June 12, 2003

Bounded Rationality

A classical conceptualization of decision making is the following:
  1. A set of goals is identified

  2. All possible parhts to the goals are articulated

  3. The complete portfolio of costs & benefits for all paths is calculated

  4. The path or paths which maiximizes (optimizes) benefit is chosen

This is a rational frame, but it relys on / assumes complete information and large computational capacity. In all but the simplest cases, the completeness of this scheme makes it unimplementable.

In place of this optimizing scheme, James March and Herbert Simon suggest that a better model is the following:
  • Rather than optimal goals, satifactory conditions are set.

  • Familiar modes of action are tried first, without global cost / benefit analysis

  • To the extent that costs and benefits are considered, they are evaluated relative to local conditions using local knowledge

  • When a satisfactory state is achieved, action stops; there is no attempt to maximize utility, only to attains some acceptable level.

In this frame actions do not have complete information and different actors have different (and not necessarily consistent) information. Achieving a satifactory state can be accopmplished in two ways. First the activities and decisions that are undertaken can advance the system to a new and satisfactory state. Second, if progress toward a satisfactory conditions is too difficult or slow, the standards themselves may be adjusted to accept achievable states as satisfactory.

Innovation can be aded to this program by constraining when statisficing conditions can be changed and including exploratory behaviors in the portfolio of "familiar actions."

This sounds a lot like genetic algorithms - stand by while I explore that further.

Complex behaviors can be achieved without global knowledge at any single points if satificing levels, action programs and problems are parsed (broken into chunks) in the right way. Of course there is no prescription for how to find "the right way" and in a satisficing world, there are likely to be more than one. The details of the "complex behavior" will of course depend on which "right way" is found and operationalized.