduler must weigh many factors quickly, computer-based tools can and do help. Past tools, however, have always focused on one particular task. IBM's team faced the particular challenge of devising a global scheduling system that would consider all factors and objectives together. The problem of scheduling is made more complex by the fact that the schedules must deal with multiple corporate objectives.
Indeed, among the many choices facing the scheduler, no single type of solution is always the best. Market conditions, manufacturing variations and a customer's willingness to tolerate a late delivery all change with time. Often, the different objectives oppose each other. For example, satisfying customers' requirements for different grades of paper requires frequent changes in the paper-making process; this disrupts the manufacturing flow and reduces output.
One step at a time
Although the paper industry was new to the Research team, such process disciplines as logistics, scheduling, planning, information systems and quality control were not. Still, instead of launching directly into a development project, IBM researchers offered to undertake a three-month paid study of Madison's operation. "We needed to understand all elements of a paper company: the raw materials, the manufacturing process, order-entry system, shipping, everything," says Yeh.
Holding a string of half-day interviews, researchers Alok Aggarwal, Ranga Jayaraman and Sesh Murthy picked the brains of Madison's workers. Using their previous experience with the steel industry as a basic template, they ended up gaining a deep understanding of paper scheduling.
When that consulting phase ended, the researchers produced three reports. One described Madison's business and how the scheduling could be improved; the second was a functional description of a proposed scheduling system; and the third explained how that system would mesh with Madison's existing management-information system. Chinn thought that the report described his company so well that he wanted all the employees to read it. "It showed them," says Yeh, "that we spoke their language and understood their concerns."
The proposed system, says Murthy, was no ordinary customer application. At its core, it has an artificial intelligence technique called A-teams (see "Calling in the A-teams"). Murthy claimed that, in harnessing the technique to the paper-scheduling process, the research team could build an integrated scheduling system that would take into account the paper machine, winding operation and load planning over a six-week period. Analysis showed that the proposed system could recoup its investment in about 18 months; comparable capital improvements can take up to three years to earn back. For IBM, which would retain the intellectual property rights to the system, the project would produce a valuable product for the paper industry.
Already, a suite of products derived from the scheduling system is being offered to customers by the Intelligent Production Decision Solutions group of IBM's Manufacturing Industry Solutions unit. "In benchmarking tests," says Robert Daigle, a brand manager in IBM's forest products segment, "the trimming software has outperformed similar software from other vendors."
Scientifically, the project represented a major advance in the field of optimization. The concept of A-teams was the subject of Murthy's thesis at Carnegie-Mellon University. Originally, it applied to problems that had only one objective. But, by drawing on his previous use of A-teams to build robotic manipulators, Murthy knew that it was possible to give an A-team system a set of objectives instead of just one. He also realized that, rather than specify a given procedure, the A-team system could offer several choices.
"Previous approaches put you in kind of a straitjacket: you either accepted the system's decision or not," says Murthy. "My main aim was to figure out how the machine could be a good assistant." To do that, concluded Murthy, the machine must first produce multiple solutions against multiple objectives, which is exactly what the A-teams do. In addition, the system had to let the operator modify a given solution using automated tools, which it does through a unique graphical user interface developed by Raymond Henry, a member of the project team.
Team-oriented approach
Dubbed a collaborative decision-support system, the product has helped employees with different agendas work in harmony. "The scheduling system gives all the people in the company a common view of the plant. It enables a team-oriented approach to paper scheduling," says Watson researcher John Rachlin.
The system permits everyone to see the scheduling options and the trade-offs they make, so that they can decide as a team what they can live with. "Some of the options are very different," says Watson researcher Fred Wu. "One gives a great width of paper, but makes several orders late. Another may have no lateness but disrupts the manufacturing schedule. They're all good in different ways. That's why people make the ultimate decision." To that end, says Aggarwal, "it lets them see the effects of different trade-offs and gives them the flexibility they need to work with their customers. Sometimes customers can accept a later delivery date, other times they can't." To optimize the company's output, the system even schedules pending orders and saves them in inventory until the actual order is placed.
How is the system working for Madison? "It exceeds our expectations," says Dan King, the paper company's information services manager. Even at this early stage, he estimates, the integrated scheduler is saving the company 1,400 tons of paper a year on its trimming operation alone. Other savings come from load balancing (i.e., shipping the finished products to make the most cost-effective use of both rail and truck transportation). Projecting present results, the annual savings from that source, says project specialist Paul Michaud, should amount to about another 10 percent. Perhaps the best judge is Adam Stearns who, as Madison's manager of customer service and logistics, is the company scheduler. "All the mechanical things I had to do before, the system does for me," he says. "It's like a dream come true."
Gil Bassak is a freelance science and technology writer based in Ossining, New York.
Seeking New Opportunities
The early 1990s saw marked changes within IBM Research, which led to the harnessing of its technological prowess to solve a broad range of customer problems directly. In industry terms, Research would focus on providing customer services, applications and solutions - a segment of the computer industry that accounted for 30 percent of the total revenue pie and was growing at a double-digit rate.
The best candidates were those for which solutions could be replicated within and across industries. In all regards, the integrated scheduling system for Madison Paper Industries presented a perfect example of the kind of project sought by Research.
More important, the same principles of optimization embodied in the system are applicable to nearly any industrial or commercial process. "It's something we can do to make our customers more competitive," says James Yeh, a department group manager at the Thomas. J. Watson Research Center. "It can help them reduce their costs, improve their time to market and respond to their market demands."
Pulp Fact: The Paper Trail
Paper starts as pulp, a wet mix of ground-up wood and other fibrous material. Blended with various ingredients to produce a specific grade and weight, called basis weight, the pulp is sprayed onto the moving wire, or web, of a paper-making machine. There it is pressed, dried and wound onto huge reels.
A machine can make only one grade of paper at a time, and it takes time to change over to another grade. Because a paper machine capable of making 1,000 tons of paper per day costs up to $500 million, a mill manager tries to maximize productivity by keeping the machine running as long as possible. He or she tries to minimize the transition time by shifting from one grade to a similar grade. Some of the paper produced during the transition may not meet the quality standards of either order. If the substandard paper cannot be sold, it is fed back to become pulp. Paper recycled in this way is called broke.
At Madison Paper Industries, the width of the paper leaving the paper machine, called the deckle, is almost 24 feet and must be trimmed to meet each order. The paper passes through a winder, which unrolls the reel and, using "slitters," trims the deckle into several narrower widths. To optimize production, a mill manager tries to trim the different orders in a way that makes the most efficient use of the deckle. Although some waste is inevitable, much of it can be recycled into new paper.
Calling in the A-teams
Asynchronous teams, or A-teams, are sets of computer programs that work independently of each other - asynchronously - to generate several good solutions to a problem. Each solution is evaluated and, based on that evaluation, may be kept, improved or eliminated.
According to Watson researchers Robert Fuhrer and Rama Akkiraju, the A-team architecture offers three unique advantages to tackling complex optimization problems such as those faced by a paper company. By using a broad-spectrum approach to solving complex problems, the system is better able to produce solutions than any single method could. In addition, multi-agent approaches to software development are naturally modular. If a new or better algorithm is later devised for solving a particular problem, a single agent to address the problem can be easily added without requiring the whole system to be redesigned. Finally, systems based on cooperative agents are more reliable. If one agent fails to work well on a particular set of inputs, other agents can pick up the slack.
An A-team consists of a population of solutions and three types of agents: constructors, improvers and destroyers. Constructors devise an initial solution based on the description of the problem; improvers work to make it better and, if successful, add the improved version to the population; and destroyers delete bad or redundant solutions.
Destroyers are needed because computer memory and processing capacity are finite; they could become bogged down by too many solutions. Destroyers aren't entirely destructive, however. While they are programmed to jettison solutions that, for example, cost too much, they are also designed to keep promising solutions that, with modification, could become good ones.
Improvers work to maximize a predefined set of desired conditions within a solution. In the case of a paper company, these characteristics include profitability, paper quality, on-time delivery and a smooth manufacturing flow. For example, an improver might select one solution and try to maximize the number of orders produced while reducing the number of disruptions to the paper-making machine.
The idea for A-teams stems from several problem-solving methods pioneered during the past 20 years, such as genetic algorithms and simulated annealing (a technique developed at the Thomas J. Watson Research Center in the early 1980s).
Simulated annealing is a technique for finding one or more global optimal solutions for a system that does not have a proven "best" solution. Similarly, the A-team working in the integrated scheduling system for Madison Paper Industries does not seek a single "best" solution. Instead, it looks for several very good solutions that produce better results than others in the population.
As an illustration, consider a population of production schedules evaluated by on-time production and paper quality. Here, schedule A "dominates" schedule B, because it has lower tardiness and lower quality at risk. However, schedules A and C are nondominated because, although schedule A has a lower quality, schedule C has a lower tardiness. Within that population, all the nondominated schedules lie on the Pareto frontier (a curve named for the 19th-century economist Vilfredo Pareto). If any one schedule is on the Pareto frontier, no other schedule can be better than it in all objectives.
The solutions on the Pareto frontier are presented to the individuals who schedule the work. From that point on, the final decision of which schedule to pick is left to the schedulers.
The A-team approach can be applied to various kinds of process industries. Murthy, who brought the concept of A-teams to the integrated scheduling system, once applied the related concept of simulated annealing to the production of steel, and Watson researcher Martin Sturzenbecker has developed an object-oriented implementation of the A-team architecture that is independent of the problem domain. Watson researcher Alok Aggarwal, for example, intends to use it for solving problems in petroleum refining and nursing home-care scheduling.
