Online AI planning with graph neural networks and adaptive scheduling

Automated Planning


Published on


Automated planning is concerned with devising goal-oriented policies executed by agents in large-scale state models. Since no single planner can work well for all tasks and domains, portfolio-based techniques have become increasingly popular in recent years. While offline portfolio approaches focus on finding a single invocation schedule that is expected to work well across all planning tasks, online methods learn to choose the right planner for each given task.

In our paper, Online Planner Selection with Graph Neural Networks and Adaptive Scheduling, published in AAAI 2020, we propose a new online algorithm that can help finish more planning tasks within a limited time. We utilize graph deep learning for planner selection. We propose a two-stage adaptive scheduling approach that enhances the likelihood of task solving within the time limit, over the usual approach of using a single planner for the whole time span. The most interesting part is to split the planning time into two stages, we use a second planner to if she “senses” that the selected one unlikely completes in time. Mathematically, this means designing a contextual graph neural network based on the result of the first-stage graph neural network planner selection model. A future line of work work may be developing new graph neural networks to make the training of these large graphs more efficient and robust.

Please cite our work using the BibTeX below.

title={Online Planner Selection with Graph Neural Networks and Adaptive Scheduling},
author={Ma, Tengfei and Ferber, Patrick and Huo, Siyu and Chen, Jie and Katz, Michael},
Close Modal