Hey y’all, in a few short days, I’ll be in Dublin for the annual international conference on automated planning and scheduling, I’ll be presenting some preliminary work on triangle search and how it performance in automated planners at the heuristic search for domain independent planning workshop. Abstract for that paper is below. If you’ll be in Dublin next week, reach out! Maybe we can grab a beer and or a coffee, depending on time of day.
LAMA has long been a standard reference point for satisficing domain-independent planning, combining planning-specific enhancements such as landmarks, preferred operators, and multiple queues with search algorithms like restarting weighted A* and greedy best-first search.
In this paper, we revisit triangle search as a base-level search strategy for planning and compare it directly against the strategies that commonly underlie modern planners. In controlled, head-to-head comparisons using the same heuristics with planning-specific enhancements disabled, triangle search consistently improves over greedy best-first search and rectangle search on the Autoscale satisficing benchmark suite.
While LAMA performs best overall, our results suggest that search strategy alone has more room for improvement than is commonly assumed in planning. On its own, triangle search already reproduces some of the behavior of these enhancements, which motivates integrating them explicitly in future work.