ACT-R models of information foraging in geospatial intelligence tasks

We describe the development of computational cognitive models that predict information selection behavior in simulated geospatial intelligence tasks. These map-based tasks require users to select layers that visualize different types of intelligence, and to revise probability estimates of attack by hypothetical insurgent groups. Our first model has vast amounts of task-specific declarative memory and selects information layers that provide maximum expected information gain. This first model exhibits layer selection sequences that are almost identical to a rational (Bayesian) model, but fails to predict the layer selection sequences of human participants' performing the tasks. Our second model integrates instance-based learning with reinforcement learning and information foraging theory to predict the selection of information layers. The second model replicates the distribution of participants' layer selection sequences well. We conclude with some limitations that our current ACT-R model has and the role of cognitive models in the intelligence analysis tasks.