Edmund Durfee and Partial Global Planning: Coordination Without Omniscience
Edmund Durfee's work on partial global planning showed how distributed problem solvers could coordinate using incomplete, local and sometimes outdated views of one another's plans.
Distributed agents rarely possess a complete global picture
Distributed problem solving creates an uncomfortable fact: each participant sees only part of the system. Communication may be delayed, bandwidth may be limited, and the world can change before everyone agrees on a common plan. Edmund Durfee’s work with Victor Lesser treated those imperfections as normal rather than exceptional.[1]
The resulting framework, partial global planning, was designed so cooperating problem solvers could coordinate from incomplete summaries of one another’s activities.
Partial global plans are deliberately incomplete
Durfee and Lesser’s 1987 work describes agents summarizing local plans, exchanging selected information and constructing models of network activity. From those models they form partial global plans representing how several local efforts contribute to larger goals.[2]
The word ‘partial’ is crucial. A plan did not need to describe every agent or every future action to improve coordination.
A useful model can be stale and still improve behavior
Because communication takes time, an agent’s view of others may be incomplete or out of date. The framework therefore focuses on generating coordination that remains useful even when perfect agreement is impossible.[1]
Selective communication is part of the algorithm
Agents do not simply broadcast everything they know. They exchange summaries relevant to coordination, reducing communication overhead while retaining enough structure to identify relationships among activities.
The framework connected local plans to network goals
A partial global plan contains information about objectives, major activities, interactions and status. Later UMass descriptions explain how agents can identify that separate local goals contribute to a larger network goal and then organize exchanges and scheduling around that relationship.[3]
This was a sophisticated answer to decentralization: preserve local autonomy while creating enough shared structure for coherent group behavior.
The DVMT gave the theory a demanding testbed
Durfee’s research grew within the UMass distributed vehicle monitoring testbed tradition, where multiple processing nodes interpreted geographically distributed sensor data. Such environments made coordination costs visible because information arrived at different places and times.
The 1991 partial global planning paper applied the framework to distributed hypothesis formation and documented its role as a coordination mechanism rather than merely a planning notation.[3]
Coordination is a tradeoff, not a free resource
More communication can improve shared understanding but consume time and bandwidth. More rigid global agreement can improve predictability but reduce responsiveness. Durfee’s work repeatedly examined these tradeoffs instead of assuming that maximal coordination is always best.
Negotiation could be layered onto partial global planning
Durfee and Lesser also explored negotiating task decomposition and allocation through the partial-global-planning framework. The work shows that coordination mechanisms need not be isolated: planning, negotiation and task assignment can be combined when agents discover dependencies among their activities.[4]
Durfee turned the research into a systematic account
Durfee’s 1988 book Coordination of Distributed Problem Solvers presented the subject as a coherent engineering problem. Its chapters move from distributed problem solving and local planning through partial global goals, coordination and experimental evaluation.[5]
That organization is historically important because it shows distributed coordination becoming a design discipline rather than a collection of one-off demonstrations.
Partial global planning anticipated later orchestration problems
Today’s multi-agent systems may use language models, APIs and cloud tools instead of 1980s distributed-sensing nodes, but they face related questions: what does each agent know, what should it communicate, how are dependencies detected, and when is local autonomy preferable to global synchronization?
Partial global planning offers a vocabulary for systems where complete central knowledge is expensive or impossible.
The central insight is bounded coordination
The framework does not seek omniscience. It seeks enough mutual awareness to improve collective behavior. That principle remains relevant in large agent systems, distributed robotics and workflow orchestration.
Why Edmund Durfee belongs in multi-agent history
Durfee helped turn the vague idea of cooperating distributed programs into a concrete theory of how partial knowledge, plan exchange and network-level objectives could be managed. The UMass record shows the work evolving across technical reports, conference papers, journal articles and a dedicated monograph.[1][5]
His contribution is a reminder that coordination is not synonymous with centralization. Intelligent systems can coordinate through carefully chosen partial views.
Works Cited
- 01
- 02
- 03
- 04
- 05Springer — Coordination of Distributed Problem Solvers link.springer.com
CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.
Submit a research lead