Emergent Planning with Philosophical Agents

Future mission plans for the space program call for complex autonomous systems that must be able to perform with high levels of reliability. Features of proposed missions include wide-scale distribution of agents, partial observability, and interaction between large numbers of active platforms. These domain features prove to be difficult for traditional planning algorithms. We are developing a planning architecture, based on the integration of continual and negotiated distributed planning techniques, that is capable of supporting reliable execution within such domains. Agents within our architecture perform utility-based localized planning. Alternative plans are evaluated with respect to an agent’s philosophy, an ordered layering of principles implemented through utility functions. Loose collaborations between agents are formed dynamically during execution. Through agent interactions, an implicit global plan is developed for the mission. This plan is never fully realized or known by the agents, but is rather an emergent property of localized interactions. We illustrate these ideas through a comparison of our proposed architecture to CASPER, a continual planning approach from NASA JPL, in the context of a typical mission