Why Greatness Cannot Be Planned

Picbreeder is a website where people evolve pictures by clicking whichever offspring they like and leaving the branch for the next person. One image comes to look like an alien face. Branched further by one of the site's own researchers, the alien's eyes become the wheels of a car. Nobody was trying to make a car. Stanley, whose lab built the site, and Lehman then report the harder finding: the skull, another image users had found there, never arrives when the same algorithm is aimed straight at it.

The claim is that objectives are deceptive: the stepping stone to a distant goal usually looks nothing like it, and a search that scores every step by closeness will refuse the step that looks wrong. The demonstration is a robot in a maze, scored only on how far it ends up from the exit in a straight line. It settles into a corridor that dead-ends close to the exit by that measure, with a wall between, and every step back out scores worse. Novelty search, which has no goal and rewards whatever behaves differently from anything tried before, gets through. The historical version is the vacuum tube, built by nobody trying to compute.

They apply it to schools measured by test score, to grant proposals that must state their result in advance, and to careers planned backwards. It lands for someone stuck inside a plan they cannot argue with. The reservation is that the jump from a maze simulator and a picture-breeding site to education policy is made by analogy, the point is restated many times, and collecting stepping stones is a disposition rather than a method. Objectives, the authors concede, work when the goal is close and the route is known.