Abstract

Automated planning is one of the foundational areas of AI. Since a singleplanner unlikely works well for all tasks and domains, portfolio-basedtechniques become increasingly popular recently. In particular, deep learningemerges as a promising methodology for online planner selection. Owing to therecent development of structural graph representations of planning tasks, wepropose a graph neural network (GNN) approach to selecting candidate planners.GNNs are advantageous over a straightforward alternative, the convolutionalneural networks, in that they are invariant to node permutations and that theyincorporate node labels for better inference. Additionally, for cost-optimal planning, we propose a two-stage adaptivescheduling method to further improve the likelihood that a given task is solvedin time. The scheduler may switch at halftime to a different planner,conditioned on the observed performance of the first one. Experimental resultsvalidate the effectiveness of the proposed method against strong baselines,both deep learning and non-deep learning based.

 

Introduction (beta)

None

 

Conclusion (beta)

None