Adaptive agent tracking in real-world multiagent domains: a preliminary report

Intelligent interaction in multi-agent domains frequently requires an agent to track other agents' mental states: their current goals, beliefs and intentions. Accuracy in thisagent-trackingtask is critically dependent on the accuracy of the tracker's (tracking agent's) model of the trackee (tracked agent). Unfortunately, in real-world situations, model imperfections arise due to the tracker's resource and information constraints, as well as due to trackees' dynamic behavior modification. While such model imperfections are unavoidable, a tracker must nonetheless attempt to be adaptive in its agent tracking. This article identifies key issues in adaptive agent tracking and presents an approach called DEFT. At its core, DEFT is based on discrimination-based learning. The main idea is to identify the deficiency of a model based on tracking failures, and revise the model by using features that are critical in discriminating successful and failed tracking episodes. Because in real-world situations the set of candidate discriminating features is very large, DEFT relies on knowledge-based focusing to limit the discrimination to those features that it determines were relevant in successful tracking episodes?with an autonomous explanation capability as a major source of this knowledge. This article reports on experiments with an implementation of key aspects of DEFT in a complex synthetic air-to-air combat domain.