Deliberate Robbery, or the Calculating Samaritan

In the usual approaches to planning in AI, a planning agent is provided with a description of some state of aaairs, a goal state, and charged with the task of discovering (or performing) some sequence of actions to achieve that goal. Reasoning about goals is non-trivial, because: Goals are context-dependent. When a robot's goal to bring me some cooee cannot be achieved, for example because the cooee machine is broken, then its alternative goal is to get me some tea. In preference-based approaches, the goal to get me some tea if the cooee machine is broken is a preference for tea if I do not get cooee, i.e. a preference of tea without cooee over neither tea nor cooee. Goals can connict. Goals only impose partial preferences, i.e. preferences given some objective and given some context. Objectives can connict and, as a consequence , goals with overlapping contexts can connict. Recently, several more or less ad hoc preference-based logics for goals and desires have been proposed. This raises a demand for criteria to evaluate the properties of the logics, in particular with respect to the dependence on contexts and connicts. In vdTT98] we therefore introduce the deliberating robber, a relative of the gentle murderer of deontic logic. The formalization of both persons is problematic, because they specify diierent degrees of sub-ideal behavior. The deliberating robber is a more informative benchmark problem for preference-based logics than the gentle murder in two respects. First, the deliberating robber is subject to decision-theoretic instead of moral norms. Second, the deliberating robber is described in more detail than the gentle murderer. There is an indeenite number of distinct degrees of robbing (instead of two degrees of killing) and there is a distinction between contexts. Moreover, the number of degrees or contexts may change (instead of keeping it xed). The deliberating robber example is used to analyze the behavior of two preference-based logics proposed in qualitative decision theory by respectively Boutilier and Tan & Pearl, and we show that counterintuitive conclusions follow. Consider the following preferences (for an interpretation see vdTT98]).