DRAFT: Multi-Agent Decision Support Via User-Modeling

Decision-support requires the gathering and presentation of information, but is subject to many kinds of resource restrictions (e.g. cost, length, time). Individual users differ not only in the resources they have available to expend, but also in the priorities they place on different kinds of information. While it is straightforward to represent these differing priorities and related constraints in a user model, using that model to allocate resources for an unseen task across multiple agents in a dynamic environment is not as simple. Before the information gathering process begins, it is not known which agents will be able to usefully participate, or how much utility they will ultimately be able to provide. MADSUM is a distributed adaptive system that uses a negotiation process to solicit and organize agents to produce information, and a presentation assembly process to coherently assemble the information into text for decision support. MADSUM assumes poor predictive models of ultimate information utility and thus requires dynamic organizational management in response to run-time information failures. A user model, including content preferences, deadlines, and length constraints, informs both processes. An evaluation demonstrates that the influence of the user model on content selection and presentation improves system output, and that the organization responds appropriately and predictably in the presence of inevitable information failures.