Guided self-organization: perception–action loops of embodied systems

Abstract

In general, self-organization is defined as the transition of asystem into an organized form in the absence of external orcentralized control. Thus, one may emphasize two keyfeatures of a self-organized system or process: (i) anincrease in organization (structure and/or functionality)over some time, and (ii) the local interactions are notguided by any external agent. At the first glance, the secondfeature (the absence of external guidance) immediatelyplaces the idea of Guided Self-Organization (GSO) under aserious doubt. The term almost sounds self-contradictory orparadoxical. However, there is a simple resolution of theapparent inconsistency.Let us first illustrate this with an example provided bystudies of optimal path formation within artificial ant col-onies. Optimal paths (sometimes a network such as mini-mal spanning tree; Prokopenko et al. 2005) connecting thenest and some distributed food sources is a well-knownoutcome of self-organization that involves pheromone-depositing ants. Each individual ant uses only local infor-mation, without reference to the global network, and thelatter self-organizes after multiple stigmergic interactionsbetween the ants and the environment. The first feature ofself-organization—the increase in organization overtime—is manifested by such an optimal network (paths).The second feature demanding that the local interactionsare not guided by any centralized control or external agentis also evident: every ant acts independently and locally,without tracing some predesigned blueprint.At this stage, we extend this example by using theresults reported by Van Vorhis Key and Baker (1982) whostudied odor-conditioned anemotaxis exhibited by Argen-tine ant workers, Iridomyrmex humilis. They experimentedwith a specific trail pheromone component that was pre-sented in two ways: as a wide, relatively uniform swath ofpermeated air, and as a point source creating a time-aver-aged plume downwind. One of the main observations wasthat ants traveled significantly farther toward the phero-mone source in wind than without wind. That is, theexternal pressure provided by the additional point source ofthe specific trail pheromone guided the ants in a particularway. Such guidance, however, was not provided as acontrol input to individual ants, i.e., without any modifi-cations of the ants’ neural circuitry. One may thereforeargue that the resulting optimal paths still appeared as anoutcome of self-organization driven by pheromone-depos-iting ants, and not by any specific blueprint—but at thesame time, the paths were affected (guided) toward aspecific goal or task. One may imagine, for instance, thatdeploying other point-sources of the trail pheromonecomponent may guide the resulting optimal paths in vari-ous ways.Crucially, the external pressure provided by the addi-tional point source of the trail pheromone was not appliedvia some explicit change to control logic of the localagents. In the second example, the paths still formedoptimally subject to the additional constraint within theenvironment, while the inner workings of local agents(ants) stayed the same—it is just that the extra pheromoneaffected some of the local decisions by changing the