Computational Decision Support: Regret-based Models for Optimization and Preference Elicitation Craig Boutilier 1 Introduction Decision making is a fundamental human, organizational, and societal activity, involving several key (and sometimes implicit) steps: the formulation of a set of options or decisions; information gath- ering to help assess the outcomes of these decisions and their likelihood; some assessment of the relative utility or desirability of the possible outcomes; and an assessment of the tradeoffs involved to determine an appropriate course of action. Advances in information and communication technol- ogy have changed the nature of the decisions that face individuals, professionals, and organizations. Sophisticated devices, ubiquitous sensors, and the increasing use of online mechanisms to mediate personal, professional, and business communications and transactions has resulted in an explosion in the data available to support decision making. Apart from data proliferation, the range and complexity of the options facing decision makers has also increased: networked communication has drastically reduced search costs, and sophisticated computational models have greatly expanded the capacity for reasoning about complex decision structures (e.g., policies, configurations of decision variables, etc.). Computer-aided decision support is vital for several reasons. First, computational methods are critical in helping decision makers wade through of huge volumes of data to extract relevant infor- mation. Tremendous strides in information retrieval and machine learning have offered ever-more sophisticated algorithms for data mining and relevance detection. Second, while computers have long been used for decision support, the new richness of data sources and complexity of decision spaces bring with them demands for more sophisticated forms of algorithmic optimization to help decision makers determine suitable courses of action. However, even with sophisticated algorithms at their disposal, decision support tools require considerable information about the decision maker’s prefer- ences to function effectively. This is especially true when we consider putting decision support in the hands of the “masses,” that is, people who make decisions without having (or needing) a rich understanding of the underlying domain or its dynamics. Many ingredients of a decision scenario can be fixed—options, likelihood of outcomes, relevance of information sources, etc., may be the same for each user of a decision support tool. What varies are the preferences of the users, each of whom has different goals or objectives in mind. Hence, this information cannot be coded into the system in advance. Unfortunately, no decision support system can recommend decisions without some idea of what these preferences are, giving rise to the preference bottleneck: how do we get the preferences of the user (or organization) “into” the decision support system? The preference bottleneck is one of the greatest impediments to the wide-scale deployment of 1