Title:	Learning General Policies with Deep Reinforcement Learning
DNr:	LiU-gpu-2026-13
Project Type:	LiU
Principal Investigator:	Martin Funkquist <martin.funkquist@liu.se>
Affiliation:	Linköpings universitet
Duration:	2026-10-25 – 2027-07-01
Classification:	10210
Keywords:

Abstract

Classical planning systems rely on symbolic search and handcrafted heuristics, which are computationally expensive and require the same effort even for problems solved many times before. This project uses deep reinforcement learning (DRL) to learn general policies for classical planning domains: neural policies that map symbolic states to actions and transfer across instances, reducing or eliminating search at inference time. We combine PDDL-defined domains with relational graph-based state encodings, such as relational graph neural networks (R-GNNs), and modern RL algorithms. Policies are trained on small problems and tested on larger ones to encourage generalization rather than memorization. Key challenges include generalization across MDPs, sparse rewards, and long horizons. Training GNN policies with RL requires millions of batched passes over graph-structured states, repeated across many domains, algorithms, and seeds, so a GPU cluster is needed. The expected outcome is a scalable learning-based planning framework that complements classical methods, with applications in robotics, automation, and AI planning.