Computational Approaches to Preference Elicitation
2006

Introduction

This is a survey of preference (or utility) elicitation from a computer scientist's perspective. Preference elicitation is viewed as a process of extracting information about user preferences to the extent necessary to make good or even optimal decisions. Devising effective elicitation strategies would facilitate building autonomous agents that can act on behalf of a user.

Artificial intelligence researchers have always been interested in developing intelligent decision aids with applications ranging from critical financial, medical, and logistics domains to low-stakes processes, such as product recommendation or automated software configuration. Decision theory provides solutions given the system dynamics and outcome utilities. However, user utilities are often unknown and vary more widely than decision dynamics. Since obtaining full preferences is usually infeasible, this presents a serious problem to the deployment of intelligent agents that make decisions or recommendations for users with distinct utilities. Therefore, preference elicitation emerges as one of the more important current challenges in artificial intelligence.

In this report, we consider both historical and current approaches to preference elicitation, concentrating on a few key aspects of the problem. We view preference elicitation and decision making as an inseparable sequential process. Although utility functions are hard to assess, partial information of user preferences might suffice to make good or optimal decisions. Preference elicitation can be driven to explore utility regions that are relevant for making decisions. On the other hand, knowledge of system dynamics and action constraints helps avoid eliciting useless utility information.

In the first part of the report, we describe "classical" decision analysis, consider ways to represent uncertainty over utility functions, and summarize various criteria for decision making with partial utility information. If uncertainty is too great to make good decisions, further preference elicitation is needed. One issue that arises in the elicitation process is which query to ask next. Intelligent querying strategies steer the elicitation process according to some agreed criteria and should be viewed as part of the combined preference elicitation and decision making process. We should note that our use of the term "query" also encompasses more implicit interactions with a user such as changing the presentation of a web page and observing the link followed by a user.

The second part of the report surveys research fields where preference elicitation plays a central role. Imprecisely specified multiattribute utility theory (ISMAUT) is one of the earlier attempts to consider decision making under partial preference information in classical decision analysis. Its extensions to engineering design and configuration problems have been influential in spurring recent interest in preference elicitation among artificial intelligence researchers. Conjoint analysis and analytical hierarchy process (AHP) methods were developed largely in isolation in the fields of marketing research and decision analysis; nonetheless, many issues involving preference elicitation are common. We finish by describing some of the recent advances in preference elicitation in AI.

@TechReport{Braziunas-PreferenceElicitation,
  author = 	 {Darius Braziunas},
  title = 	 {Computational Approaches to Preference Elicitation},
  institution =  {Department of Computer Science, University of Toronto},
  year = 	 {2006}
}