How can we trust a policy system to make the best decision?

Policy-based systems are becoming increasingly common; the emerging areas of autonomie [5,6] and on demand [7] computing are accelerating the adoption of such systems. As the requirements on policy-based systems become more complex, traditional approaches to the implementation of such systems, such as relying entirely on simple "if [condition] then [actions]" rules, become insufficient. New approaches to the design and implementation of policy-based systems have emerged, including goal policies[7,11], utility functions [7], data mining, reinforcement learning, and planning. Unfortunately, these new approaches do nothing to reduce administrators' skepticism towards policy-based automation - how is an administrator to believe that a policy-based system will help his systems perform better? Unless policy-based systems are trusted at least as much as traditional systems, it is unlikely that the acceptance of the policy-based systems will increase. In this report, we describe an approach by which a policy-based system can win the trust of its users, and can continuously adjust itself to make better decisions based on the users' preferences. © 2005 IEEE.