Achieving Efficient Cooperation in a Multi-Agent System: the Twin-Base Modeling

The Virtual Secretary 2 project (ViSe2) focuses on the construction of a multi-agent cooperation system. As a research vehicle, we have chosen to build intelligent agents that perform secretarial tasks for their users either by themselves or via cooperation. An individual ViSe2 agent has limited knowledge and problem-solving capabilities. To act better for its user, the agent interacts with other peers to solve problems. In this sense, an agent's ability to reason about the other agents' activities and thus nd the peer becomes a key issue. In this paper, we propose a twin-base (cooperator-base U task-base) modeling for e cient cooperation in a small agent group. The cooperator-base collects stable information of the others and acts as an auxiliary base to the task-base. The task-base provides direct mappings between tasks and relevant expert agents that can perform such tasks. A capability revision process is proposed for keeping the mapping information consistent. With such twin-base modeling, when an agent receives a task that is beyond its capabilities, the agent can directly retrieve the best quali ed peer from its task-base, and ask the peer to perform the task. To test the validation of the twin-base modeling, we have implemented a prototype of ViSe2 multi-agent cooperation system. The experimental results show that the system achieves the anticipated functionality: an individual agent performs the user's task by either retrieving results from its local knowledge base system, or consulting peer agents to take over the job. More precisely, to verify our intuition that the twin-base modeling achieves e cient cooperation, we compare the performance of our model with other cooperation approaches, i.e., the contract net protocol [2], the assisted coordination approach [4], and the acquaintance model approach [14, 7]. Results received so far indicate that our method achieves the most e cient cooperation with high on-line performance.