Self-Improving WA* Learning with Relational Graph Neural Networks for Planning

A new paper on arXiv (2605.25720) proposes a self-improving WA* learning framework combined with a Relational Graph Neural Network (RGNN) value heuristic to address combinatorial generalization in Deep Reinforcement Learning (DRL). The approach uses best-first search methods like A* to solve planning problems from scratch, without relying on expert demonstrations or random walks. The heuristic guides search, and resulting search data updates the heuristic via Q-learning, creating a loop that yields general policies capable of solving new instances. The work highlights the challenge of sparse-reward domains where standard RL exploration is ineffective.