Extending AI Planning to Solve more Realistic Problems

Abstract

Applying AI planning to solve real-world problems is still difficult despite many attempts done in this area. Classical planning systems are able to handle a limited number of symbolic data elements, without taking into account the numerical aspect of many realworld problems. Some recent planners have moved to solve more realistic problems involving resource consumptions and time management (Bresina et al., 2002; Bacchus & Ady, 2001; Edelkamp, 2002), but the most of these planners deal with the numerical side of the planning problem as an assisting feature to a main symbolic problem, without being able to tightly mix the two sides of the problem to be solved as one homogeneous problem. However, there are still many real-world problems that involve dominant even absolute numerical processing (Zalaket & Camilleri, 2004b; Hoffmann et al., 2007) and for which planning problem representation and data type handling are to be extended. In order to cover this latter type of problems, we propose many extensions that allow the application of the planning process evenly over symbolic and numerical data that can constitute any realistic planning problem. Despite the multiple extensions that have been made to the Planning Domain Definition Language (PDDL) (Ghallab et al., 1998; Fox & Long, 2002; Gerevini & Long, 2005), this language is still insufficient for representing real-world problems that need complex action representation and complex state transformation expression. This lack motivates the organizers of the sixth international planning competition (IPC-6) to request a new extension to PDDL (the PDDL3.1 version). Inspiring from the continuous extensions to PDDL, we propose our first extension that concerns the data representation, in which we introduce the concept of using non-invertible functions to update the numerical and nonnumerical data throughout a planning process. This type of functions allows the integration and the handling in an easy way of uncertainty as well as of temporal and numerical knowledge into planning. As non-invertible functions can be only applied in forward traversal in a search space, hence we focus on the adaptation of forward planning systems to support the application of this kind of functions. We show that our technique can be used by any forward planner with a minor expansion, and we detail the extension of the Graphplan (Blum & Furst, 1995) structure and algorithm to support the execution of functions. The advantage of the