A queueing model of visual search

Understanding how attentional resources are deployed in visual processing is a fundamental and highly debated topic. As an alternative to theoretical models of visual search that propose sequences of separate serial or parallel stages of processing, we suggest a queueing processing structure that entails a serial transition between parallel processing stages. We develop a continuous-time queueing model for standard visual search tasks to formalize and implement this notion. Specified as a finite-time, single-line, multiserver queueing system, the model accounts for both accuracy and response time (RT) data in visual search on a distributional level. It assumes two stages of processing. Visual stimuli first go through a massively parallel preattentive stage of feature encoding. They wait if necessary and then enter a limited-capacity attentive stage serially where multiple processing channels (“servers”) integrate features of several stimuli in parallel. A core feature of our model is the serial transition from the unlimited-capacity preattentive processing stage to the limited-capacity attentive processing stage. It enables asynchronous attentive processing of multiple stimuli in parallel and is more efficient than a simple chain of two successive, strictly parallel processing stages. The model accounts for response errors by means of two underlying mechanisms, namely, imperfect processing of the servers and, in addition, incomplete search adopted by the observer to maximize search efficiency under an accuracy constraint. For statistical inference, we develop a Monte-Carlo-based parameter estimation procedure, using maximum likelihood (ML) estimation for accuracy-related parameters and minimum distance (MD) estimation for RT-related parameters. We fit the model to two large empirical data sets from two types of visual search tasks. The model captures the accuracy rates almost perfectly and the observed RT distributions quite well, indicating a high explanatory power. The number of independent parallel processing channels that explain both data sets best was five. We also perform a Monte-Carlo model uncertainty analysis and show that the model with the correct number of parallel channels is selected for more than 90% of the simulated samples.