Preference Handling for Artificial Intelligence

Article excerpt

Most problems studied in artificial intelligence involve some form of choice. For example, a robot has to choose among alternative plans to reach given goals, a web-based recommender system should choose a configuration that pleases the user, and an automatic translation system has to struggle with the multiple meanings of words. All these problems may have huge spaces of possible decisions that significantly differ in general criteria, such as cost, quality, simplicity, as well as domain-specific criteria. Preferences are a convenient way to compare the options a priori and then to use them to make best choices in a multitude of problems with different decision spaces.

Preferences are thus essential for making choices in a rational (and intelligent) way. Preference models have been necessary in many fields of AI, including multiagent systems, combinatorial auctions, knowledge representation and reasoning, planning, diagnosis, and design. Moreover, preference modeling is central to decision theory, social choice, and game theory, which, more and more, are cross-fertilizing with AI. AI brings new problems to these classic fields and often needs new forms of preference handling beyond classic utility-based models such as graphical and logical preference representations, new forms that can directly be used in preferencebased problem-solving algorithms.

Preference handling has become an intense area of research in AI. The AAAI-07 workshop continued a series of successful workshops on preference handling at AI conferences (AAAI-02, a Dagstuhl-seminar in 2004, IJCAI-05, and ECAI '06). The one-day workshop had a dense program of 14 presentations and a discussion about the role of preference for AI. Furthermore, AAAI-07 included a tutorial, an invited talk, and many technical papers about preferences.

The workshop papers applied preferences to a variety of AI problems such as e-commerce and combinatorial auctions, intelligent assistants, winner determination in majority voting, game theory, search for solving combinatorial problems, configuration, meeting scheduling, peer-to-peer query answering systems, geographic map generation, and conference paper assignment. The last two problems unveiled interesting preference structures. The geographic map generation chooses one among several satellite images for each position in a grid while imposing preferences on adjacent images in addition to preferences on the choice of a single image. …


An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.