A behavioural analysis of signal-detection performance

Abstract

Choice behaviour of pigeons on a psychophysical task was analysed by applying the generalized matching law to the standard signal-detection theory payoff matrix. This approach views the yes-no detection task as two concurrent schedules each operating under a distinctive stimulus. Differential control by the discriminative stimuli causes biased matching performance in each concurrent schedule. Four experiments are reported, each designed to investigate empirically this matching approach to detection performance. The experimental procedure was an analogue of the discrete-trial yes-no method frequently used in contemporary psychophysical studies. One of two light stimuli of varying intensity or duration was presented, on each trial, on the center key of a standard three-key operant chamber. The animal was trained to report which stimulus was presented by emitting one of two responses. A left-key response was defined as "correct" and reinforced if one stimulus (S1) had been presented, and a right-key response was defined as "correct" and reinforced if another stimulus (S2) had been presented. Correct responses produced a 3-sec magazine light accompanied intermittently by wheat presentation. Incorrect responses, defined as right-key responses after S1 presentations and left-key responses after S2 presentations, produced either a 3-sec blackout (Experiments 1-3), or a 3-sec magazine light plus access to wheat (Experiment 4). During blackout, all chamber lights were extinguished and responses were ineffective. Discriminability, or the accuracy with which the animal detected S1 relative to S2, and response bias, or the degree to which the animal emitted one response more frequently than the other, were studied as a function of four main classes of experimental variables: stimulus-presentation probability, relative stimulus value, and reinforcement rate for correct and incorrect choice responses. Discriminability was a function of stimulus values. It was, however, curvilinearly related to the logarithmic ratio of stimulus values, and a more detailed analysis showed that a stimulus-difference model gave a more linear relation. Although discriminability was not a function of the relative reinforcement rate for correct choice responses, conventional detection theory measures of stimulus discriminability decreased as the probability of reinforcement for incorrect responses increased. Response bias was a function of the relative reinforcement rate for both correct and incorrect responses, but was not a function of stimulus-presentation probability or stimulus values. The present research provided empirical support for a matching model of discrete-trial detection performance, but did, however, suggest some necessary modifications. The model was also shown to fit published data on multiple and multiple concurrent free-operant behaviour. Description of detection performance in terms of the matching of response ratios to reinforcement ratios offered both accurate predictions of complex behaviour, and measures of discriminability with wide generality.