Insects that live in colonies—ants, bees, wasps, termites—have long fascinated everyone from naturalists to artists. Maurice Maeterlinck, the Belgian poet, once wrote, “What is it that governs here? What is it that issues orders, foresees the future, elaborates plans and preserves equilibrium?” These, indeed, are puzzling questions.
Each insect in a colony seems to have its own agenda, and yet the group as a whole appears to be highly organized. Apparently the seamless integration of all individual activities does not require any supervision. In fact, scientists who study the behavior of social insects have found that cooperation at the colony level is largely self-organized: in numerous situations the coordination arises from interactions among individuals. Although these interactions might be simple (one ant merely following the trail left by another), together they can solve difficult problems (finding the shortest route among countless possible paths to a food source). This collective behavior that emerges from a group of social insects has been dubbed “swarm intelligence.”
A growing community of researchers has been devising new ways of applying swarm intelligence to diverse tasks. The foraging of ants has led to a novel method for rerouting network traffic in busy telecommunications systems. The cooperative interaction of ants working to build their nests leads to more effective control algorithms for groups of robots. The way in which insects cluster their colony's dead and sort their larvae can aid in analyzing banking data. And the division of labor among honeybees could help streamline assembly lines in factories.Virtual Foraging
ONE OF THE EARLY STUDIES of swarm intelligence investigated the foraging behavior of ants. It had long been known that the ant “highways” often seen in nature (and in people's kitchens) are laid down by individual ants depositing pheromone, a chemical attractant, which increases the probability that other ants will follow the same path to the food source. In the 1990s Jean-Louis Deneubourg of the Free University of Brussels in Belgium, a pioneer in the field, showed that the trail-laying and trail-following behavior of ants was also a good strategy for finding the shortest path between a nest and a food source.
In experiments with the Argentine ant Linepithema humile, Deneubourg and his colleagues constructed a bridge with two branches, one twice as long as the other, that separated a nest from a food source. Within just a few minutes the colony usually selected the shorter branch. Deneubourg found that the ants lay and follow trails of pheromone as they forage. The first ants returning to the nest from the food source are those that have taken the shorter path in both directions, from the nest to the food and back. Because this route is the first to be doubly marked with pheromone, nestmates are attracted to it.
If, however, the shorter branch is presented to the colony after the longer branch, the ants will not take it because the longer branch has already been marked with pheromone. But computer scientists can overcome this problem in an artificial system by introducing pheromone decay: when the chemical evaporates quickly, longer paths will have trouble maintaining stable pheromone trails. The software ants can then select a shorter branch even if it is discovered belatedly. This property is highly desirable in that it prevents the system from converging on mediocre solutions. (In L. humile, the pheromone concentrations do decay but at a very slow rate.)
In a computer simulation of pheromone evaporation [see middle illustration on page 42], researchers presented identical food sources to an artificial colony at different distances from the nest. At first the virtual ants explored their environment randomly. Then they established trails that connected all of the food sources to the nest. Next they maintained only the trails of the sources closest to the nest, leading to the exploitation of those supplies. With the depletion of that food, the software ants began to raid the farther sources.
Extending this ant model, Marco Dorigo, a computer scientist at the Free University of Brussels, and his colleagues devised a way to solve the famous “traveling salesman problem” [see box on preceding page]. The problem calls for finding the shortest route that goes through a given number of cities exactly once. This test is appealing because it is easy to formulate and yet extremely difficult to solve. It is “NP-complete”: the solution requires a number of computational steps that grows faster than the number of cities raised to any finite power (NP stands for nondeterministic polynomial). For such problems, people usually try to find an answer that is good enough but not necessarily the best (that is, a route that is sufficiently short but perhaps not the shortest). Dorigo showed that he could obtain near-optimal routes by using artificial ants that are tweaked so that the concentration of pheromone they deposit varies with the overall distances they have traveled.
Similar approaches have been successful in a number of other optimization tasks. For instance, artificial ants provide the best solution to the classic quadratic assignment problem, in which the manufacture of a number of goods must be assigned to different factories so as to minimize the total distance over which the items need to be transported between facilities. In a related application, David Gregg of Unilever in the U.K. and Vincent Darley of BiosGroup in Santa Fe, N.M., reported that they developed an ant-based method for decreasing the time it takes to perform a given amount of work in a large Unilever plant. The system must efficiently schedule various storage tanks, chemical mixers, packing lines and other equipment.
In addition to solving optimization problems that are basically static, or nonvarying, antlike agents can also cope with glitches and dynamic environments—for example, a factory where a machine breaks down. By maintaining pheromone trails and continuously exploring new paths, the ants serendipitously set up a backup plan and thus are prepared to respond to changes in their environment. This property, which may explain the ecological success of real ants, is crucial for many applications.
Consider the dynamic unpredictability of a telephone network. A phone call from A to B generally has to go through a number of intermediate nodes, or switching stations, requiring a mechanism to tell the call where it should hop next to establish the A-to-B connection. Obviously the algorithm for this process should avoid congested areas to minimize delays, and backup routes become especially valuable when conditions change dramatically. Bad weather at an airport or a phone-in competition on TV will lead to transient local surges of network traffic, requiring on-the-fly rerouting of calls through less busy parts of the system.
To handle such conditions, Ruud Schoonderwoerd and Janet Bruten, both then at Hewlett-Packard's research laboratories in Bristol, England, and Owen Holland, then at the University of the West of England, invented a routing technique in which antlike agents deposit bits of information, or “virtual pheromone,” at the network nodes to reinforce paths through uncongested areas. Meanwhile an evaporation mechanism adjusts the node information to disfavor paths that go through busy areas.
Specifically, each node keeps a routing table that tells phone calls where to go next depending on their destinations. Antlike agents continually adjust the table entries, or scores, to reflect the current network conditions. If an agent experiences a long delay because it went through a highly congested portion of the network, it will add just a tiny amount of “pheromone” to the table entries that would send calls to that overloaded area. In mathematical terms, the scores for the corresponding nodes would be increased just slightly. On the other hand, if the agent went quickly from one node to another, it would reinforce the use of that path by leaving a lot of “pheromone”—that is, by increasing the appropriate scores substantially. The calculations are such that even though a busy path may by definition have many agents traveling on it, their cumulative “pheromone” will be less than that of an uncongested path with fewer agents.
The system removes obsolete solutions by applying a mathematical form of evaporation: all of the table entries are decreased regularly by a small amount. This process and the way in which the antlike agents increase the scores are designed to work in tandem so that busy routes experience more evaporation than reinforcement, whereas uncongested routes undergo just the opposite.
Any balance between evaporation and reinforcement can be disrupted easily. When a previously good route becomes congested, agents that follow it are delayed, and evaporation overcomes reinforcement. Soon the route is abandoned, and the agents discover (or rediscover) alternatives and exploit them. The benefits are twofold: when phone calls are rerouted through the better parts of a network, the process not only allows the calls to get through expeditiously but also enables the congested areas to recover from the overload.
Several companies have explored this approach for handling the traffic on their networks. France Télécom and British Telecommunications took an early lead in applying ant-based routing methods to their systems. In the U.S., MCI WorldCom (now part of Verizon) investigated artificial ants not only for managing the company's telephone network but also for other tasks such as customer billing. The ultimate application, though, may be on the Internet, where traffic is particularly unpredictable.
To handle the demanding conditions of the Net, Dorigo and his colleague Gianni Di Caro, now at the Dalle Molle Institute for Artificial Intelligence in Lugano, Switzerland, increased the sophistication of the ant agents by taking into account several other factors, including the overall time it takes information to get from its origin to its destination. (The approach for phone networks considers just the time it takes to go from one node to another, and the traffic in the reverse direction is assumed to be the same.) Simulation results indicate that Dorigo and Di Caro's system outperforms all other routing methods in terms of both maximizing throughput and minimizing delays. In fact, extensive tests suggest that the ant-based method is superior to Open Shortest Path First, the protocol that the Internet currently uses, in which nodes must continually inform one another of the status of the links to which they are connected.A Swarm of Applications
OTHER BEHAVIORS of social insects have inspired a variety of research efforts. Computer scientists are studying insect swarms to devise different techniques for controlling a group of robots. Several applications being investigated are inspired by the coordination of traffic along pheromone trails or the formation of self-assembled chains in ant colonies [see box on pages 44 and 45]. Using such approaches, engineers could design relatively simple and cheap robots that would work together to perform increasingly sophisticated tasks. In another project, a model that was initially introduced to explain how ants cluster their dead and sort their larvae has become the basis of a new approach for analyzing financial data [see box on opposite page]. And research investigating the flexible way in which honeybees assign tasks could lead to a more efficient method for scheduling jobs in a factory [see box at right].
Additional examples abound. Applying knowledge of how wasps construct their nests, Dan Petrovich, then at the Air Force Institute of Technology in Dayton, Ohio, designed a swarm of tiny mobile satellites that would assemble themselves into a larger, predefined structure. H. Van Dyke Parunak of NewVectors in Ann Arbor, Mich., deploys a variety of insectlike software agents to solve manufacturing problems—for example, scheduling a complex network of suppliers to a factory. Paul B. Kantor of Rutgers University developed a swarm-intelligence approach for finding information over the World Wide Web and in other large networks. Web surfers looking for interesting sites can, if they belong to a “colony” of users, access information in the form of digital pheromones (essentially, ratings) left by fellow members in previous searches.
Indeed, the potential of swarm intelligence is enormous. It offers an alternative way of designing systems that have traditionally required centralized control and extensive preprogramming. It instead boasts autonomy and self-sufficiency, relying on direct or indirect interactions among simple individual agents. Such operations could lead to systems that can adapt quickly to rapidly fluctuating conditions.
But the field is in its infancy. Because researchers lack a detailed understanding of the inner workings of insect swarms, identifying the rules by which individuals in those swarms interact has been a huge challenge, and without such information computer scientists have had trouble developing the appropriate software. In addition, although swarm-intelligence approaches have been effective at performing a number of optimization and control tasks, the systems developed have been inherently reactive and lack the necessary overview to solve problems that require in-depth reasoning techniques. Furthermore, one criticism of the field is that the use of autonomous insectlike agents will lead to unpredictable behavior in the computers they inhabit. This characteristic may actually turn out to be a strength, though, in that it could allow such systems to adapt to solve new, unforeseen problems—a flexibility that traditional software typically lacks.
Many futurists predict that chips will soon be embedded into thousands of mundane objects, from envelopes to trash cans to heads of lettuce. Enabling all these pieces of silicon to communicate with one another in a meaningful way will require novel approaches. As high-technology author Kevin Kelly puts it, “Dumb parts, properly connected into a swarm, yield smart results.” The trick, of course, is in the proper connection of all the parts.THE AUTHORS
ERIC BONABEAU and GUY THÉRAULAZ study the behaviors of social insects and the application of those behaviors in the design of artificial “swarm intelligence” systems. Bonabeau is chief scientific officer at Icosystem Corporation in Cambridge, Mass. He received a Ph.D. in theoretical physics and advanced degrees in computer science and applied mathematics from the University of Paris XI (Paris-Sud). Théraulaz is a research director at the Research Center on Animal Cognition of CNRS at Paul Sabatier University in Toulouse, France. He received a Ph.D. in neuroscience and animal behavior from the University of Provence.