Edmund H. Durfee, Thomas A. Montgomery
We describe how a behavior hierarchy can be used in a protocol that allows AI agents to discover and resolve interactions flexibly. Agents that initially do not know with whom they might interact use this hierarchy to exchange abstractions of their anticipated behaviors. By comparing behaviors, agents iteratively investigate interactions through more focused exchanges of successively detailed information. They can also modify their behaviors along different dimensions to either avoid conflicts or promote cooperation. We explain why our protocol gives agents a richer language for coordination than they get through exchanging plans or goals, and we use a prototype implementation to illustrate our protocol. We argue that our hierarchical protocol for coordinating behaviors provides a powerful representation for negotiation and can act as a common foundation for integrating theories about plans and organizations.