|
|
|
|
|
|
|
prediction. They enable a system to make improvements in the absence of overt payoff or detailed information about errors. Whenever a model's prediction fails to match subsequent outcome, there is direct information about the need for improvement. An appropriate credit (blame) assignment algorithm can even determine what part(s) of the model should be revised. This is a tremendous advantage in most real-world situations where the rewards for current action are usually much delayed. Internal models enable improvement in the interim. |
|
|
|
|
|
|
|
|
Though we readily ascribe internal models, cognitive maps, anticipation, and prediction to humans, we rarely think of them as characteristic of other systems. Still, a bacterium moves in the direction of a chemical gradient, implicitly predicting that food lies in that direction. The repertoire of the immune system constitutes its model of its world, including an identity of "self." The butterfly that mimics the foul-tasting monarch butterfly survives because it implicitly forecasts that a certain wing pattern discourages predators. A wolf bases its actions on anticipations generated by a mental map that incorporates landmarks and scents. Because so much of the behavior of a complex adaptive system stems from anticipations based on its internal models, it is important that we understand the way in which such systems build and use internal models. |
|
|
|
|
|
|
|
|
A general theory of complex adaptive systems that addresses these problems will be built, I think, on a framework that centers on three mechanisms: parallelism, competition,and recombination. Parallelism lets the system use individuals (rules, agents) as building blocks, activating sets of individuals to describe and act upon changing situations (as described in the discussion of classifier systems). Competition allows the system to marshal its resources in realistic environments where torrents of mostly irrelevant information deluge the system. Procedures relying on the mechanism of competitioncredit assignment and rule discoveryextract useful, repeatable events from this torrent, incorporating them as new building blocks. Recombination underpins the discovery process, generating plausible new rules from building blocks that form parts of tested rules. It implements the pervasive heuristic that building blocks useful in the past will prove useful in new, similar contexts. Overall, these mechanisms allow a complex adaptive system to respond, instant by instant, to its environment, while improving its performance. In so doing, as with classifier systems, the system balances exploration with exploitation. |
|
|
|
|
|
|
|
|
When these mechanisms are appropriately incorporated in simulations, the systems that result are well founded in computational terms, and they do indeed get better at attaining goals in perpetually novel environments. It should be possible to take a first step toward a general theory of complex adaptive systems by formalizing |
|
|
|
|
|