Abstract
Humans constantly adapt to new tasks and new circumstances. Such adaptations include changes in both physical behavior and cognitive state. In each case, adaptation requires more than just sequencing components, but balancing costs associated with each component to develop sequences that are consistent with human capabilities while meeting task demands. For instance, it is well known that many cognitive states (e.g., vigilance) cannot be maintained indefinitely and thus optimizing task performance requires adapting a dynamic system that does not allow all possible sequencies. Here we use advanced cognitive modeling and prediction tools, built using strategies from the generative AI community, to analyze and assess cognitive state adaptations during a task in which 150+ participants (individually and in pairs) must search for, identify, and defuse simulated bombs dispersed in a visually complex environment. The prediction tool first estimates cognitive state directly from physiological data. These estimates are then passed to a generative model that analyzes the sequence and predicts how the state will unfold in the future. The generative tool was trained offline on a corpus of data that does not include the current task or participant pool.
We then analyze data around bomb detection and defusal events. The results show that over the course of the experiment a consistent pattern of cognitive state changes emerge across participants. During the first task run, the emergent pattern is not well-predicted by the generative model. In other words, the bomb response appears more exogenously driven and not a function of ongoing dynamics. However, by the second run, the model’s prediction errors decrease, and the bomb response appears to shift from exogenously to endogenously driven. The results highlight how cognitive state adaptations synchronize to task demands in order to maintain task performance while operating within the bounds of the system’s limited capabilities.