Visual foraging: Quitting behavior when searching aerial maps follows the Marginal Value Theorem
61.11, Wednesday, May 15, 8:15 - 9:45 am, Royal Ballroom 1-3
Todd Horowitz1, Kilian Semmelmann2,3, Sage Boettcher3, Jeremy M Wolfe3,4; 1Basic Biobehavioral and Psychological Sciences Branch, National Cancer Institute, 2Department of Psychology, Ludwig-Maximilians-Universität München, 3Visual Attention Laboratory, Brigham and Women's Hospital, 4Department of Ophthalmology, Harvard Medical School
Consider a radiologist searching a mammogram for tumors, a baggage screener searching for weapons, or an intelligence analyst poring over satellite imagery of North Korea. In each of these visual search tasks, each image contains an unknown number of targets and there are many images to get through. How do observers choose when to move to the next image? These tasks can be thought of as foraging tasks, which have been extensively studied in behavioral ecology. According to the Marginal Value Theorem (MVT), an organism leaves a patch when the local rate of acquiring energy drops below the global rate. Here we test whether MVT applies to a novel visual map foraging task. Stimuli were 50 maps sampled from ten major US metropolitan areas. From each metropolitan area, we selected five 3.4 square km images ranging from dense downtowns to sparse rural areas. Observers (N = 29) searched maps for gas stations using a custom interface created via the Google Maps API, allowing them to zoom in and out as well as pan around map images. Observers were given 50 minutes to find as many gas stations as possible. Our primary finding was that observers patch-leaving (i.e., map-leaving) strategy was well-described by MVT. The instantaneous reward rate for the last gas station located on a map was approximately the average global reward rate for that observer. Second, we found that observers were more likely to miss gas stations in maps that had more of them, independent of literal "crowding" by urban density. This may reflect "satisfaction of search". Third, we identified three within-patch search strategies: "knowledge-based" (e.g., searching along major thoroughfares); systematic "scanning" (e.g., searching along a grid); and "random sampling". Knowledge-based strategies were most successful. Interestingly, use of the knowledge-based strategy declined with age.