Planning Graph Based Heuristics for Automated Planning

One of the most successful algorithms in the last few years for solving classical planning problems is Graphplan [3]. This algorithm can be seen as a disjunctive version of forward state space planners. The algorithm has two interleaved phases: a forward phase where a polynomial-time data structure called ”planning graph” is incrementally extended, and a backward phase where that planning graph is searched for a solution. In this research proposal, we will show that the planning graph data structure extracted from Graphplan is a rich source for extracting effective and admissible heuristics for controlling search under different planning frameworks. Specifically, we have developed a state space planner called AltAlt [36, 43], which uses such heuristics to drive a regression state-search engine. Our technical contributions include (i) a description of a large family of heuristics derived from the planning graph, (ii) methods to control the cost of computing the heuristics and limit the branching factor of the search, (iii) an extension to our base planner AltAlt to generate parallel plans using distance-based heuristics, and (iv) a plan compression algorithm used to improve the quality of the parallel plans generated. We will describe the technical details of AltAlt, as well as its variant AltAlt [41], which generates parallel plans. We will also present an extensive empirical evaluation for both approaches and their heuristics. Our proposed work will address the application of planning graph based heuristics to more complex scenarios. Specifically, some ongoing work will introduce such heuristics to lifted representations to reduce the branching factor of our planners.