Swarm Intelligence

2/16/2010

Swarm intelligence is the term for complex behavior arising from a group of similar individuals

following a set of simple instructions and often sharing local information. Insects like ants and bees are

known to show aspects of swarm intelligence, easily finding the best sources of food, optimal paths

between different areas, and exhibiting social behavior like cooperation. People have also used the

properties of swarm intelligence to solve a vast variety of problems. Examples of such applications

include new ways to organize packing lines, problem solving algorithms and swarm robotics. There

are a few commonly used terms to describe different aspects of a swarm, and what makes it intelligent.

A swarm of individual units, called agents, is said to be intelligent if the swarm is able to

complete it's goals regardless of the situation it finds itself in, within the environment it was intended

for. A top down definition like this amounts to an exploratory approach to the swarm's environment,

where it eventually settles on the best course of action through agent–agent interactions. This contrasts

with the traditional workings of artificial intelligence where classically AI involves an explicit

calculation of all possible choices and paths leading to an optimum course of action, a bottom up

approach. In the course of a swarm operating it is often the case that a few or many individual agents

will be lost or become non-functional, the swarm's ability to complete it's goals without the lost agents

is called robustness. A swarm can be extremely flexible and reallocate agents to compensate for the

lost agents, while any singular intelligent agent that finds itself in a situation with lost functionality

probably will not be able to complete it's task.

The ability of a swarm to change it's collective behavior to suit it's environment is called

flexibility, while an individual agent's ability to change according to its environment is called plasticity.

While the two properties may seem interrelated, they are independent of one another. A swarm's

behavior comes from the interactions between agents, and it's flexibility can come solely from the rules
guiding the agents, without any need for individual plasticity. One crucial concept of swarm

intelligence is emergent behavior, this is group behavior that is not explicitly ordered or programed, but

simply emerges from the group interactions. Another concept critical to swarm intelligence is self

organization, or the ability of a swarm to spontaneously organize itself without any management or a

command system.

There are many examples of swarm intelligence in nature, and swarm intelligence was in fact

inspired by the complex behavior of social insects. Other examples of swarm behavior include

flocking birds and schooling fish. Ant colonies are one very common example of an intelligent swarm.

This is because of their ability to defend their territory, hunt for and retrieve the best food sources,

respond to weather changes, and build complex nests all without any instruction. One tool ants employ

for finding optimum paths is a pheromone trail each ant leaves behind wherever it travels that

evaporates over time. Whenever an ant comes across the pheromone trail of another ant it is more

likely, depending on how strong the trail is, to follow the path of the pheromones. This means that

should an ant find food and travel back to the nest the same way it went to the food, it's path would

have about twice the normal level of pheromones and draw more ants to help gather the food. With

each additional ant more pheromones are laid down leading to a positive feedback loop, and plenty of

ants to carry the food back to the nest quickly.

One other naturally occurring example of swarm intelligence is a flock of birds. A flock of

birds operates with no leader or command structure but is still able to find a roost for the night, and

food in the morning. When flying there are only a few simple rules that the birds tend to follow, which

are to stay at a comfortable distance away from other nearby birds, and to fly roughly with the rest of

the flock. These simple rules however let a majority rule of sorts emerge, for example if most of the

members want to go in a certain direction, or roost in a certain place, the flock will follow.

There are a few algorithms designed for certain problems that are almost exact copies of natural

swarm intelligence. One of these is ant colony optimization (ACO) and another is particle swarm
optimization (PSO). ACO draws its inspiration from the way ants lay and 'read' pheromones to find

optimum paths between two points, so it is used to find routs in problems that can be represented by a

vertex–edge graph. The way it works is by randomly placing virtual ants all across the representative

vertex edge graph, and then having each one pick a random edge to travel across. After each ant has

moved over one edge the pheromone levels of the edges are updated with concentrations by a set

amount of pheromones per ant divided by the distance covered on the edge. Then the random walk

step is repeated followed by another pheromone update, now including a pheromone evaporation step

to let old trails go cold. This process is repeated many times and eventually the ants will, most likely,

all settle on the most efficient path between the starting and finishing points.

Particle swarm optimization takes it's inspiration from a flock of birds, and is used to solve

problems of optimization with respect to multiple continuous variables. Each particle is randomly

initialized in the solution space with a position and a velocity, then a fitness value is calculated telling

how close the particle is to the desired value. There are two optimizing values, one is the best value of

the swarm, and one is the global best value discovered so far. As the swam moves around, each step

updates the individual particle velocities depending on the current global and flock best value. As the

algorithm goes on the swarm will eventually settle on the desired result, or find a local minimum value

that is close to the desired result.

Here is an example of how a small change in rules can have a large change in the emergent

behavior of a group. Imagine you are in a ballroom with around one hundred other people and

everyone decides, at the same time, to play a game governed by one rule that everyone follows. This

rule is to choose two other people in the room at random, one is your aggressor the other your defender,

and you must move so that your defender is between you and your aggressor. After everyone gets tired

of the first game the rules are changed a little, this time the two people chose randomly are an aggressor

and a victim, and you must move to be between the two as the defender. It turns out that the difference

in behavior can be surprising. Following the first rule everyone would have been moving around
seemingly randomly without any order or direction. However, following the second rule everyone

would have converged in to one relatively stationary group, where everyone jostles around for position,

but as a whole do not move very much. With only a slight change in rules the behavior of a group can

change dramatically, this shows just how difficult it might be to devise a set of rules to make a group

perform an arbitrary task like running a corporation, or doing repairs on a submarine.

There are many examples of real life applications of swarm intelligence, two of which are

swarm robotics and running a business. Swarm robotics is a natural extension of swarm intelligence,

however it is not without challenges and is still in it's infancy. Some of the problems to overcome are

with the mechanics of putting together a group of robots that will work together communicateing with

each other and with the cost of implementing such a system. If the robots are not communicating

enough or the rules are not well designed there may be competition arising between the agents instead

of cooperation. However if the challenges can be overcome there are many situations that would

benefit greatly from the application of swarm robotics. The strengths of swarm robotics and swarm

intelligence are well suited to distributed problems or tasks without much in the way of forward

intelligence, areas not well suited to the single robot approach. These properties make swarm robotics

well suited to patrolling, exploring, and searching operations.

The application of swarm intelligence to the operation of large corporations might seem odd at

first but it is a very natural extension and, if done well, is a win–win for the corporation and employees.

The running of a business has underlying principles that are similar to those behind swarm intelligence,

where there are many individuals working towards the common goals of the group. Corporations also

share the robustness property with intelligent swarms in that the success of the individual is not

necessarily tied to the success of the group. One example of swarm intelligence application is at a

book packing warehouse. The warehouse originally had a system that placed one worker at one station,

filling the part of the order from the station and then passing the order down the line. The problems

with this were that the quicker workers were under worked, while the slower workers were often
stressed. These conditions resulted in bottlenecks in the production, and a high stress environment.

The production line rules were changed to just one rule: Do the order you are working on until the

person in front of you takes it over, then move back and take over the work behind you. The line was

restructured to have the slower workers toward the front of the line finishing the orders, and the faster

workers in the back of the line starting the orders. This led to a natural balance in work load where

everyone could work at their own pace, while also reducing the amount of management that was

necessary for the workers, increasing productivity and reducing stress.

A second interesting application on swarm intelligence to a large corporation is in Capital One's

IT division. In 1994 Capital One got a new CIO who implemented four simple rules to help govern the

IT division's employees. The rules were: first, to keep the company's goals in mind when working;

second, spend money like it's your own; third, be flexible or don't box yourself into one mindset;

fourth, empathy or think of others' situations when you are asked to do something you don't want to do.

After a year of reinforcing these values, they had become engrained into the employees habits. The

results were fantastic with increased productivity and an attrition rate of 4%, well below the national

average of 20%.

Even with all of this, swarm intelligence is still a new field with much room for advancement.

Swarm intelligence has a great deal of potential and it offers possible applications ranging from

industry, to military, to the economy and even traffic control systems. Even though it is often very

difficult to design a set of rules for a specific outcome, much research is being done to unravel the

chaotic nature of this phenomenon.

Martinoli, Alcherio. Swarm Intelligence in Autonomous Collective Robotics: From Tools to the
Analysis and Synthesis of Distributed Control Strategies. Doctoral Thesis, Microprocessors and
Interface Laboratory, Sweden 1999

Thampi , Sabu m . Swarm Intelligence. LBSITW

Bonabeau, Eric; Meyer, Christopher. Swarm Intelligence: A Whole New Way to think About Business
Harvard Business Review May 2001 R0105G. C 2001 Harvard Business School Publishing
Corp.