Electrical Engineering and Computer Science

AI Seminar & Defense Event

Agent Aware Organizational Design

Jason Sleight

Monday, April 20, 2015
08:30am - 10:30am
3316 EECS

Add to Google Calendar

About the Event

As cooperative multiagent systems (MASs) increase in interconnectivity, complexity, size, and longevity, coordinating the agents' reasoning and behaviors becomes increasingly difficult. One approach to address these issues is to use insights from human organizations to design structures within which the agents can more efficiently reason and interact. Generally speaking, an organization influences each agent such that, by following its respective influences, an agent can make globally-useful local decisions without having to explicitly reason about the complete joint coordination problem. For example, an organizational influence might constrain and/or inform which actions an agent performs. If these influences are well-constructed to be cohesive and correlated across the agents, then each agent is influenced into reasoning about and performing only the actions that are appropriate for its (organizationally-designated) portion of the joint coordination problem. In this dissertation, I develop an agent-driven approach to organizations, wherein the foundation for representing and reasoning about an organization stems from the needs of the agents in the MAS. I create an organizational specification language to express the possible ways in which an organization could influence the agents' decision making processes, and leverage details from those decision processes to establish quantitative, principled metrics for organizational performance based on the expected impact that an organization will have on the agents' reasoning and behaviors. Building upon my agent-driven organizational representations, I identify a strategy for automating the organizational design process~(ODP), wherein my ODP computes statistical estimates of the qualities of optimally coordinated behaviors and then searches through those statistics for patterns of coordinated interactions that it specifies as organizational influences to the MAS. Evaluating my ODP reveals that it can create organizations that both influence the MAS into effective patterns of joint behaviors and also streamline the agents' decision making. Finally, I use my agent-driven approach and ODP implementation to develop mathematical foundations for multiagent metareasoning through organizational design, and abstract organizational influences, which were topics only informally understood and/or utilized in previous research.

Additional Information

Sponsor(s): Professor Edmund Durfee

Open to: Public