Transactional Suppor t for Cooperation in M ultiagent-based Information Systems

Agent technology is a promising solution for coping with the ever-growing complexity of modern information systems involving highly distributed and error-prone information sources. Unfortunately, current prototypes often lack the robustness necessary for real-world deployment. We propose transaction mechanisms as a possible remedy. For this, we divide Multiagent Systems into two layers: (1) a planning layer, whose actions are idempotent, do not cause any real-world side-effects and are therefore easily recoverable, and (2) an execution layer, which is responsible for all real-world actions and runs under transactional protection to achieve robustness. The execution layer consists of special Execution Agents, one for each agent in the planning layer. A Planning Agent develops its local plan and transfers it to its Execution Agent in the form of a transaction tree that models the execution dependencies between agent actions. The Execution Agent executes this tree and provides recovery mechanisms to deal with possible failures according to a predefined contingency behavior within the agent plan. In this paper, we focus on the social aspects of agent planning and illustrate how to map cooperatively developed shared plans with their underlying cooperation patterns to a set of coordinated transaction trees to be executed by the Execution Agents.