Skeleton planning spaces for non-numeric heuristic optimization*

Abstract

The AFTERMATH system implements a heuristic technique for improving long solutions (up to about 250 steps) for robot planning problems. AFTERMATH transforms the given solution into a skeleton solution that focuses attention on repetitious and opposite moves. AFTERMATH attempts to obtain an alternate, improved skeleton. From the alternate skeleton, an alternate solution is built (if possible) to the original problem. If the alternate solution is an improvement, AFTERMATH accepts it as input, and cycles. Although not guaranteeing optimality, AFTERMATH improves many solutions, sometimes gradually in several cycles. Examples can be built for which AFTERMATH obtains an arbitrarily large improvement in one cycle.