A mapping model of bow effects in absolute identification (Q1916538)

A typical absolute identification experiment involves a set of \(N\) stimuli that vary along some unidimensional psychological continuum. On each trial of the experiment, one of the \(N\) stimuli is presented, and the subject is required to identify it by selecting the previously specified unique ``correct'' response for that stimulus. The possible responses are usually key presses, the keys being associated with the numbers 1 through \(N\), and with the usual order of the numbers being in agreement with the order of the stimuli according to their magnitudes on the psychological dimension. One of the major phenomena in this area, and probably the one that leads to researchers' fascination with it, is that no matter how widely spread the stimuli are on the relevant sensory dimension, people are usually only able to identify the stimuli up to an accuracy that corresponds to correctly identifying about seven stimuli. Although our results in this paper address the above-mentioned issue of the basic limits on absolute identification performance, our major focus is on a second set of phenomena referred to variously as the end-anchor effect, the bow effect, and sometimes the serial position effect. These terms refer to the fact that the stimuli at the end of the presented range in an absolute identification experiment are normally more accurately identified than are the other stimuli in the experiment. The structure of the paper is as follows. In the next section we suggest that the so-called ``encoder problem'' is a useful formal instantiation of a unidimensional absolute identification task. We then present a general solution to the encoder problem using a feed-forward network with one hidden unit and a maximum-selecting output network. We next present two different parametrizations of the general solution, each developed independently by one of the authors. The general solution is then expanded to incorporate noisy and dynamic (real time) processing. Next we report simulation results showing some properties of the model, and the performance of the model is evaluated against data in terms of its predictions for choices made and time to make them. Finally, revisions are made to the model to account for some sequential effects, and we end by suggesting directions for future theoretical work.