Distractor location frequencies better account for the instantiation of learned distractor suppression than do reinforcement learning prediction errors

Stimulus-driven attentional capture is reduced when a salient distractor regularly appears at a predictable spatial location (e.g., Wang & Theeuwes, 2018). This phenomenon is consistent with a growing body of work that suggests that selection history plays a powerful role in shaping the instantiation of attentional priority. However, the underlying mechanisms of distractor suppression learning remain poorly understood. In the current study, we fitted behavioral response time (RT) data from a variant of the additional singleton paradigm with a series of computational models to test how individuals harness previous experiences to guide attentional deployment. As in previous studies, we observed robust evidence of learned distractor suppression such that RTs were shorter when the distractor appeared at a high probability location than when it appeared at a low probability location. Furthermore, RTs were also longer when the target appeared at the high probability location relative to the low probability location. Next, we adjudicated whether distractor suppression was best explained by (a) the tracking of distractor location frequencies or (b) a reinforcement-learning (RL) prediction error mechanism. Under the location frequency account, individuals decrease the priority afforded to a particular location with each successive presentation of a distractor at that location. However, while participants did not receive explicit rewards, accurate performance is intrinsically rewarding. The RL account assumes that individuals attempt to maximize performance by increasing or decreasing the priority for a particular location depending on the size and direction of the trial-by-trial difference between expected and observed distractor location likelihoods. We used Hierarchical Bayesian Inference to simultaneously fit and compare models, finding that the distractor location frequency model best accounted for the data. Together, these results suggest that a simple frequency model outperforms models that nudge predictions up and down based on trial-by-trial outcomes.