GRM-0267 Learned Heuristics for Efficient A* Search: Improving Pathfinding in Combinatorial Pathfinding Problems

GRM-0267 Learned Heuristics for Efficient A* Search: Improving Pathfinding in Combinatorial Pathfinding Problems

https://www.kennesaw.edu/ccse/events/computing-showcase/fa25-cday-program.php

This work proposes a learned heuristic framework designed to improve planning efficiency in deterministic pathfinding tasks. Building on the classic 8-puzzle as an initial test domain, supervised models were trained to approximate heuristic values and guide node ordering during A* search. The proposed approach focuses on modifying and enhancing traditional heuristics such as Manhattan distance by incorporating learned corrections that reduce search depth and node expansions. Experimental results show that the learned heuristic consistently improves search efficiency over standard baselines. Although this evaluation began with the 8-puzzle, the framework establishes a foundation for scaling to significantly larger and more practical domains, such as multi-agent pathfinding. This work will demonstrate how data-driven heuristic refinement can extend classical search methods into more complex, real-world planning scenarios.