As automated decision support becomes increasingly accessible in a wide variety of AI applications, addressing the preference bottleneck is vital. Specifically, since the ability to make reasonable decisions on behalf of a user depends on that user's preferences over outcomes in the domain in question, AI systems must assess or estimate these preferences before making decisions. Designing effective preference assessment techniques to incorporate such user-specific considerations (that is, breaking the preference bottleneck) is one of the most important problems facing AI.
In this brief survey, we focus on explicit elicitation techniques where a system actively queries a user to glean relevant preferences. (1) Preference elicitation is difficult for two main reasons. First, many decision problems have exponentially sized outcome spaces, defined by the possible values of outcome attributes. As an illustrative example, consider sophisticated flight selection: possible outcomes are defined by attributes such as trip cost, departure time, return time, airline, number of connections, flight length, baggage weight limit, flight class, (the possibility of) lost luggage, flight delays, and other stochastic outcomes. An ideal decision support system should be able to use, for example, precise flight delay statistics and incorporate a user's relative tolerance for delays in making recommendations. Representing and eliciting preferences for all outcomes in a case like this is infeasible given the size of the outcome space. A second difficulty arises due to the fact that quantitative strength of preferences, or utility, is needed to trade off, for instance, the odds of flight delays with other attributes. Unfortunately, people are notoriously inept at quantifying their preferences with any degree of precision, adding to the challenges facing automated utility elicitation.
Within AI, decision analysis, operations research, marketing, and other areas of research, a number of elicitation techniques have been developed that attempt to address these problems. In this article, we survey a selection of these techniques with a specific focus on two key themes. First, we focus on the use of factored utility models (Fishburn 1967; Keeney and Raiffa 1976): these models decompose a utility function over outcomes (all combinations of attributes) into local utility functions over subsets of attributes, which are then combined to produce a "global" utility function. This breaks the combinatorial explosion in representation size and can dramatically reduce the number of preference parameters that need to be assessed. Widely used additive models are a prime example of this.
Second, we deal with methods that provide an explicit representation of utility function uncertainty. An important trend in preference elicitation, especially within AI, is the recognition of the tradeoff between obtaining more utility information--and thus making a better decision--and the cost of further elicitation effort. Elicitation costs can be cognitive (human effort in answering questions), computational (for example, calculating a value of certain alternative by performing intensive optimization or simulation), financial (for example, hiring a team of experts to analyze potential business strategies), or involve time and opportunity costs. Often elicitation costs will outweigh the "value" of the information it provides. In such a case, decisions should be made with partial utility information. For instance, suppose a travel-planning agent has narrowed the list of potentially optimal flights (measured by, say, willingness to pay) to Flight A, whose utility is between $800 and $1000, and Flight B, whose utility is between $980 and $1200. Further utility elicitation could determine that A is optimal, but at best it can be only $20 "better" than B (whereas B may be up to $400 better than A). If the cost (time, cognitive, annoyance, and so on) of differentiating these two flights further is greater than $20, it is best to simply recommend Flight B. …