The Chess Variant Pages

[ Help | Earliest Comments | Latest Comments ]
[ List All Subjects of Discussion | Create New Subject of Discussion ]
[ List Latest Comments Only For Pages | Games | Rated Pages | Rated Games | Subjects of Discussion ]

Comments/Ratings for a Single Item

Later Reverse Order Earlier
Variants playable against the diagram's AI. Index of variants that can be played against the interactive diagram.[All Comments] [Add Comment or Rating]
H. G. Muller wrote on 2021-11-22 UTC

OK, I added those to the list.

Some time ago I slimmed down the Quiescence Search of the diagram's AI, because in large variants with many powerful pieces it was still prone to search explosion. (It now unconditionally searches only Low x High captures that capture the last-moved piece, or punish moving away a (soft-)pinned piece; other captures need to consume part of any remaining depth budget.) Because of this diagrams that used impractically long thinking time might now perform acceptably. For instance, after this change I had no problems having the Diagram play Macadamia Shogi against Ai Ai. (But I still don't want to add Macadamia Shogi to the list, because it is not possible to enter the triple capture of the Lion Dog. Even though in practice there probably will never be an opportunity to make such a capture.)

Adam DeWitt wrote on 2021-11-22 UTC

Hectochess now has a diagram with playable AI. I also made a new variant, Banner Xiangqi, which has its own playable diagram as well.

H. G. Muller wrote on 2021-05-08 UTC

However, I noticed that Ryugi was missing.

Oh, I completely missed that one.I added it now. 9 hours is pretty extreme. That was on 2 ply? There must be an enormous search explosion in quiescence search. I guess the problem is jumping generals, which can always make captures, mostly bad ones. And the diagram's AI doesn't use a transposition table yet. So if you have several generals, it will be trying all their captures in combinatorially many orders. In my Tenjiku Shogi engine 'Inferno' I solved that by riguorously pruning any H x protected L capture in QS.

Anyway, the diagram behaves badly for long thinking times, because the search is done by the main thread. This triggers complaints from the browser ("a script is slowing down your browser") if it takes more than 10 sec. I mus figure out how to delegate the search to a 'worker thread' that runs in the background; this is what Jocly does for ist AI. (In the 2-vs-1 checkmating appletI solved it by generating the EGT in steps, having each step schedule the next through a timeout, so that it could attend to other browser functions (such as actually displaying and formatting the step results) in between.)

I left diagrams where I was in doubt out of the list. That included most large Shogi variants, also my own 13x13 variants. These can be added later if they check out OK. Perhaps dai Shogi should not even be in the list; I never tried that one. Sac Chess also seems to give problems, at least at > 2 ply.

Adam DeWitt wrote on 2021-05-07 UTC

This is a very nice index you have so far. However, I noticed that Ryugi was missing. While it is admittedly not my favourite game, it is just as simple for AI to play as Yangsi, if not simpler. If you can add Ryugi to the list that would be great.

Also, my larger Shogi variants have diagrams that an AI can play as well, though not including these is sort of understandable, as the AI would of course take a very long time to make a move. I actually had the AI play a game of Taishin Shogi (my largest game) for the first 60 or so moves, and in the worst case it took upwards of 9 hours for the AI to make a move due to the sheer size and complexity of the game, which only increased as more moves were made, although this would theoretically lessen as the board thins.


4 comments displayed

Later Reverse Order Earlier

Permalink to the exact comments currently displayed.