Abstract: |
Traditionally, representation has been perceived as a necessityfor producing intelligent behavior. Once the right representationis in place to drive it, behavior unfolds as the system²s dynamicsinteract with what is usually a fixed, structural entity. For manykinds of systems this approach can be successful. However, as aprescription for building increasingly complex adaptive systems,it often fails. An alternative perspective that is underinvestigation in our Starcat project suggests that representation isnot what drives behavior but rather what is left over by thesystem²s dynamics after concepts have been activated andbehavior has emerged.There are numerous examples of patterns emerging fromunderlying dynamics. In an ant colony, for example, stigmergicbehavior arises from the colony²s dynamics; but when viewedfrom outside the system, the pattern reveals the coupling betweencolony behavior and the environment. We could, from thatperspective, consider the pheromone and ant trails as a kind ofrepresentation. Particles in cellular automata offer anotherexample that demonstrates how coupling with an externalenvironment (here via the fitness function) draws particularbehavior out of the system²s dynamics. Again, from an externalperspective, these appear to represent information about theenvironment. Prigogine²s dissipative structures describe a similarphenomenon in physical systems far from equilibrium. Thefamiliar Bénard cells could be said to represent a certain level ofheat flow through viscous oil. The dynamic nature of these cellsfurther highlights the likelihood that emergent representationsare very fluid and likely to change as pressure from theenvironment changes. Representation in natural systems maywell be an emergent phenomenon, a consequence of the system²sdynamics, an echo of the coupling between the system²s behaviorand the pressures from its environment.There is a family of cognitive architectures, related to Mitchelland Hofstadter²s Copycat, which explores these possibilities. TheStarcat project attempts to generalize Copycat, bringing severalapplications under the same design. Starcat is intended toaddress problems in embodied cognition, where the systeminteracts with an environment and must produce behaviorindefinitely, in the face of changing pressures. It is anarchitecture for components that produce and consume codelets.The components swim in a virtual sea of different kinds ofcodelets. The components ignore some codelets and act uponothers, while frequently introducing new ones. Some Starcatcomponents couple to the environment, allowing the supply ofavailable codelets to be regulated externally. Each codelet is ashort-lived agent that may run and then die. Codelets are by theirnature small; and there are many different kinds associated withthe system. The primary job of a codelet is to build up or teardown perceptual structures. So codelet activity leaves an echobehind in the form of transient data structures. These datastructures ³represent´ Starcat²s perceptions and applications ofconcepts.An interesting consequence of Starcat²s emergentrepresentation is that the system²s myriad micro-behaviors drivethe representation rather than, as in traditional systems,knowledge representation driving behavior. Additionally, thecoordinated aggregate behavior typical of complex adaptivesystems·arising from among the multitudes of interacting localagents·is coupled externally with the environment. In this way,viewed from the outside, the building up and tearing down ofmicrostructures looks like intentional representation.Knowledge representation in Starcat does not capture concepts,nor does it simply get in the way as Brooks² has asserted.Representation is what is leftover once concepts have emerged. Itallows the system to be affected by what it is already doing.Once a behavior is done, the representation can erode becausethe representation that had built up to support the recentlycompleted behavior is likely to have parts that are irrelevant tothe next behavior. New representation soon builds up as part ofthe next round of behavior, and the cycle continues.The system experiences pressure from outside, and thispressure changes what the system must do to continue tofunction, even though the specifics of those changes are dictatedentirely by its existing internal dynamics. The environmenttriggers behavior, but it does not specify behavior. This is thenotion of autopoeisis. We suspect that at an importantrelationship exists between autopoeisis, autonomy, and the kindof behaviorally coupled emergent representation that we havebeen describing. Our various applications and correspondingexperiments continue to examine if this is the case.We have developed a simulation of ant colony behavior thatuses a degenerate version of the Starcat architecture. Theexperiments reveal just how much the microbehaviors of anemergent system can produce macroscopic features that arecoupled to the environment. We are about to undertake a newexperiment in which the slipnet plays a shaping role for thecolony. This may be the lynchpin that ties the externalapprehension of representational behavior back into the system.We have an application that will use the perception of patterns inmusical input to shape the ongoing structural changes in theslipnet. We also have a set of agents that are each driven by anindividual slipnet but which interact with one another in acollective workspace. Finally, we are in the process ofreimplementing both Copycat and one of its successors, Madcat,using the Starcat framework. |
