In these experiments, Marcia Spetch and I presented university students with a computer betting game. On any one trial, one location was marked by a blue square on a computer monitor (not five, as in the figure). Students were invited to bet 0 to 4 points. A winning location (S+, there was only one for each student) won them four times the points betted. Other locations shown during training (zero to four across experiments), were losing locations (S-s).
After sufficient training, they were occasionally tested in various ways. They were asked to bet on a range of locations, including those used during training. They were shown a location on the monitor and asked if it was the one that won them the most points. And, without anything shown on the screen, they were asked to point to S+ (the most winning location).
We obtained generalisation gradients on betting after training without any S-s.
These were exponential in shape, supporting Shepard's law. See related story on spatial generalisation in bees.
When betting after peak shift training (some S-s),
subjects betted more on the S+ side than on the S- side, a pattern that is called area shift.
When asked whether a location was the most winning one,
subjects under some conditions showed a systematic error (peak shift): they were more likely to name another location than S+ as the most winning one. This happened when they had sufficient betting experience on the S- side but not on the S+ side.
When asked to point to the most winning location, subjects made systematic errors.
In this figure, subjects with vertical training acted as a control group. Their S-s were vertically displaced from S+. The horizontal group had S-s horizontally displaced from S+. While both groups showed some systematic errors in the horizontal direction, the horizontal group showed significantly more. This indicates peak shift induced by the training.
Cheng, K., & Spetch, M.L. (2002). Spatial generalization and peak shift in humans. Learning and Motivation, 33, 358-389.