Current trends in AI planning
Résumé
Few planners have addressed the problem of planning actions with context-dependent effect. In this paper, we present a model of actions with context-dependent effects, stemming from a structured representation of world's states based on the notion of state-variables. From a logical point of view, we distinguish between two types of facts in the description of the world in order to define such a model of actions. Some particular facts are directly asserted by actions. Others are derived from general world properties, and represent context-dependent effects. In the second part of this paper, we present a nonlinear planning algorithm, based on an abduction principle, which uses our state and action representation to deal correctly with the presence of context-dependent effects. This algorithm is sound and complete with regard to the state space, producing minimal length solution plans, and it can be seen as a generalization of the now well-known SNLP algorithm.