Pauline M. Berry, Melinda Gervasio, Bart Peintner, Tomás Uribe, Neil Yorke-Smith
Distributed scheduling systems have used a variety of mechanisms to maximize the quality of the global solution. Techniques include negotiation frameworks based on market economies, game theoretic algorithms, and global or shared evaluation functions. Our position is that these techniques do not adequately address the situation where self-interested cooperative agents maintain schedules on behalf of users operating in unbounded environments. Here, satisfaction is more important than optimality, and personalized preferences are paramount to users. In our approach, global utility is secondary; the agent aims to maximize the utility of its user, but folds the positive utility of cooperating with others into that utility. Thus, we treat the utility of others as a single component of a personalized and context-dependent multi-criteria evaluation function, which the agent learns through interactions with its user.