Goal Evaluation Problems and Solutions.pdfVIP

 Goal Evaluation Problems and Solutions Moshe Tennenholtz Industrial Engineering and Management Technion Israel Institute of Technology Haifa Israel Abstract We discuss the problem of estimating the exact agents preferences on outcomes This problem has b een almost neglected in the recent years although its solution is a crucial step in the design of many AI systems We show that a straightforward use of the classical metho ds for determining the exact preferences on outcomes is not satisfactory in a multiagent setting We dene a basic setting where the problem of goal evaluation can b e addressed discuss p ossible solutions and show how the existence of dierent p ersp ectives aects goal evaluation in a nontrivial manner Intro duction The notion of a goalseeking agent is fundamental to AI as well as to other areas such as decision theory and mathematical economics The goal of an agent is in general a preference relation on the p ossible outcomes of her actions Given an accurate evaluation of her goals the agent will try to achieve her goals in an ecient manner Hence in order to act in a satisfactory manner the agent has to take care of two basic tasks  This work was carried out partially when the author was in Stanford University Devise an accurate evaluation of her goals Devise appropriate plans to satisfy her goals These two tasks are b oth imp ortant although much more eort has b een devoted in the recent years to the latter task The eld of planning as well as much of