|| Checking for direct PDF access through Ovid
There has been a long-running debate over whether humans match or maximize when faced with differentially rewarding options under conditions of uncertainty. While maximizing, that is, consistently choosing the most rewarding option, is theoretically optimal, humans have often been observed to match, that is, allocating choices stochastically in proportion to the underlying reward rates. Previous models assumed matching behavior to arise from biological limitations or heuristic decision strategies; this, however, would stand in curious contrast to the accumulating evidence that humans have sophisticated machinery for tracking environmental statistics. It begs the questions of why the brain would build sophisticated representations of environmental statistics, only then to adopt a heuristic decision policy that fails to take full advantage of that information. Here, we revisit this debate by presenting data from a novel visual search task, which are shown to favor a particular Bayesian inference and decision-making account over other heuristic and normative models. Specifically, while subjects’ first-fixation strategy appears to indicate matching in aggregate data, they actually maximize on a finer, trial-by-trial timescale, based on continuously updated internal beliefs about the spatial distribution of potential target locations. In other words, matching-like stochasticity in human visual search is neither random nor heuristics-based, but attributable specifically to fluctuating beliefs about stimulus statistics. These results not only shed light on the matching versus maximizing debate, but also more broadly on human decision-making strategies under conditions of uncertainty.