Introducing a method for calculating the allocation of attention in a cognitive "two-armed bandit" procedure: Probability matching gives way to maximizing

Abstract

We tested whether principles that describe the allocation of overt behavior, as in choice experiments, also describe the allocation of cognition, as in attention experiments. Our procedure is a cognitive version of the "two-armed bandit choice procedure." The two-armed bandit procedure has been of interest to psychologistsand economists because it tends to support patterns of responding that are suboptimal. Each of two alternatives provides rewards according to fixed probabilities. The optimal solution is to choose the alternative with the higher probability of reward on each trial. However, subjects often allocate responses so that the probability of a response approximates its probability of reward. Although it is this result which has attracted most interest, probability matching is not always observed. As a function of monetary incentives, practice, and individual differences, subjects tend to deviate from probability matching toward exclusive preference, as predicted by maximizing. In our version of the two-armed bandit procedure, the monitor briefly displayed two, small adjacent stimuli that predicted correct responses according to fixed probabilities, as in a two-armed bandit procedure. We show that in this setting, a simple linear equation describes the relationship between attention and correct responses, and that the equation's solution is the allocation of attention between the two stimuli. The calculations showed that attention allocation varied as a function of the degree to which the stimuli predicted correct responses. Linear regression revealed a strong correlation (r = 0.99) between the predictiveness of a stimulus and the probability of attending to it. Nevertheless there were deviations from probability matching, and although small, they were systematic and statistically significant. As in choice studies, attention allocation deviated toward maximizing as a function of practice, feedback, and incentives. Our approach also predicts the frequency of correct guesses and the relationship between attention allocation and response latencies. The results were consistent with these two predictions, the assumptions of the equations used to calculate attention allocation, and recent studies which show that predictiveness and reward are important determinants of attention.