Improving Local Search for Resource-Constrained Planning

The need to economize limited resources such as energy, fuel, money, and/or time, is ubiquitous in many of the classical applications of planning. Although planning with resources is a long-standing topic, the progress made in solving such problems has been limited. Relaxation heuristics solve a relaxed planning problem in which all “negative effects” of the operators, such as resource consumption, are ignored. Almost all state of the art heuristic functions for planning are based on such relaxations and therefore completely ignore the need to economize resources. To address this difficulty, one could try to develop better-suited heuristics. We consider the alternative of developing better-suited search methods. Let C denote the ratio between the amount of available resources and the minimum amount required to solve an instance. Planning problems become intuively “harder” as C approaches 1. (Hoffmann et al. 2007) observed that otherwise very successful planners perform very poorly when C is close to 1. The research started in this current work attempts to improve this sad situation. Examples of resource-based planning benchmarks from the International Planning Competitions (IPC) are Mystery and Mprime from IPC’98, which encode transportation with fuel consumption. Trucks in IPC’08 features time as a resource, since actions take time and trucks have strict delivery deadlines. Satellite in IPC’02 features fuel consumption as a resource; however, resources are missing in the STRIPS version of this domain. Several other IPC domains feature resource consumption not as a hard constraint, but only as an optimization criterion, which is ignored by most satisficing planners. Resource constrainedness C is not a controlled quantity in these IPC benchmarks. In most cases, C is not even known. A well-suited banchmark must provide a range of instances with different values of C, while keeping all other settings the same. The main empirical basis of our investigation is the simple transportation domain of Hoffmann et al (2007), called NoMystery here. In NoMystery, a truck moves in a weighted graph; a set of packages must be transported between nodes; actions move along edges, and load/unload