Signal detection by human observers: A cutoff reinforcement learning model of categorization decisions under uncertainty.

Abstract

Previous experimental examinations of binary categorization decisions have documented robust behavioral regularities that cannot be predicted by signal detection theory (D.M. Green & J.A. Swets, 1966/1988). The present article reviews the known regularities and demonstrates that they can be accounted for by a minimal modification of signal detection theory: the replacement of the "ideal observer" cutoff placement rule with a cutoff reinforcement learning rule. This modification is derived from a cognitive game theoretic analysis (A.E. Roth & I. Erev, 1995). The modified model reproduces all 19 experimental regularities that have been considered. In all cases,it outperforms the original explanations. Some of these previous explanations are based on important concepts such as conservatism, probability matching, and "the gambler's fallacy" that receive new meanings given the current results. Implications for decision-making research and for applications of traditional signal detection theory are discussed.

Decision noise: An explanation for observed violations of signal detection theory

“…We use this model not as a replacement for SDT (and do not create new measures of sensitivity and bias based on it), but as an extension of classic SDT that can illustrate how different sources of noise may affect measurable statistics. This model incorporates confidence ratings and encapsulates aspects of decision uncertainty present in numerous previous models (see, e.g., Busemeyer & Myung, 1992;Erev, 1998;Kac, 1969;Schoeffler, 1965;Treisman & Williams, 1984), but does so at a level that does not incorporate learning and other trial-by-trial dynamics present in many of these previous models. This simplification allows us to evaluate the role of decision noise in general, independent of the specific assumptions of these theories (i.e., learning scheme, response mapping, criterion sampling/drift, etc.).…”

“…Some theorists have suggested that the decision criterion drifts along a sensory continuum from trial to trial, perhaps in response to error feedback (see, e.g., Kac, 1969). Others have suggested that decision criteria are sampled from a distribution on each trial (e.g., Erev, 1998), and still others have suggested that the observer learns a probabilistic function mapping sensory evidence onto the response (e.g., Schoeffler, 1965). Exactly how noise enters into the decision process is not important for our argument; thus, we aspresent substantial challenges for SDT and are not just complications caused by degenerate criterion placement, as was suggested by Treisman.…”

On the generality of optimal versus objective classifier feedback effects on decision criterion learning in perceptual categorization

“…A second alternative explanation is that optimal classifier feedback reduces overreaction to the objective feedback. Because the objective classifier is, by definition, correct on 100% of the trials, and since no decision criterion exists that can achieve this level of performance, in hill-climbing (e.g., Busemeyer & Myung, 1992) and error correction models (e.g., Erev, 1998;Erev, Gopher, Itkin, & Greenshpan, 1995;Roth & Erev, 1995), the errors for objective classifier feedback might be understoodas indicativeof a need to continue adjusting the decision criterion. Although it is unclear why this would always lead to greater conservatism in cutoff placement for objective classifier, relative to optimal classifier, feedback, we tested this notion in two ways.…”

A test of the optimal classifier’s independence assumption in perceptual categorization

“…As suggested by many researchers, suppose that the observer adjusts the decision criterion based (at least in part) on the change in the rate of reward, with larger changes in rate being associated with faster, more nearly optimal, decision criterion learning (e.g., Busemeyer & Myung, 1992;Dusoir, 1980;Erev, 1998;Erev, Gopher, Itkin, & Greenshpan, 1995;Kubovy & Healy, 1977;Roth & Erev, 1995;Thomas, 1975;Thomas & Legge, 1970). To formalize this hypothesis one can construct the objective reward function.…”