Attentional shifting and learning are rational
Pages 1

Attention shifts and learned attention are good for the rapid learning of new associations without damaging previously learned associations. While this accelerates learning, it can also lead to apparently irrational behaviors. The irrationality of intradimensional shift advantage, blocking, and the inverse base rate effect has been described above. There are many other examples. Consider a situation wherein a cue is only imperfectly correlated with an outcome. The extent to which people (and other animals) learn to utilize the cue decreases when other, irrelevant, cues are added (e.g. Castellan, 1973; Wagner, Logan, Haberlandt, & Price, 1968). For an optimal learner, the presence of irrelevant information should not affect the ultimate utilization of relevant information, yet for natural learners it does. Kruschke & Johansen (1999) reviewed a number of related phenomena in probabilistic category learning, and addressed a panoply of irrational behavior with a connectionist model called RASHNL (which stands for Rapid Attention SHifting ‘N’ Learning). The rash shifts of attention facilitate the rational goal of rapid learning, but also lead to over-or under-commitments to various sources of information. Thus, a model that is driven purely by rapid error reduction can generate a number of seemingly irrational behaviors, just like people and many other animals. The pervasiveness of these learning phenomena, across situations and across species, suggests that it would be irrational for connectionist modelers to ignore attentional learning.