Description:
<jats:p>The AI planning field has a long history of introducing yet another search algorithm that is believed to be the best in all domains. Some recent examples are NONLIN, TWEAK and SNLP. In this paper we show, by directly comparing the above three planners, that the quest for an overall winner is doomed to fail. We first argue that these three planners differ in a variety of ways, including methods for termination check, causal-link protection, and subgoal selection. These differences entail different search behaviors in terms of factors such as the branching factor, the search depth and the time spent by the algorithm on each search node. Furthermore, the search behavior is closely related to the characteristics of the problem domain in which a planner is operating. In this paper we identify one such domain feature, expressed as the ratio of the number of negative threats to the number of positive threats. We present an artificial domain where we can control this ratio and show that in fact the planners show radically different performance as the ratio is varied in this domain. The implication of this result for someone implementing a planning system is that the most appropriate algorithm will depend on the types of problems to be solved by the planner.</jats:p>