When will Hex AI be as accessible as Chess, etc.? Hex, Havannah
3 replies. Last post: 2020-08-06Reply to this topic Return to forum
add3993 ★ at 2020-07-16
I really like Hex, and am intrigued by other connection games here like Twixt, but the gameplay on LG is slow and limits one’s rate of learning patterns. With a frequent 12+ hours' wait between moves, I even tend to forget my own analyses. (This is not a complaint to my opponents, I play slowly too.)
So lately I’m playing more Go and Backgammon, because the AI is so good and so easy to play and learn from (KataGo + KaTrain; GNUBackgammon or XG). I want any ambitious developer to have a clear sense of why they’re so good, and I want strong players to discuss as well, because I think it’s very important for recruitment and the development of a larger expert player base.
Here is my “ideal checklist” for a game being in really good shape: there should exist software that is
-playable offline on PC
-easy for an average person to download and install without knowledge of github or command-line
-superhuman strength-explicitly evaluates player (or game-file) moves in some useful form appropriate to the game (usually winning %, but also equity, territory, or whatever else applies)
-shows ranked list of possible next moves according to the evaluation function
Ideally, the software also
-supports creation and analysis of game-tree/variation files
-has a standard AI-engine file format, allowing AI-focused people to experiment while providing an easy interface
-offers a few skins / other quality-of-life things
This is an ambitious list (and I’m not saying anyone owes this work to anyone else); but it’s basically the status quo for Chess and Backgammon, and recently for Go as well. People have these options, and if Hex is to publicly claim its rightful place as a premier abstract game, it has to catch up. Fortunately we now know there is basically a template for the really hard part, the “superhuman” thing.
Of course, if I’ve missed good software, I’d like to know about it. And I’m also eager to learn of any more abstract games that have software approaching this ideal. For example, Piskvork for GoMoku is quite nice, even if it doesn’t deliver on the evaluation part.
spartacu5 at 2020-07-17
Play against Stephen Tavener’s AiAi if you are looking to improve rapidly.
To play against a stronger or weaker bot, just modify the thinking time.
That is the link to the download page.
It plays reasonably with 5 seconds thinking time, on my machine.
I would be really curious to hear assessments from stronger players about its strength at 1 minute per move, 5 minute per move, etc.
add3993 ★ at 2020-08-06
Like some other people, I was not able to get the AI to play on a board larger than 11-by-11.
Also, I was not impressed by its level of play with 30 sec on a fast PC. In particular, I do not trust it to help me develop a better sense of opening play. I will wait for something on the level of Mootwo.