The Center for Research on Concepts and Cognition
The Parallel Terraced Scan
The parallel terraced scan is the key characteristic of the FARG architecture. It is a nondeterministic, parallel (or at least simulated parallel) architecture in which the actions of many small agents holistically result in the attainment of a high-level goal. The cognitive process modeled may be perception, categorization, the making of analogies -- the underlying architecture used in FARG models is the same. The parallel terraced scan makes it possible to define a great number of microstrategies towards the goal, and actual paths to the goal are explored at different speeds according to their estimated promise.
A good metaphor for this architecture (and the metaphor used in Hofstadter's earlier writings about the parallel terraced scan) is the similar action of enzymes on proteins in the cytoplasm of a (biological) cell. Each enzyme is as stupid as can be, and in no way can it be said to have a picture of the ultimate end of whatever metabolic process it aids (and it may participate in many processes at once) -- and in fact the actions of enzymes in the cytoplasm are a tearing chaos, with proteins being built up and broken down at random. However, all the enzymes working together, along with regulatory mechanisms consisting of yet more enzymes, end up as coherent (-seeming) high-level processes, and the whole thing works beautifully.
Another useful metaphor is the action of ants or termites. Each insect performs an insignificant amount of the work of (say) building a nest, but the end result can be a surprisingly regular, sophisticated structure. Especially the nests of termites, which even have mechanisms for regulation of interior temperature, the effect is of some crafty intelligence, the seat of which is obviously not in any one termite.
The FARG models, then, use this sort of processing to get their jobs done. The structures they construct are stored in the Workspace (in earlier models, this was often called the Cytoplasm), which corresponds to human short-term memory. The insignificantly small agents are called codelets, each of which performs a small task, say, the comparison of two items in memory and the subsequent suggestion of a link between them explicitly proclaiming their simlarity (or difference). Codelets have the ability to propose new codelets to continue a promising pathway, and the likelihood of any one codelet getting processor time depends on its relationship to concepts marked as "interesting." The various FARG projects approach this in different ways.
When watching Copycat run, the order-from-chaos aspect of the parallel terraced scan is obvious. The graphical presentation of the current state of the Workspace consists of various boxes around, and lines between, different items in the problem being studied. The aspect and weight of these boxes and lines correspond to their importance in Copycat's current view of the world, and there is a constant flurry of activity as these structures are proposed, take on weight, are rejected, are reproposed -- and usually, the magical result of all this chaos is a coherent view of the problem! It is difficult to watch a Copycat run without seeing the program as a (semi-)intelligent being, and rooting for it to see this or that solution to the problem. It resonates with something.
That's the parallel terraced scan in a nutshell. For any amount of further information, elucidation, postulation, cerebration, or flagellation, select any FARG publication and start reading.