ICAPS best paper runner-up award

Rostislav Horčík and Gustav Šír from Artificial Intelligence Center and Intelligent Data Analysis lab at the Department of Computer Science FEE CTU received a best paper runner up (2nd place) at the prestigious International Conference on Automated Planning and Scheduling (ICAPS) conference for their paper "Expressiveness of Graph Neural Networks in Planning Domains."

ICAP 2024

 

Horčík and Šír delve into the challenges and potential of using Graph Neural Networks (GNNs) for learning policies in PDDL domains. While GNNs offer superior generalization capabilities due to their ability to handle symmetries in input graphs, they also suffer from limited expressiveness. The paper investigates the critical role of encoding planning states as graphs and ensuring that states requiring different actions can be distinguished by the chosen GNN model. It provides a comprehensive theoretical and statistical analysis of these challenges across various encoding schemes and GNN models, contributing to our understanding of how to effectively leverage GNNs in planning tasks.

The paper is presented at the (A*) conference ICAPS 2024 held in Canada on 4 June 2024, read the full paper here.