In 2004, when the Belgian postal department turned to optimization software from the field of logistics, the result was no better. The software produced a set of routes that should have taken the department’s employees 5 percent less time to complete. Unfortunately, Bonabeau says, “the mailmen hated it, and they dragged their feet,” taking 30 percent longer to complete the new routes. “There are failed implementations of optimization software all over the world in postal organizations–because it’s not an optimization problem at all,” he says.
It is impossible, Bonabeau is saying, to design a web of postal routes around a city that perfectly fits the sensibilities of both managers and mailmen. There are only better solutions and worse solutions. It’s even difficult to say what makes one solution better than the next, since “sensibilities” can’t always be systematically predicted or quantified. One postal worker may prefer a longer route to a shorter one, if it goes by the patisserie where the owner greets him every day by name. One worker may prefer to walk downhill all day because of his fallen arches, while another likes a steep uphill grade because he’s training for a triathlon.
No software can take all these things into account. So programmers shouldn’t even try, argues Bonabeau. Instead, they should look to nature, in which random mutation and natural selection produce the features that fit an organism to its environment. Evolutionary algorithms start with a few solutions and a set of constraints, then breed generation after generation of new solutions by recombining selected elements from the previous generations (see “Technology Design or Evolution?” July/August 2006). In order to steer the solutions in a certain direction, the algorithms need only be told which solutions in each generation to recombine. The choices can be made automatically, or they can be made by humans.
And that is where a program like the Hunch Engine comes in. Last year, the French postal agency La Poste hired Icosystem to spend a week combing the kinks out of the route system in a small city two hours north of Paris. On the first day, Bonabeau recounts, Icosystem handed a set of six route options to each of the city’s 40 mail carriers. The carriers handed in their ranked choices the next morning, and these were fed into the Hunch Engine. The program randomly recombined the routes to generate new ones, then picked out the ones with the highest-ranked “parents.” That night, the carriers ranked the new routes, and the cycle continued. After a few days of this, a persistent web of routes began to emerge; the program had homed in on a solution. And when the new routes were put into practice, the letter carriers said they were happier. Or less malheureux, as the case may be.
The astonishing thing, says Bonabeau, is that “the mailmen never say why they like this route or that one.” Bonabeau believes that programs based on the Hunch Engine can help people with a particular class of problems: the kind where they have a hard time expressing exactly what they want, but they know a solution when they see it.