On Sample-Efficient Generalized Planning via Learned Transition Models

Abstract

In this work, we formulate generalized planning as transition-model learning: a neural model approximates the successor-state function and generates plans by rolling out symbolic state trajectories. To achieve size-invariant generalization, we evaluate multiple state representations, including graph embeddings. Our results show that learning explicit transition models yields higher out-of-distribution success than action-sequence prediction in multiple domains, despite significantly fewer training instances and smaller models.