"26.8% fewer search steps than standard A∗ search"
So, slightly better than A* which is far from SOTA on Sokoban (https://festival-solver.site/).
What is impressive in this paper? Why is this on Hacker News?
So, slightly better than A* which is far from SOTA on Sokoban (https://festival-solver.site/).
What is impressive in this paper? Why is this on Hacker News?
However, sometimes the specific domain you are searching has other constraints that can be exploited to do better than A*. An example of this being Jump Point Search that exploits certain properties of grid searches if you can only move in certain ways.
If you were able to write a general searching algorithm that can effectively exploit the whatever the specific properties of the underlying domain "automatically" without you having to actually work out what they are, that would be useful right?
A* can't solve even the first level of original Sokoban 90 levels.
This paper is interesting (to me) as an example of how to use transformers for decision making. I don't particularly care if it is up to A* standards at the moment.
What is the scientific contribution of the paper?
They trained transformer on pairs of <sokoban_level, A*_optimal_solution>.
That's not a question you ask other people, that's a bullet point at the top of the outline you created while reading the paper for yourself. You should see the creation of said outline as a measure of your actual interest in the subject.
The research team identified ways to tokenized path finding algorithms for two tasks maze solving and sokoban (a game where a crate has to be pushed to goal) and then trained a model on the execution traces of these algorithms.
The insight this provided was that the "searchformer" model was about 26% faster than the traditional methods. If that is applied to Route-planning, Robotics, and Game Development, it could have tangible performance benefits.
IMHO, it is not a wild breakthrough but an interesting solution to a real-world problem.
Anything that is on HN is on HN because the community likes it
Why wouldn't this be the case for research generally? Has our community really devolved to the point where things should only be noteworthy insofar as they optimize for SOTA for a given problem?
What a sad thought.
Comments like these are antithetical to a strong technical sharing community.
"26.8% fewer search steps than standard A∗ search" For reference of prior art, it's slightly better than A*, which is far from SOTA on Sokoban (https://festival-solver.site/).