Making Hill-Climbing Great Again through Online Relaxation Refinement and Novelty Pruning

Abstract

Delete relaxation is one of the most successful approaches to classical planning as heuristic search. The precision of these heuristics can be improved by taking some delete information into account, in particular through atomic conjunctions in the hCFF heuristic. It has recently been shown that this heuristic is especially effective when these conjunctions are learned online in a hill-climbing search algorithm. In this work, we devise a natural extension to this approach using novelty pruning, a recently-developed technique that prunes states based on whether they contain facts not seen before in the search. We evaluate our extension on the IPC benchmarks, where it beats LAMA, Mercury, and Dual-BFWS on many domains.

Configurable Heuristic Adaptation for Improving Best First Search in AI Planning

“…Preferred actions are indirectly used in our approach, since an evaluation of a preferred action is inconsistent only if it is determined by a penalisation associated to Equation (2). In addition, similarly to our approach, partial delete relaxation and in particular delete-relaxation with conjunctions techniques [23]- [25] aim to refine the delete relaxation heuristic by analysing when it makes mistakes in an online fashion. Moreover, in the case of learning conjunctions from the flaws in the relaxed plan, they have the guarantee that the heuristic value will eventually converge to a real plan and we plan to further elaborate on the connections with our approach.…”

“…A local minimum is a plateau of level 0 < l < ∞ that has no exits [21]. Problems related to local minima have been extensively studied by different authors [22]- [25]. In [22], the author defines the exit distance for a local minima and proves that in many domains there are no local minima at all and are proved to be easily solvable by enforced hill climbing with h f f .…”