Decision Making in Complex Multiagent Contexts: A Tale of Two Frameworks

82 AI MAGAZINE Akey feature of an autonomous system, here onwards referred to as an agent (Russell and Norvig 2010), is the capability to choose optimally between different lines of action available to it. This capability of normative decision making becomes demanding in diverse informational and physical contexts, which may range from having precise information about all aspects of the problem to partial knowledge about it and multiple interacting agents. For illustration, consider a toy problem involving an autonomous unmanned aerial vehicle (AUAV) tasked with intercepting a fugitive in its theater of surveillance that is divided into a grid of large sectors. Interception requires the AUAV to move to the sector occupied by the fugitive. Consistent sighting reports, which may be noisy, could lead the AUAV to the sector containing the fugitive. However, a false positive intercept in a proximal sector due to the noise in the sightings would cause the alarmed fugitive to flee his true sector and reappear in some other random sector that is further away. Of course, the AUAV’s objective is to succeed in as many such missions as possible over a time period. Clearly, the AUAV faces a tough decision-making problem because of the uncertainty in sightings and the consequences of making a bad decision. Furthermore, presence of another AUAV in the theater — whether the AUAV is helpful or not — could potentially complicate the decision-making problem. If this