Basically: the less accurate AStar's heuristic function is, whatever the domain, the more nodes it has to expand to reach the goal. That's because with a poor h

Basically: the less accurate AStar's heuristic function is, whatever the domain, the more nodes it has to expand to reach the goal. That's because with a poor heuristic the fringe is expanded in all directions rather than strictly in the direction of the goal. Manhattan and Octile are not great heuristics because they assume the path to the goal is always unblocked. So you end up exploring dead-ends and other areas that look promising but which the optimal path does not cross. You can see the difference in performance by comparing different heuristic functions; for example the special case of AStar where h(n) = 0 (i.e. Dijkstra's algorithm) vs the Manhattan distance heuristic. The latter is more informed and the search proceeds faster. The paper to which I refer builds a database of costs between every node and certain landmark points. Given such information you can now better estimate the distance between two points by computing a distance differential. Bottom line: their search is several times faster than vanilla AStar using a plain-jane Octile or Manhattan heuristic. More generally, the more landmarks you use, the better the heuristic. However, with better performance comes the need to store more costs which can introduce a significant memory overhead. The details are in the papers I mentioned if you want to know more.