The value of information in automated negotiation: A decision model for eliciting user preferences

Consider an agent that can autonomously negotiate and coordinate with others in our stead, to reach outcomes and agreements in our interest. Such automated negotiation agents are already common practice in areas such as high frequency trading, and are now finding applications in do-mains closer to home, which involve not only mere financial optimizations but balanced trade-offs between multiple issues, such as cost and convenience. As a simple example, a smart thermostat controlling a heat pump could provide demand response to the electricity grid if the inconvenience is offset by the grid relieve incentives. In such situations, the agent represents a user with individual and a priori unknown preferences, which are costly to elicit due to the user bother this incurs. Therefore, the agent needs to strike a balance between increasing the user model accuracy and the inconvenience caused by interacting with the user. To do so, we require a tractable metric for the value of information in an ensuing negotiation, which until now has not been available. In this paper, we propose a decision model that finds the point of diminishing returns for improving the model of user preferences with costly queries. We present a reasoning framework to derive this metric, and show a myopi- cally optimal and tractable stopping criterion for querying the user before a fixed number of negotiation rounds. Our method provides an extensible basis for interactive negotiation agents to evaluate which questions are worth posing given the marginal utility expected to arise from more accurate beliefs.