Eliciting Utilities by Refining Theories of Monotonicity and Risk

Interest in such diverse problems as development of useradaptive software and greater involvement of patients in medical treatment decisions has increased interest in development of automated preference elicitation tools. A design challenge of these tools is to elicit reliable information while not overly fatiguing the interviewee. We address this problem by using domain background knowledge in a fle xible manner. In particular, we use knowledge-based artificial neural networks to encode assumptions about a decision maker’ s preferences. The network is then trained using answers to standard gamble type questions. We explore the use of a domain theory encoding simple monotonicity assumptions and another additionally encoding assumptions concerning attitude toward risk. We present empirical results using a data set of real patient preferences showing that learning speed and accuracy increase as more domain knowledge is included in the neural net.