Swarm intelligence : SI overview – Part 1


The current SI debate suffers from two problems; a lack of clear criteria to recognize SI and confusion over commonalities and differences in human SI and animal SI. Here, we give examples of ways to apply our definition to case studies and we group SI into different categories that make clear where the parallels are between SI in humans and animals. The first category covers cases in which SI is utilized via direct social interactions in both humans and animals; whereas the second category only applies to human cases that involve different forms of electronic information processing for SI purposes. Direct interaction Both animal and human groups can gain SI advantages through direct social interactions for a range of cognitive problems, such as navigating accurately over short or long distances, efficient taxis in a noisy environment or finding a suitable new nest site (Box 1; quorum formation in social insects). Navigational accuracy increases with group size based on the many wrongs principle. The assumption underlying this principle is that all individuals have a common target destination that they want to reach but that each individual navigates with some error. If group members average over each other’s directional preferences, then the error with which the group moves toward the target decreases as a nonlinear function of group size. In this example, the individual errors follow a normal distribution. However, the principle of the many wrongs is not restricted to this type of distribution. As long as the mean of the individual vectors approximates the desired target direction, there are many other types of distribution that could produce a similar outcome (i.e. reducing navigational error with group size). Likewise, there is no particular reason why an averaging principle should necessarily be used and other ways of processing the information held by individuals are possible.

Evolutionary perspective
Traditionally, biological texts compile long lists of diverse benefits that group living enables. Rather than listing effects in this way, individual-based modelling has opened up the possibility of an agent-based approach, in which each additional behaviour or capability that we allow the agents facilitates the emergence of new system properties. This view is different and interesting because the additional capabilities at the individual level could be seen as mutations that then open up a new set of social interactions and subsequent fitness gains to groupliving animals. Let us consider several such capabilities to understand the mechanism by which the systems properties of groups could arise:
(i) In their simplest form, group-living agents are attracted by conspecifics and are repulsed at close range. In addition, they might respond to the body orientation of near neighbours by alignment. Implicit in the above rules is that each agent can independently gather information and also copy the behaviour of others. These simple rules alone would provide the basis for the encounter-dilution effect, the many eyes effect and the confusion effect. This means that the agents can benefit from others detecting a predator, food and so on, even if they do not know which cue other group members respond to at any moment in time. Simple copying of local neighbours already opens up a range of potential collective properties for the group. Interestingly, none of this requires any form of social learning or signalling: simply a follow-response to others in their vicinity. Agents with these properties should be capable of some forms of SI, such as quorum decision-making. It might also result in improvement in navigational abilities and efficient taxis in noisy environments.
(ii) If we allow the agents to store and recall information (i.e. to learn), then they could recognise other agents and remember outcomes of social interactions. This ability could lead to cooperative interaction networks and potentially enable collective solutions to challenging problems to emerge.
(iii) If we bring signalling into play, then it becomes possible for agents to inform others (e.g. about food locations) without physically guiding them (e.g. the waggle-dance in honey bees). Signalling in combination with learning is an effective way of widely distributing information within a group.
(iv) So far, we have assumed that all individuals carry out identical behaviours and that there are no differences in social roles. Furthermore, we could introduce division of labour, but assume that every individual is still capable of adopting any task or role, and see what additional properties emerge. For all of the above four categories, we could explore the importance of group size. How does the performance of a group of a given size compare to that of an average single animal? Or how does the performance of a group compare to a high-ability performer in the population (provided that individuals can be ranked in terms of their individual cognitive abilities)? If the collective behaviour can be measured as the time that it takes to solve a problem or the accuracy of a solution, we could, for instance, look for non-linear relationships between group size and problemsolving time. However, this criterion alone is probably not sufficient to indicate SI.