Features, Projections, and Representation Change for Generalized Planning

  • 2018-05-15 12:12:04
  • Blai Bonet, Hector Geffner
  • 0


Generalized planning is concerned with the characterization and computationof plans that solve many instances at once. In the standard formulation, ageneralized plan is a mapping from feature or observation histories intoactions, assuming that the instances share a common pool of features andactions. This assumption, however, excludes the standard relational planningdomains where actions and objects change across instances. In this work, weextend the standard formulation of generalized planning to such domains. Thisis achieved by projecting the actions over the features, resulting in a commonset of abstract actions which can be tested for soundness and completeness, andwhich can be used for generating general policies such as "if the gripper isempty, pick the clear block above x and place it on the table" that achieve thegoal clear(x) in any Blocksworld instance. In this policy, "pick the clearblock above x" is an abstract action that may represent the action Unstack(a,b) in one situation and the action Unstack(b, c) in another. Transformationsare also introduced for computing such policies by means of fully observablenon-deterministic (FOND) planners. The value of generalized representations forlearning general policies is also discussed.


Introduction (beta)



Conclusion (beta)