Up: Pic1: A Visual Database
Previous: The microfeature extractor
In genetic algorithms (Goldberg, 1989),
one central problem is how to maintain some degree of exploration
of the search space while simultaneously exploiting
information already garnered.
Pic1 has the same problem; it is handled in three ways:
- Converge slowly:
- The system will not converge as a classic -ary search would,
trimming the search space down to th the size with each selection.
First, because the user is not providing exact information,
and second because that would provide too little flexibility to similar
judgement calls, mildly deceptive anomalies, an evolving conception of
the target, or simple user error.
- Preserve some randomness:
- In addition to the n images provided by the chooser,
Pic1 presents a few extra images
that are either totally random or exemplars of
classes of images not currently being seriously considered by the guesser
(that is, not having been selected so far).
- Provide an explicit escape:
- If the user sees that the system is not getting closer to the target,
he should be able to either backtrack or explicitly broaden the range.
Gregory J. E. Rawlins