Practical reasoning as a generalized decision making problem

LAMSADE : Laboratoire d'analyse et modélisation de systèmes pour l'aide à la décision ; International audience ; Decision making, often viewed as a form of reasoning toward action, has been considered from different points of view. Classical decision theory, as developed by economists, has focused mainly on identifying criteria such as expected utility for comparing different alternatives. The inputs of this approach are a set of feasible atomic actions, and a function that assesses the value of their consequences when the actions are performed in a given state. One of the main practical limitation of this approach is the fact that it reduces the whole decision problem to the availability of two functions: a probability distribution and a utility function. This is why some researchers in AI have advocated the need for a different approach in which all the aspects that may be involved in a decision problem (such as the desires of an agent, the feasibility of actions, etc) are explicitly represented. Hence, BDI architectures have been developed. They take their inspiration in the work of philosophers who have advocated practical reasoning. Practical reasoning mainly deals with the adoption, filling in, and reconsideration of intentions. However, these approaches suffer from a lack of a clear formulation of decision rules that combine the above qualitative concepts to decide which action to perform. In this paper, we argue that practical reasoning is a generalized decision making problem. The basic idea is that instead of comparing atomic actions, one has to compare sets of actions. The preferred set of actions becomes the intentions of the agent. The paper presents a unified setting that benefits from the advantages of the three above-mentioned approaches (classical decision, 15 BDI, practical reasoning). More precisely, we propose a formal framework that takes as input a set of beliefs, a set of conditional desires, and a set of rules stating how desires can be achieved, and returns a consistent subset of desires ...