Planning with Preferences

ularly as they relate to deciding how to act. As such, it comes as no surprise that preferences also play a significant role in AI automated planning, providing a means of specifying those properties of a plan that distinguish it as high quality. Given some task to be achieved, users may have preferences over what goals to achieve and under what circumstances. They may also have preferences over how goals are achieved— properties of the world that are to be achieved, maintained, or avoided during plan execution, and adherence to a particular way of doing some or all of the tasks at hand. Interestingly, with the exception of Markov decision processes (MDPs), nontrivial user preferences have only recently been integrated into AI automated planning. In classical planning, the oldest and best-known planning paradigm, we are given a description of a dynamic domain, an initial state, and a description of a goal. The problem is to find a course of action (that is, a plan) that, when performed in the initial state, leads to a state where the goal is achieved. In a typical planning problem, there may be many plans that achieve the goal, but the problem specification provides no additional criteria to distinguish between good and bad plans. Planners are often constructed to find shortest plans first. This has been extended to minimal cost plans where the cost of a plan is the sum of the cost of its constituent actions. Preference-based planning (PBP) (see for example, Son and Pontelli [2006]; Bienvenu, Fritz, and McIlraith [2006]) is an extension to the wellknown classical planning problem. In PBP we are provided with a criterion to determine when a plan is preferred to another. To this end, we normally consider the relative merit of properties that desirable plans would satisfy. By way of illustration, consider a logistic planning domain in which packages can be transported between cities using trucks. In the classical planning setting, one is typically interested in finding a sequence of actions that results in a state were all packages are located at Articles