Title:GABAR: Graph Attention-Based Action Ranking for Relational Policy Learning

Abstract:We propose a novel approach to learn relational policies for classical planning based on learning to rank actions. We introduce a new graph representation that explicitly captures action information and propose a Graph Neural Network architecture augmented with Gated Recurrent Units (GRUs) to learn action rankings. Our model is trained on small problem instances and generalizes to significantly larger instances where traditional planning becomes computationally expensive. Experimental results across standard planning benchmarks demonstrate that our action-ranking approach achieves generalization to significantly larger problems than those used in training.