An ACT-R model of sensemaking in geospatial intelligence tasks

Abstract

We developed an ACT-R model of sensemaking in a geospatial intelligence task based on two widely used learning processes in ACT-R: instance base learning and reinforcement learning. This map-based task requires users to select (make visible) layers that visualize different types of intelligence, and to revise probability estimates about which groups might commit a future bombing attack. The model (a) evaluates layers during the simulation, (b) selects layers based on the evaluation of all layers, and (c) adjusts probability estimates for all groups based on new evidence. The model exhibits layer-selection patterns that are comparable to participants (N = 45) studied on this task and both model and people deviate from a rational model based on greedy maximization of expected information gain. The model also exhibits an anchoring bias in updating belief probabilities based on revealed evidence, which corresponds to the average participant.