Picture of a sphere with binary code

Making Strathclyde research discoverable to the world...

The Strathprints institutional repository is a digital archive of University of Strathclyde research outputs. It exposes Strathclyde's world leading Open Access research to many of the world's leading resource discovery tools, and from there onto the screens of researchers around the world.

Explore Strathclyde Open Access research content

Genetically evolved macro-actions in AI planning problems

Newton, M. A. H. and Levine, J. and Fox, M. (2005) Genetically evolved macro-actions in AI planning problems. In: Proceedings of the 24th Workshop of the UK Planning and Scheduling Special Interest Group (PlanSIG 2005). UNSPECIFIED, pp. 163-172.

Full text not available in this repository. (Request a copy from the Strathclyde author)

Abstract

Despite recent progress in planning, many complex domains and even simple domains with large problems remain hard and challenging for current planners. A macro-action, defined as a group of actions applied at one time, can make jumps to reach a goal at less depth in the search tree and thus problems, not solvable within a given time limit, might become solvable. FF Style planners like Macro-FF and MARVIN showed some improvement with macro-actions. But both of them somehow need knowledge about the domains and the search algorithms as Macro-FF uses static facts and MARVIN uses plateau escaping sequences to generate macroactions. There is no known method capable of learning good macros without any significant structural knowledge about the domains or the planning algorithms. Genetic algorithms are automatic learning methods that require just a method to seed the initial population, definitions of the genetic operators on the populations, and a method to evaluate individuals across the populations but no structural knowledge about the problem domains and the search algorithms. Genetic algorithms have promising results in learning control knowledge for a domain and some success in generating plans but have not yet been tried to evolve macro-actions. This paper presents initial results of applying a genetic algorithm to learn macro-actions in planning problems.