The centerpiece of this project is the design, development, and evaluation of computational representations and algorithms for making software agents that are organizationally adept. An organizationally-adept agent is not only aware of its role(s) in an organization, but can also monitor how well it is fulfilling its organizational responsibilities and can proactively adapt its behaviors to meet organizational needs better. Organizationally adept agents evaluate their behaviors based not on their (agent-centric) self-interests but rather on their (organization-centric) responsibilities to each other, and autonomously adapt to achieve organizational objectives emergently.
Elaboration and adaptation by organizationally adept agents means that the ultimate organization design is formed by a combination of top-down design (to produce a "ballpark" organization) and emergent refinement processes. Further, this combination can be iterative and ongoing, where organizationally adept agents can detect tension between top-down and emergent influences, and inform the design processes of runtime interaction patterns and environmental tendencies that suggest useful top-down organization restructurings.
The intellectual problems being pursued are central to practical issues in scaling multi-agent systems to help solve complex, long-term, global problems. Many critical challenges facing society including climate change, health care, and sustainable energy|require a prolonged commitment to monitoring and managing distributed activities. Networked computer systems populated by software agents, which can be constantly measuring, comparing, and interpreting information to understand and respond to wide-scale phenomena, promise to address such challenges, but need the kinds of innovations proposed in this project. Second, while the project is specifically looking at organizations for computational agents, our results will inform, and be informed by, research on human organizations. To stimulate sharing insights and results, the investigators will organize a multi-disciplinary symposium on organization-centric reasoning, and will train an inter-disciplinary cohort of graduate students.
Victor R. Lesser (PI) Daniel D. Corkill (Co-PI) The ability to create and maintain effective multi-agent organizations will play a major role in the development of larger and more complex multi-agent systems. If multi-agent applications are to grow to encompass hundreds or thousands of agents—especially applications in which intelligent software agents are operating with and within human organizations—research efforts must shift from an individual-agent-centric view of coordination and control to an organization-centric one. The project focused on developing organizationally adept software agents (OAAs) that can reorient their local activities based on their interpretation of organizational intent, allowing emergent and adaptive organizational behavior within designed organizations. This research addressed scaling and performance issues involved in constructing multi-agent systems by adding new capabilities to agents so that they can operate effectively in an organizational context by being able to modify supplied organizational guidelines should those guidelines become ineffective. We have been primarily interested in emergent organizational behavior where agents identify interactions and local control decisions that have been effective in the past and give preference to similar decisions in the future. One of the novel ideas we explored is the use of annotated organizational guidelines that provide performance expectations that are used by an OAA to improve its local decision-making and help it detect when its organizational guidelines are no longer appropriate for the current environment. We constructed and evaluated an OAA software architecture that uses organizational guidelines and performance expectations as an integral part of its operation. Our OAA architecture can: 1) operate even without organizational guidelines; 2) adjust its operational decisions to conform with organizational guidelines, if supplied; 3) assess the appropriateness of the organizational guidelines based on deviation from annotations describing the task and environmental assumptions used when the guidelines were designed; 4) stop following guidelines deemed to be inappropriate; and 5) propose and negotiate agreements with other agents to use in place of inappropriate guidelines. Using this architecture, we showed how an OAA could adjust its operational decisions to conform to organizational guidelines if they are made available, and demonstrated that OAAs can operate without organizational directives. We also showed how an OAA uses expectation belief values contained in guideline annotations to inform its operational decision-making and to identify deviations from the task and environmental assumptions that were used when the guidelines were designed (and to stop following guidelines deemed to be inappropriate). Additionally, we explored the learning of policies that decide when organizational adaptation should be done and how much effort should be invested in adaptation as opposed to continuing with the current organizational directives. We used a decentralized, multi-agent reinforcement-learning algorithm, developed by a graduate student on our project, within each agent to learn agent policies. The agent is also augmented with a heuristic rule-based algorithm that uses information provided by the reinforcement learning algorithm to resolve conflicts among agent policies from a local perspective during both learning and execution stages.
