The CAMERA Project is defined by four key technical insights:
· It is essential to have configurable negotiation mechanisms between agents, which we support through a concept of negotiation management protocols, in order for the behavior of an agent collection to be adaptive to particular problems despite the individual agents potentially participating in a number of different collections;
· Maintaining focused behavior over time in an agent collection requires that we complement those negotiation management protocols with mechanisms that enable the emergent community of agents to detect problems and negotiate self-correcting behaviors;
· This, in turn, requires representing commitment and coordination relations among agents;
· These capabilities are most effective when coupled with additional infrastructure needed for maintaining consistency between agents during arbitrarily long transactions.
To act on these insights, we take a three-pronged approach: (1) produce an architectural specification to guide internal development and facilitate integration with other ANT program projects; (2) provide useful building blocks via reference implementations of key architectural components; and (3) test, evaluate and demonstrate those components in significant applications.
The CAMERA architecture provides services enabling agents to: (1) describe capabilities and register availability to perform services; (2) send, receive, and process requests for services; (3) exchange reports on status, progress and priorities to support coordination; (4) track dependencies and maintain joint commitments through a scheme that allows configurable resource management policies and maintains shared context information; (5) maintain security among the agent collection via a scheme based on mobile objects whereby access rights are embedded as active behavior within objects being exchanged; (6) establish and reconfigure specialized, efficient communication/control paths between agents enabling them to execute negotiated emergent behaviors efficiently; and consequently, (7) negotiate (and renegotiate) emergent organizations of the active members of an agent collection.
CAMERA provides critical infrastructure to make systems of interacting, negotiating component services viable and protect them from going unstable. This requires addressing two critical problems. The first is to represent explicitly commitment, coordination, role and responsibility relations among agents. These enable self-monitoring and reporting services that help the collection determine whether or not it is on track, self-correcting negotiations among agents to put the joint behavior of an agent collection back on track, and mechanisms for managing security and authority in dynamically changing collections. The second problem we address is to track dependencies and maintain consistent assumptions between agents about their tasks and responsibilities. We then can significantly increase the likelihood that the system of agents will function in a coordinated fashion despite the potential for desynchronization due to communication disruptions inherent in a distributed system of autonomous components.
To turn these technical impacts into practical impacts, there will be a running implementation that supports our own claims through application to significant cases selected jointly with DARPA. We are currently focused on multiple applications to Harrier aircraft safety enhancement: via optimized exchange of parts and services, mission-sensitive maintenance management, time-sensitive flight scheduling, and/or intelligent guidance to pilots. This work is being pursued with cooperation and support from the Joint Strike Fighter Program Office, Marine Air Group 13 at Marine Corps Air Station Yuma in Arizona, and a number of commercial organizations.
Our goal is to build incrementally, producing a working framework that can begin to support applications as early as possible, then enhancing the performance and functionality of the framework and the applications over time.
This strategy is reflected in the goals and milestones for each year:
· Year 1: Focus on initial formation of agent collections, application development, and emprical evaluation of negotiation protocols/framework. Agents find each other by broadcasting requests for service and transmitting responses. Negotiation is point-to-point between agents, and is not yet visible to other agents. That visibility in later years will enable other agents to affect the system either through negotiating to become a participant in the collection or by serving monitoring functions that trigger renegotiation among existing participants. Part of the challenge in later years is to provide that visibility without bogging down the system with high overhead for storing and processing information about itself.
· Year 2: Initial support for regrouping within collections and corrective renegotiation of emergent behavior, continued application and empirical evaluation efforts. Processing overhead will be reduced by using agent capability descriptions and recorded commitments among agents to narrowcast service requests, negotiation messages, and progress reports. Initial visibility of these activities to outside agents will be supported recording them in a passive, centralized blackboard-like data structure. Data recorded to support renegotiation will be determined via explicit scheduling of checkpoints for evaluating anticipated possibilities for renegotiation. (This is a temporary expedient to break implementation into simpler problems and allow testing of features before the full implementation is in place; it will be eliminated in Year 3.)
· Year 3: Support for regrouping and self-adaptation through corrective
renegotiation, under full control of the agent system, application
consolidation and continued empirical evaluation. Processing overhead
will be still reduced, but the relatively rigid scheme used in the preceding
year will be replaced by a scheme in which active Status Monitoring Agents, a
specialized kind of ANT, are spawned to represent commitments and dependencies. The Status Monitoring Agents, which rely on
underlying infrastructure for constraint management and dependency tracking,
replace the rigid blackboard system.
Renegotiation will be triggered on an as-needed, when-needed basis. The system will determine dynamically what
information needs to be kept.
Last Modified by Pedro Szekely on Wednesday, October 06, 1999 (contact szekely@isi.edu)