We developed a language called AgenTalk for describing coordination protocol for multiagent systems. Many coordination protocols such as the contract net protocol have been proposed, and as many software agents have started to be built, many application specific protocols are expected to be required. However, no general framework has been proposed for defining and implementing such coordination protocols. AgenTalk allows the incremental definitions and customization of protocols by incorporating the inheritance mechanism. This paper describes the design policies of AgenTalk and shows how the contract net protocol can be described in AgenTalk. In addition, the multistage negotiation protocol is also described in AgenTalk as an extension of the contract net protocol using the inheritance mechanism. The comparison with KQML, a language and protocol for exchanging information and knowledge between agents, and relationships with other research in distributed artificial intelligence are also discussed.
Figure 15: AgenTalk and Socia
To support desktop meetings that use workstations connected to a high-speed network, we adopt software agents that negotiate on their users' behalf to arrange a meeting at a time convenient for all attendees. Our meeting system Socia was implemented with desktop vision agents, which employ the following facilities: Non-committed meeting scheduling, wherein the agents do not commit to any plan, and thus do not govern their users' schedule; Human behavior recognition, wherein the agents can discern if a human is in front of the computer and recognize their behavior, such as presence, absence etc.
Each vision agent consists of several processes, such as schedule management, conferencing tool management, and face recognition, and adaptively controls the scheduling process. Users can hold meetings using multimedia-conferencing tools, nv and vat, in this system.
The Contract Net Protocol (CNP) assigns a subtask to agents which are involved in multiagent problem solving. Although the logical aspects of the negotiation protocol have been analyzed, the properties of protocol dynamics remain unclear. This paper introduces our quantitative analysis of protocol dynamics which is essential for constructing continuous realtime applications.
In this work, a simulation independent of any application has been carried out in order to analyze the dynamic properties of the CNP. We have obtained the following results: the contractor utility increases together with the system load while the manager utility decreases; when the number of agents increases, the contractor utility rises while the manager utility does not change; the uniformity of agents causes the concentration of ``bids'' and ``awards,'' and thus decreases the manager and the contractor utility.
Furthermore, we apply our simulation results to analyze Enterprise, a famous application of the CNP, and point out the problem which might arise in Enterprise.