Show HN: Sokoban AI Solver(mkornreich.me) |
Show HN: Sokoban AI Solver(mkornreich.me) |
I suspect that the next big AI breakthrough will result at least in part from constraining LLM decisions with old AI approaches. Frank Coyle presented the idea of ontologies constraining LLM output about a month ago: https://www.youtube.com/watch?v=Sir59K8ZDPU
Going beyond that, I wonder if an agent could keep a running list of assumptions & known facts (with confidence levels/intervals), test them (actively & passively), update them when observations contradict them, and act based on them -- not merely as an emergent behavior, but as a provably correct (old AI based) algorithm embedded in the transformer architecture.
AFAIK bridging deductive and inductive AI has been understood as the trick for "AGI" for a long time, probably even before it was called AGI.
I really want this winter of deductive AI to be short. We need both sides and can't afford a long winter like the one inductive AI suffered.
> Going beyond that, I wonder if an agent could keep a running list of assumptions & known facts (with confidence levels/intervals), test them (actively & passively), update them when observations contradict them, and act based on them
I don’t think that’s “going beyond that”. It’s a blackboard system from the 1980s (https://en.wikipedia.org/wiki/Blackboard_system)
That's not really what I was referring to. I meant more of a system of working assumptions (call them first-class axioms, propositions, facts, whatever) that govern behavior and can be updated when observations occur that contradict them -- not a task queue or a database of all known facts. Sort of like human core beliefs that dictate how we see the world.
Interesting, thanks for sharing!
I plan to give it my own shot with vibing my own harness; I plan to use https://glean.software/ as the central database for everything code, then for documentation. After that, it could be used for bookkeeping for reasoning in a style similar to what you describe; after all, Angle, the query language of Glean, is a https://en.wikipedia.org/wiki/Datalog with some extras (I don't know which yet), so stuff like onthologies could be queried quite well, I would guess.
Seems to be AI in the older sense from 10 years ago?
In 2015, https://en.wikipedia.org/wiki/AlphaGo came around and latest from there on, AI was associated heavily with NNs, deep learning and so on (but not with the transformer architecture which became popular later, the foundational paper itself was published in 2017: https://en.wikipedia.org/wiki/Attention_Is_All_You_Need).
If you squint a little, the linked project is basically a https://en.wikipedia.org/wiki/A*_search_algorithm with optimized implementation, heuristics and so on. I also think that A* was associated with AI due to its use in path finding in early robotics - But I am not sure!
Anyway, one thing that's fun here is that you can trigger the AI solve from any board state. So in particular on puzzle 12 I was interested to see that an initial push (to escape from the 'box' where you start) I'd written off as untenable turns out to be the optimal solution. Then of course it's fun to watch the solver tackle the initial conditions I solved under (and still beat my number of moves).
Might be kind of fun to play with "pessimizing" the puzzle - like, how can you move blocks around to provide a maximally adversarial place to hit the "solve with AI" button? (obviously you don't get to count your initial moves around the board, or you could just move back and forth to get the most pessimum (thanks, Mel) solution.)
Edit: Puzzle 14 feels odd. Super easy, why is it at 14? Maybe something tricky about it that I'm not seeing, perhaps the shape of the arena makes A* harder or something?
Also, 15 is interesting and highlights a theme I'd noticed, which is that often the initial moves of a puzzle seem pretty locked in, and the place where the AI shaves moves off my solution in in some clever approach to the "stacking" of boxes onto the goals. I guess that seems kind of obvious when I write it out.
Anyway, thanks for something to noodle on this morning!
Kinda tempted to go play with it now and see if it becomes more obvious...
I wonder: maybe the state is overly compressed? Could it speed things up to store (boxes, [every position the keeper can reach without pushing]) rather than (boxes, representative keeper position), so we can reduce recomputation of the keeper walking around?
I wonder: maybe A* is counterproductive, as obvious heuristics have traps? Maybe BFS is better?
I wonder: the search doesn't actually "skip over" walking states, it just hides them in the processing of each element in the queue, so adding them to the queue might actually be faster?
I wonder: are there any other simple pruning techniques that you could incorporate? Any learnings from state-of-the-art Sokoban solvers, like this one? -- https://ieee-cog.org/2020/papers/paper_44.pdf
Many interesting questions... sadly, the webpage is written by AI, so there's zero discussion of these tradeoffs, future avenues, or rejected ideas, in favor of meaningless self-congratulatory copy about the "provable optimum" and silly claims like a bucket queue being allocation-free.
http://www.sokobano.de/wiki/index.php?title=Solver_Statistic...
Also, what counts as "complexity" in Sokoban puzzles. Does it plateau at a point, where board size/box count starts scaling the solving time more linearly?
The reasoning behind it is that I want to put as much logic as possible into the harness where it is deterministic, controled and fast - My goal is to make agentic coding usable for my purposes even with small local LLM models. Those might botcher generating Prolog queries sometimes, but they might might be able to used my API/interface/whatever reliably (and then the harness does everything else under the hood, including instantiating the Prolog templates and running the queries)
While I have a feeling that subject 'difficulty' is necessarily a slippery concept, they focus on 'context switching' as a key element of difficulty. In sokoban terms - how often you have to alternate between pushing one box and pushing another. This too can be gamed/trivialized, but, when I used it as a heuristic is was very good at generating the most horrifically difficult levels, much moreso than just going for 'solution length'.
On more general notions of complexity. In sokoban terms, the number of crates trumps everything else - for solvers I've written you quickly get exponential explosions with the number of crates. Nothing else really is significant.
I've also been working on solvers for more general classes of these games (puzzlescript games) and it's surprising how powerful generic solvers still are. PuzzleScript+MIS https://dekeyser.ch/puzzlescriptmis/ (not by me) is one powerful tool that uses PuzzleScript as a basis. I've worked on speeding up the solver a bunch (not currently integrated), figuring out good general heuristics for different kinds of games ( https://github.com/increpare/puzzlescript-labs has various experiments in this direction, including a modded version of PS+MIS). It's a nice optimizaiton problem for focusing on making numbers go down - there are lots of games to test on.
I think the crates trumping other complexity metrics isn't entirely obvious to me. For problem 15 in the OP's post, author says it was too expensive to compute at runtime in the browser. From a human perspective, it's not apparent why, as a large part of the solution is very repetitive. It feels as if there should be a more condensed representation for iterating over problems like that one.
If I may gauge your opinion on it, have you looked into MazeBench? It comes from LLM benchmarking circles, but seems to suggest a search space that's too difficult for LLMs, even with tools, to solve. Curious how much overlap the PS/MIS solvers would have with solving something like this.
"Solvability" would mean number of turns to solve the game. It is known for some puzzles and can be found by brute force, otherwise you need to figure out a proof.
I'm not sure how 'in' you are, you might know this already, but for sokoban IIRC the recommended way is to treat things is topologically - you always do a flood-fill empty space from the player so that the decision isn't whether to go up/down/left/right on this turn but which accessible side of which crate to push. It decomposes quite well, and maybe makes the complexity a bit more obvious? Hmm...
I haven't looked into mazebench, but yes this is very related stuff.
In either case, I think when trying to assess complexity it might also be useful to consider the "narrowness" of the winning move sequence. Positions where the number of moves that win/make progress towards the goal is a small fraction of the number of available moves would arguably be harder or more complex than positions where a larger percentage of the moves win/make progress. In other words, finding a smaller needle and/or in a larger haystack makes the problem harder / more complex.
I feel like there's an interesting tradeoff of storing/computing cheap representations vs exploring a smaller tree.