Mercury Planner : Pushing the Limits of Partial Delete Relaxation

Mercury is a sequential satisficing planner that is based mainly on the red-black planning heuristic. Red-black planning is a systematic approach to partial delete relaxation, taking into account some of the delete effects: Red variables take the relaxed (value-accumulating) semantics, while black variables take the regular semantics. Prior work on red-black plan heuristics has identified a powerful tractable fragment requiring the black causal graph – the projection of the causal graph onto the black variables – to be a DAG; but all implementations so far use a much simpler fragment where the black causal graph is required to not contain any arcs at all. We close that gap here, and we design techniques aimed at making red-black plans executable, short-cutting the search. Mercury planner is entered into sequential satisficing and agile tracks of the competition.