At first glance the default looks great! But steps seem too small.
You know how Wordle has a long list of words it will accept as guesses, but pulls from a much shorter list of words it will use as the solution?
Maybe it would make sense to allow any joseki in whatever database as guesses, but only use AI-optimal joseki as the possible solution?
This is a cool idea but I donāt think thereās enough data in OJE to make it easy to filter for AI-optimal-ness. For instance this one doesnāt mention AI at all. So we would need some more dataā¦
Awesome stuff, @Jon_Ko!
How do people feel about removing the jump forward / backwards 10 moves buttons?
I mean these ones:
Another cool feature would be move forward / backward (one move) by swiping right / left on the board.
Whatās involved in getting more data? Hand-checking each variation against AI with a set number of playouts? I guess even then AI can disagree lol. Also a lot of these would require larger board-states that make sense, right?
Maybe āAI-optimalā isnāt the answer, perhaps just modern joseki then? Of course that research might be even harder to do.
The list is relatively arbitrary though, so it surely couldnāt be unreasonable to just put joseki onto the list that easily fit certain criteria (no loss greater than x points at 5000 playouts, with each other corner having the star points occupied, for example)?
Iām kind of out of my depth here but if this is the kind of research best done without automation I could probably help (slowly) to compile a list
I honestly donāt know much about joseki, despite diving into a project that is entirely revolving around joseki.
I think @okonomichiyaki just means that itās not immediately accessible which joseki in the OJE are considered āAI-optimalā. I guess one could use an AI tool to check, but knowing how to properly do that is beyond my knowledge.
Either way, I think the tricky part about building the solution sets is the seeming need for some level of human-guided curation. An initial thought was to have the solution set be āmore commonā joseki, but there was also interest expressed earlier for a āharderā set of solutions. I think that having a āmodernā or āAI-inspiredā set would also be interesting. Maybe all of these could be choices somehow, but I donāt know how to properly curate these sets.
Yes right now the pool of puzzles is very arbitrary, just what I and Jon_Ko hand assembled. Iām in the process of getting much more data from OJE for checking guesses against, and this will enable more interesting collection(s) of puzzles. For instance a āhard modeā of longer joseki only, etc.
Right now I have data up to depth 12 or so (over 450 joseki and variations), Iām trying to get to at least 13 to include some well known longer joseki like the 3-3 invasion I just linked. So soon will rethink the pool of daily puzzles
Automated testing of the joseki against local Kata or something is pretty out of my depth tooā¦
Remember when one of the selling points of this game was that it was simple?
If i understand well, itās one problem pro day (like for wordle)
Any way to play with higher frequency?
Yes, if you want to peek ahead you use this link. Just add &puzzle=7
in the address bar, &puzzle=8
, etc. There are currently around 40 and some hard ones in there, beware (did not filter avalanche out yet)
Maybe include a kind of archive with older problems for players who discover the game lately, missed one of them or for players who wants to replay?
You can also go backwards, using the same method, here is the first puzzle
It would not be too hard to make a page with links to these, if thatās what you mean. Actually Iām keeping track of them so we can avoid repeats right away when changing the collection.
I canāt believe it took me so long to finally do thisā¦
Pull request sent
Iāve pushed something I meant to do at the start:
Keeping track of whether todayās puzzle has been solved in local storage. If it has, it will pull the history of submissions out and restore them. I also am piggybacking off this feature to keep track of the brand new dark mode and zoom levels and restore those options on page load.
Check it out with ?debug=true
, Iāll turn this on by default soon.
Also with this flag on, the app will check the latest data I have from OJE, which covers some decent ground: 524 joseki up to depth 14 or so. It wonāt prevent submitting paths that missing, but will drop a message in the hints about it
OJE flags moves as ideal, good, trick plays and mistakes, those flags could be used for filters.
Oh and OJE allows to tenuki, thatās something that needs to be considered too. Does Josekle allow passes?
We could add a pass button, if needed, but are there actually joseki that involve passing in OJE, or is that just for exploring continuations.
Or we could adopt the convention that the solution will never include a pass.
Passes are required for double approaches and such. It would be nice to have those sequences too. It would make it harder to guess, but it is part of joseki knowledge to know when tenuki is a valid possibility.
Regarding different difficulties, one fun part of Wordle is that everybody does the same word on the same day. So it would be cool, if that could be somehow preserved.
For example consider the avalanche josekis.
3-4, high approach, attach underneath, bump, extend back and hane is the same for all difficulties. But then the easy solution is to connect at 3-3 or to extend down. Small avalanche would be more advanced and large avalanche the most.
That sounds like a really cool idea. However, that would require quite careful curation to synchronize different difficulty levels like that.
Nice idea. It would give a bit of hint too on which branch to try according to the difficulty level.
Drawback is that the database will consist of sets of 3 josekis probably selected manually, it would be hard to have any automatic selection.