Genetic Algorithms Attack Insoluble Business Problems A significant challenge in managing today’s enterprise is quickly and effectively choosing among a large number of complex options and alternatives with subtle trade-offs. Fortunately, new computer applications using genetic algorithms can do a much better job at making these choices than a person using traditional analytical techniques.
How do they do this? Genetic algorithms allow people to “breed” a population of computer programs using the concept of natural selection.1 They are typically used as part of a computer simulation in which alternate solutions to a problem are selected, combined with other solutions, and changed to reflect new learning. Then each newly created potential solution is either kept or dropped, depending on how “fit” they are at solving the problem based on the criteria that were established at the start of the process.
In the beginning, a genetic program spawns a wide range of possible solutions at random. Some of these are a poor match for the target criteria and are eliminated. Other ideas match some of the criteria and survive to the next generation, where their features are combined with the features of other survivors of the first round. Just as geneticists create stronger offspring in each new generation by recombining the most favorable traits of the previous generation, genetic programmers breed better solutions by mixing the best qualities of promising ideas in each generation.
Typically, the top nine percent of the solutions are copied from one generation to the next, so that the best ideas are not lost in the next round. Another one percent of the solutions undergo mutation in the hopes of finding a radical improvement. The remaining 90 percent are the result of recombining the offspring of the previous generation of solutions.2
We believe five areas will benefit most from the power of genetic programming. The first of these is training neural networks. As the price of advanced microchips and microprocessors falls, researchers will create ever more complex neural networks. These neural networks must “learn” how to perform certain tasks. Unfortunately, they are so intricate that no human programmer could explicitly instruct them how to think in all situations.3 Instead, training software must be generated using an approach that simulates biological evolution. The neural network will evolve over many generations of random mutations and breeding to train itself for all possible situations.
The second growth area we foresee lies in autonomic computing, which we just discussed in the context of Trend #4. Autonomic computing gives networks and computer systems the ability to regulate and repair things that now require human thought and intervention. Basically, genetic algorithms allow computers to program themselves. The user states the problem, and the computer designs a program to solve it through natural selection. The software also fixes its own bugs by automatically creating generations of possible remedies to a glitch, recombining the best parts of the different solutions, and then implementing the patch that works.
The third major area that will benefit from the power of genetic programming is product design. According to Scientific American,4 this is likely to be “the first practical commercial area for genetic programming... Design work is especially well-suited to genetic programming because it presents tough problems for which people seek solutions that are very good, but not mathematically perfect. Generally, there are complex trade-offs between competing considerations, and the best balance among the various factors is difficult to foresee.” Finally, design work usually involves discovering topological arrangements of things (as opposed to merely optimizing a set of numbers), a task that genetic programming is very good at.
Several researchers and companies are already leveraging the power of genetic algorithms in new product design. For example, GE used genetic algorithms to help design the Boeing 777 engine in the late 1980s. General Motors used them to design new automobile frames that were lightweight yet absorbed more energy in crashes. More recently, researchers created a “virtual inventor”: a computer program that would use genetic algorithms to randomly come up with ideas for potential new products. When they ran the program, it suggested 15 new ideas. Amazingly, all of these ideas had already been patented ? five of them by major research institutions in the past three years! Quite an impressive performance record.
The fourth area of enormous potential for genetic programming is scheduling. Genetic algorithms are revolutionizing the use of scheduling to cut costs and gain competitive advantage.
For example, John Deere uses a genetic algorithm-based optimization tool called Evolver, which is an Excel add-on from Palisade Corporation. Evolver has helped Deere to optimize scheduling problems at some of its factories, as explained recently in the International Journal of Manufacturing Technology and Management.5
Deere wanted to find the best way to manage its transportation and delivery schedules, factory runs, and production capacity. Now, every night before the factory closes, a genetic algorithm program ? running on a PC ? generates a population of random factory schedules. The program then “breeds” this schedule and evaluates each new generation of schedules, rating them based on their ability to cope with the current backlog of demand. The best schedules are selected to participate in the next round of “breeding.”
This process produces more than 600,000 possible schedules each night and picks the best one to use the following morning. As a result, the output at the Deere factory has improved significantly.
The fifth area to benefit from the power of genetic engineering is predictive and business intelligence. Many business tools, such as data mining, work because they sort through large amounts of information to fit trends to past data points. In many cases, these tools cannot detect sudden shifts in customer behavior or external jolts suffered by the industry.
A genetic algorithm, however, takes business data and treats it like genetic material to simulate future behaviors. Based on the principles of mutation and cross-breeding, analytic software incorporating genetic algorithms can assess a broader range of potential competitive situations. Predictions of customer, supplier, and competitor trends thereby become more accurate ? and more sensitive to situations formerly thought of as implausible.
Based on all of the factors we’ve discussed, we forecast the following four developments:
Autonomic “black box” applications using genetic algorithms will become available even to small companies and individuals, dramatically improving business productivity. Analysts anticipate extremely powerful yet moderately-priced desktop workstations by the end of the decade. Today, John Deere already generates its manufacturing schedules using Evolver, an off-the-shelf software application running on current PC hardware. Fifty-gigahertz computers that perform 50 billion operations a second should be available toward the end of this decade. These workstations will run genetically programmed applications of exponentially increasing sophistication. These systems will provide sufficient computing power to make many kinds of mission-critical systems self-programming. This has the potential to dramatically improve productivity in the five application areas we discussed earlier, and potentially far beyond. However, because genetic algorithms operate as black boxes, using mutation and natural selection, rather than explicit, human-derived logic, they possess a unique ability to generate problems if not closely monitored and cross-checked. Therefore, we anticipate the failure of isolated firms whose executives decide to implement these powerful techniques without sufficiently understanding the operations of the “black boxes” they create.
As companies adopt genetic programming software tools, we will witness the rise of a new class of business professionals conversant with this technology. This will have multiple effects on the IT industry and for MIS professionals. Coupled with the off-shoring of generic programming to India and China over the next decade, autonomic programming will eventually obsolete most traditional IT jobs in the developed world. We expect an oversupply of undifferentiated computer science professionals in seven years, and a concurrent rise of a small cadre of MIS specialists who understand how to integrate genetic programming into a company’s existing IT architecture.
We anticipate an explosion in newly patented inventions. Genetic programming will be used to create invention incubators whose purpose is to discover, patent, and nurture inventions. Pioneers and fast-followers in the electronics, chemical, pharmaceutical, software, and financial trading industries ? where product design will benefit from untiring invention incubators that operate 24/7 ? will reap considerable profits. These firms will rapidly shorten the average product life-cycles in their industry sectors, and the many firms that ignore genetic programming’s ability to turbo-charge product development will fall by the wayside.
Hackers will use genetic algorithms to evolve new kinds of viruses that stay ahead of security technologies by evolving faster than they can be thwarted. These “super-viruses” will proactively search out victims, in contrast to today’s seemingly passive computer viruses. As a result, we expect an explosion in identity theft and in commercial and retail fraud on the Internet. However, this represents an enormous opportunity for leading-edge computer security firms that start preparing anti-viral products and services now to eradicate even those rapidly-evolving viruses.
References List :1. Scientific American, February 2003, "Evolving Inventions," by John R. Koza, Martin A. Keane, and Matthew J. Streeter. ⓒ Copyright 2003 by Scientific American, Inc. All rights reserved.2. New Scientist, January 9, 1999, "Clever Kitty," by Duncan Graham-Rowe. ⓒ Copyright 1999 by Reed Business Information, Ltd. All rights reserved.3. MIT Technology Review, November 2003, "From Artificial Intelligence to Artificial Biology?" by Claire Tristam. ⓒ Copyright 2003 the Massachusetts Institute of Technology. All rights reserved.4. Scientific American, February 2003, "Evolving Inventions," by John R. Koza, Martin A. Keane, and Matthew J. Streeter. ⓒ Copyright 2003 by Scientific American, Inc. All rights reserved.5. International Journal of Manufacturing Technology and Management, 2000, Vol. 2, "Complex Systems Theory: Implications and Promises for Manufacturing Organizations," by Ian P. McCarthy, and Thierry Rakotobe-Joel. ⓒ Copyright 2000 by Inderscience Enterprises, Ltd. All rights reserved.
