The overall goal of this research is to develop methods for organizational adaptation in artificial agent societies, resulting in short-term and long-term changes to the society's structure that lead to demonstrable performance improvements. Specifically, the project will develop techniques to locally adjust connections between agents and to form long-term stable teams, resulting in responsive, effective agent societies. A closely related educational objective is to develop course materials centered around the organizational learning software to be developed.
The organizational structure of a multi-agent system refers to the nature of the physical or virtual connections among agents, including their communication, familiarity, and trust and reputation relationships. Agents can adapt this organization by modifying connections, by changing their patterns of interaction with other agents, and by establishing authority relationships and subcontracts. Effective organizational adaptation requires the agents to maintain knowledge of the other agents to whom they are connected, including their capabilities, competence, resource capacities, reliability, and trustworthiness. From the system designer's perspective, developing protocols and methods by which agents can adapt their own organization requires an understanding of how organizational change affects the system dynamics at an individual and at a global level. This project will develop a theoretical framework for organizational adaptation in a simulated multi-agent society, implement this framework within an experimental testbed, and use the framework to develop techniques for two forms of organizational learning: local adaptation of network structure and contract-based approaches for forming stable teams and coalitions. These techniques will be applied to several multi-agent applications: multi-robot exploration, distributed vehicle monitoring and tracking, and supply chain management.
Software agents with varying degrees of autonomy are the focus of many current research projects. They are currently used for information gathering, e commerce, virtual entertainment, and mobile robot applications. As intelligent agents become more ubiquitous, it will be of great benefit if the resulting "agent societies" can work effectively to provide value to their users. This research will result in fundamental advances in representations, modeling, and self-organizing environments and protocols for agent societies. A primary educational objective of the work is to distribute software and benchmarks to facilitate education and research on multi-agent organizational adaptation. This distribution will include a suite of "mini-projects" suitable for classroom assignments or independent study research projects. The other educational objectives include outreach to underrepresented students at primarily undergraduate institutions and involvement in mentoring programs for doctoral students in the artificial intelligence community.
The major research activities have been (1) the development of a multi-agent systems testbed, (2) the development of several novel clustering techniques, (3) the development of a framework for the design and control of swarm systems at an abstract level, and (4) the development of trust and reputation models for intelligent agents that can be applied in a variety of contexts. The multi-agent systems testbed enabled the exploration and development of a variety of methods for agents to autonomously self-organize into team structures to support time-varying tasks in complex environments. Our Probabilistic Relational Clustering Framework uses block modularity and our relation selection method to provide clustering methods that are robust to various types of relational data, including low-density data and data in which there are multiple link types. The Agent-Based Framework (AMF) for swarm design uses machine learning to predict swarm-level behavior, given agent-level specifications.This framework has been applied to design swarm models for boids, simple geometric swarms, wireless sensor networks, particle swarm optimization, a fire-fighting domain, an AIDS epedemiological model, and a wolf-sheep predation model. Our probabilistic learning-based models permit agents in a distributed environment to model the integrity and competence of other agents, which decomposes agent behaviors in a way that leads to improved decisions about which agents are most likely to fulfill their stated obligations. This research was later extended to model indirect reputation by modeling the trustworthiness of referring agents. Education activities on the award included (1) the use of the multi-agent system testbed in course curricula, (2) the development of a game-playing web application for student learning, (3) the development of a new upper-level honors seminar on complex systems, and (4) numerous outreach activities.
