ABSTRACT
The last chapter spelled out the constraints an associationist model of statistical reasoning has to obey for both theoretical and empirical reasons. Recall that such a model has to rely on the co-occurrence of features of successive events as the sole input and it has to learn without any supervision. Its estimates of conditional probabilities should roughly conform to actual relative frequencies, but the model should produce a regression effect, be responsive to a change in relative frequency over time, become more exact with increasing sample size, and put increasing confidence in judgments as sample size increases. Moreover, it should be able to simulate the results described in this book for urn-type tasks, that is, to judge p(A & B) to be at most as likely as p(A) or p(B), and to judge p(A|B) as more probable than p(B\A) if p(A) > p(B) (and the reverse). The PASS model will be shown to achieve all of this.
