Searching for Plans with Carefully Designed Probes

Abstract

We define a probe to be a single action sequence computedgreedily from a given state that either terminates in the goalor fails. We show that by designing these probes carefullyusing a number of existing and new polynomial techniquessuch as helpful actions, landmarks, commitments, and con-sistent subgoals, a single probe from the initial state solvesby itself 683 out of 980 problems from previous IPCs, a num-ber that compares well with the 627 problems solved by FFin EHC mode, with similar times and plan lengths. We alsoshow that by launching one probe from each expanded statein a standard greedy best first search informed by the addi-tive heuristic, the number of problems solved jumps to 900(92%), as opposed to FF that solves 827 problems (84%),and LAMA that solves 879 (89%). The success of probessuggests that many domains can be solved easily once a suit-able serialization of the landmarks is found, an observationthat may open new connections between recent work in plan-ning and more classical work concerning goal serializationand problem decomposition in planning and search.

Solving Large-Scale Planning Problems by Decomposition and Macro Generation

“…CAP can be viewed as a method for identifying and connecting a tractable set of waypoints in the search space. This is somewhat related to Probe (Lipovetzky and Geffner 2011), a best first search augmented by "probes" that try to reach the goal from the current state by quickly traversing a sequence of next unachieved landmarks in a "consistent" (Lipovetzky and Geffner 2009) greedy chain. CAP and Probe work at different levels of abstraction: While the land-marks are single propositions (Probe does not handle disjunctive landmarks), CAP identifies serializable subproblems, each of which contains a set of subgoals and requires a significant amount of search by the SubPlanner.…”

“…In principle, such serializable problems should be easy for modern planners equipped with powerful heuristic functions and other search-enhancing techniques. In fact, techniques such as probes (Lipovetzky and Geffner 2011) explicitly target the exploitation of serializable subgoals and have been Copyright c 2015, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved.…”

“…Finally, although some macro-based systems such as Marvin, which generates macros to escape plateaus in the search space, are closely coupled with the search algorithm used by the planner, CAP is implemented as a wrapper for any PDDL (STRIPS + action cost) planner and as we show below, it can be easily combined with various planners, including FF (Hoffmann and Nebel 2001), configurations of Fast Downward (Helmert 2006), and Probe (Lipovetzky and Geffner 2011). CAP can even be used as a wrapper for previous macro-based approaches such as Marvin.…”

“…In our experiments, for example, CAP can solve Woodworking instances with 79 parts with 120% wood -in contrast, the largest IPC2011 instances of Woodworking has 13 parts, and the state-of-the-art Fast Downward planner (LAMA2011 configuration) can solve upto 23 parts (with 120% wood). Even Probe, which can efficiently solve IPC2011 Woodworking instances with only 31 search nodes (Lipovetzky and Geffner 2011), can solve upto 46 parts with 140% wood. The CAP source repository is at https://github.com/guicho271828/CAP .…”

Understanding and Improving Local Exploration for GBFS

“…Greedy Best First Search (GBFS) is the core search engine used in many state-of-the-art satisficing planners (Bonet and Geffner 2001;Helmert 2006;Lipovetzky and Geffner 2011;Xie, Müller, and Holte 2014b). GBFS always expands a node that minimizes a heuristic function h, without considering its g-value.…”