Is there a connection between evolution and global supply chains? Yes, and a close one, if current trends continue. Evolution and other features of the natural world are being used more and more to solve supply chain problems that we haven’t been able to tackle before.

Evolution can be a very powerful way to optimize, when it is computerized and when it is used to solve supply chain problems. In North America, genetic algorithms, invented in the 1970s by John Holland, have been used by many companies since the 1990s to find better solutions to supply chain problems. Supply chain software companies are capitalizing on the new capabilities that these algorithms (i.e. problem-solving techniques) provide us for exploiting the computer power available to us now. 

Genetic algorithms are one way of putting evolution on a computer. The algorithms mutate, cross-breed, and evolve solutions to hard problems until what is produced is very good—in many cases, better than the results that human experts can find. Genetic algorithms make it possible to solve problems that are not solvable by mathematical techniques—not, at least, in any reasonable amount of time.

Genetic algorithms are one of a family of approaches to computerized evolution, each with success at solving hard problems in the supply chain world. Other related approaches are evolution strategies and evolutionary programming. Although these approaches differ in detail, they are all members of the family of evolutionary optimization techniques, and each is a way of harnessing evolution to a computer to help us solve hard supply chain problems.

When you need a very good solution fast

Computerized evolution doesn’t necessarily find the best possible solution to a problem. Why, then, are so many people using it? Because these techniques can find very good solutions to a wide range of problems, and they can find them quickly. Relaxing the requirement to find the best possible solution makes many things possible that weren’t possible before.

As an example, Air Liquide Large Industries U.S. LP, an industrial gas company based in Houston, uses genetic algorithms in two systems created for it by NuTech Solutions. The first system optimizes Air Liquide’s production and distribution of liquid oxygen and nitrogen. The system simultaneously schedules the operation of more than forty plants and distribution by truck to more than 10,000 client sites. By optimizing both production and distribution at the same time, the system saves a good deal of money over the solutions generated by commercially-available supply chain software, which solve first one problem and then the other. Charles Harper, Air Liquide’s Director of National Supply and Pipeline Operations, says, “The system gives us a new understanding of what is possible using contemporary approaches to optimization.”

A second system created for Air Liquide optimizes pipeline operations, controlling more than 1,800 miles of pipelines that deliver oxygen and nitrogen to Air Liquide’s clients. In addition to the pipelines, the system optimizes production of oxygen and nitrogen at the plants that feed the pipeline, and takes into account demand forecasts, client contracts, power costs, and other aspects of the problem in order to create a unified production and distribution plan. The system it is different from traditional approaches is many ways. Perhaps most important is the system’s ability to control devices in the pipeline that can change their function—and the way the pipeline flows—with a signal from the Operations Control Center.

Both these systems illustrate the way that a genetic algorithm can solve the whole problem that is facing supply chain personnel, as opposed to solving sub-problems one at a time and stitching the solutions together to produce a plan that is sub-optimal. 

The techniques used in these applications can be applied across the range of supply chain problems. If we accept very good answers instead of perfect answers, we can get them very quickly. The result is that we can take on much larger problems than we could solve using mathematical techniques.

Borrowing from the experts

Although we may think of evolution as a process driven by random mutations, genetic algorithms don’t have to operate blindly. It can be very useful to add to them some of the techniques that human experts use. When this happens, the algorithms may speed up dramatically and the solutions they find can be much better. At Air Liquide, adding rules of thumb that human experts used to schedule production doubled the speed of the liquid gas optimization system.

What’s more, as the experts analyzed the solutions produced by the system, they began to learn new rules. “The system is shifting production in bulk more often than I would, but it’s saving a lot of money by doing that. It looks like it does it only outside a radius of 200 miles.” The system worked faster when it used rules like these.

If evolutionary algorithms are used in a production environment, a very interesting thing happens. The experts can learn from the system and the system can learn from the experts. Evolution happens outside the scope of the system as well as inside it.

Other ways to borrow from nature

Genetic algorithms are inspired by processes we have observed in the natural world. There are others:

Ant colony optimization is a technique that is based on the foraging behavior of ants. The Air Liquide system described above includes a genetic algorithm and an ant colony optimizer, working together to solve the overall problem of production and distribution. If you search for “ant colony optimization” on the Internet, you may be fascinated to learn how ant behavior can produce a useful computerized optimization approach.

Simulated annealing is an optimization technique that is inspired by the behavior of metals as they are heated and cooled. It is another technique for rapid solution of complicated problems. A search on the Internet will show you how, in the hands of physicists, principles from the science of thermodynamics gave rise to a very useful optimization algorithm.

Particle swarm optimization is an optimization technique that is inspired by the behavior of animals in swarms, herds, and flocks. A search on the Internet will show you how swarming and flocking behavior inspired computer scientists to produce very effective optimization algorithms.

There is a great deal of power in optimization approaches inspired by processes in the natural world, and the 21st century will probably see the use of many more of them. Genetic algorithms are the most widespread of these, but we predict that many more optimization approaches inspired by natural processes are yet to be developed.

Increasing our certainty about uncertainty

One of the biggest problems with supply chain plans is their performance when something unexpected happens. We may be uncertain about the future, but when we use our traditional problem-solving techniques we are often forced to eliminate that uncertainty as we describe the problem. We tell our computers that our truck’s travel time will be three hours and twenty minutes, that the price of power next Tuesday will be $78 per kilowatt hour, and that the level of demand next month will be 23,471 units. 

It is very unlikely that any of these predictions will be precisely accurate, and yet the solutions we produce using traditional techniques may be tied very tightly to the predictions we gave when we posed the problem.

A good way to capture uncertainty about the future is to include that uncertainty in our descriptions of our problems. Suppose, instead of telling the computer that travel time will be three hours and twenty minutes, we tell it instead that there are different possible travel times, and that they have different probabilities. We tell the computer that power prices may fall somewhere in a range next Tuesday, we tell it what the chances are for different parts of that range, and so forth.

What we gain from doing this is the ability to simulate the performance of our plan again and again, using these probabilities in each simulation to determine what happens in that particular version of the future. If we said that travel time will be four hours or more 10% of the time, then in about 10% of the simulations we would see travel times greater than four hours. A critical part of these simulations would be the reactions of the simulated dispatchers and planners when the unexpected happens, and the effects of their actions on subsequent activities as the simulation plays out.

It is possible that none of these simulations would be the same, but taken together they would give us a very good feel for the kinds of things likely to happen. What we gain by running many simulations, in other words, is a picture of the range of futures that are possible using our plan. 

There are many things we can do by studying a range of outcomes for a plan. We can look at the average performance of our plan—the mean of its performance across the range of outcomes.  We can look at the best and worst outcomes of our plan. (In order to reduce risk, it may be important to find a plan that has worse average performance but a better worst case.) We can modify our plan in order to look for a better distribution of outcomes. We can see whether there are consistent problems that arise across many of the simulations, and consider ways to reduce their impact as we modify our plan.

The way to gain certainty about uncertainty, then, is to include uncertainty in our description of the problem. We then create many simulations of the way our solution plays out. The range of outcomes of those simulations gives us some certainty about how our plan will perform, even in an uncertain world. In this way, twenty-first century companies can overcome one of the hidden pitfalls of business: making decisions based upon averages of averages.

Optimizing under certainty

In our view, the 21st century will be the century of the simulation of uncertainty, and evolutionary algorithms are just the tool to use when we are optimizing under uncertainty. Evolutionary algorithms handle ranges of outcomes well. They adapt to patterns that appear across sets of simulations. And, they can include the kinds of rules of thumb that experts use to produce better solutions. If a rule of thumb works, then the algorithm will try it out and then evolve and refine solutions based on it; if it doesn’t, the algorithm will try other approaches to improving the overall performance of the schedule.

Conclusion

In this article we have described genetic algorithms and the reasons that experts look to them to provide solutions to 21st century problems. We have given examples of their use taken from our own experience (these are the techniques that we have been using to solve supply chain and logistics problems for our clients for more than twenty years.) Finally, we have suggested that their use will spread as they are given the task of creating good plans in the presence of uncertainty. These algorithms will give us the ability to plan better in a world with increasingly complicated problems and increasing uncertainty. wt