NONCONSERVATIVE PROBABILISTIC INFORMATION PROCESSING SYSTEMS

Abstract
The report is concerned with two large-scale simulation experiments on probabilistic information processing (PIP) systems. One, a very large and prolonged study of four systems, yielded the conclusion that PIP is indeed an efficient philosophy for information-processing systems--at least twice as efficient as its next-best competitor, and four times as efficient as a representative of current processing techniques. The second PIP experiment was concerned with whether likelihood estimators in PIPs should be allowed to know the state of system opinion; the data confirm the suggestion that it might be undesirable. These experiments required the use of an on-line computer system. This comparison of PIP and its competitors clearly indicates that PIP is superior, but does not indicate how PIP compares with theoretically optimal performance since no objective model of the data-generating process was available. A smaller-scale laboratory experiment is reported that compares PIP with a posterior-odds estimation system (POP) in a task sufficiently complex to be difficult for subjects and yet allowing an objective standard of correct performance. PIP was far superior to POP. PIP and calculations of optimal performance were roughly comparable, with PIP sometimes more extreme than optimal performance and sometimes less extreme. Another small laboratory study, concerned with the development of a response mode in which subjects report on probabilities by making choices among bets, is reported. Its original purpose was to develop a response mode for one group in the first PIP experiment, but it proved to be considerably more important than that.

This publication has 0 references indexed in Scilit: