Temporal information and adaptive rationality.

Abstract

This project investigates the role that temporal structure plays in a number of well-studied psychological tasks, with particular attention to recency effects. Recency effects are defined as increased memory for recent information, as well as preferential reliance on such information in determining responses. Such behavior is often interpreted as suboptimal, indicating limitations of human cognitive architecture. However, recent research suggests that recency effects reflect an adaptation to the temporal structure of our cognitive environment, and furthermore one which is adjustable (J. R. Anderson & Schooler, 2000; R. B. Anderson et al., 1997). I elaborate on this perspective through studies of verbal working memory, category learning, and game theory. First, a computational model of working memory, based on recurrent neural networks, is presented which exhibits transient storage of information, with a characteristic decay time that adapts to the task environment. This model is extended to a model of verbal serial recall that can explain all of the most robust phenomena associated with the task. Next, data are presented from an empirical investigation of repeated cued classification, in which the serial correlation among category outcomes is manipulated. The results show that recency effects in this task are adaptable as well, and a modification to existing categorization models is proposed to account for this adaptation. Finally, recency effects in simple social interactions are investigated using the framework of iterated matrix games. It is proven analytically that under reasonable conditions it is individually rational for all players to act with an arbitrarily short memory span. This result is applied to the Iterated Prisoner's Dilemma to derive a class of rational yet cooperative policies for that game. Together, these three investigations show how choosing actions based primarily on recent events can be rational in certain situations, lead to proposals of specific psychological mechanisms for online learning and adjusting of the magnitude of recency effects, and demonstrate that inclusion of these mechanisms within models of choice and memory performance can provide unified explanations of adaptive recency effects with other established phenomena.