Want to play a smarter Pente AI?
Posted:
Mar 4, 2018, 2:53 AM
Hi all,
I love Pente, but I'm new to this community. For a college class, my group and I are planning to develop an adaptation of Google-owned DeepMind famous AlphaZero algorithm to Pente. Intuitively, that means we will tell the computer how to play itself a bunch of times and "teach itself" Pente. This approach beat the human Go champion, Lee Sedol, in 2016, and it can be applied to Pente as well.
I notice some of you are very highly rated players. Would you be interested in playing this AI?
Please either directly message me or reply to this thread; I don't know the Pente.org system very well yet. Cheers!
Re: Want to play a smarter Pente AI?
Posted:
Mar 5, 2018, 2:14 AM
Are you proposing real time games or more correspondence/turn based games with more time given to each side to analyse positions from a couple of days to a week to consider moves? The reason that I ask is that over the years many of the strongest real time players stopped playing due in large part to the first player advantage. Istvan Virag who co wrote the paper proving free renju to be a sure first player win, for example, stopped playing in 2003, when ratings were still adjusted on the basis of single game losses here, whether as first or second player. Several of the current top rated turn based players here played real time games against him and I suspect will attest to his real time playing strength. He finished playing with a 2089 rating, I believe, at a time when ratings were thought by expert players to hit a ceiling not far above that due to the first player advantage.
Dmitri Krasnonosov, considered by many in that era to be the strongest 2 day per move player hasn't played since around the same time or a bit before. Scott Justice ditto, aside from a couple of sets. Nosovs not in several years here; I think he finished with the top real time rating in the era of set based rated games. In other words, you might not be getting the best living human real time players by asking here now. As I mentioned in a semi tongue and cheek forum post here, I suspect a team of humans and current AI could play an AlphaGo zero style Pente program as first player in turn based games and win close to 100% - over 95% of the time, let's say, given a week to analyse each move. This mostly shows that Pente is generally considered to be a sure first player win even with tournament rules. If you choose to compare your project to other programs, they'll need better opening books than the program here has, as it plays sure P1 loss openings at least some of the time.
Message was edited by: watsu at Mar 5, 2018 2:14 AM
Retired from TB Pente, but still playing live games & exploring variants like D, poof and boat
Re: Want to play a smarter Pente AI?
Posted:
Mar 7, 2018, 6:17 PM
Definitely not trying to discourage the project - which I find interesting - with my last post, just trying to provide food for thought in terms of measuring the success of the project. With an expertly crafted opening book, it would likely be tough to beat the program here, even though a self taught AI might eventually prevail in the long run.
Message was edited by: watsu at Mar 7, 2018 6:18 PM
Retired from TB Pente, but still playing live games & exploring variants like D, poof and boat
Re: Want to play a smarter Pente AI?
Posted:
Mar 7, 2018, 9:04 PM
I've put together a few specific examples of where the programs opening book needs improvement, since it sometimes play into the e positions as player 1: 1. K10 L9 2. N10 N9 3. K8 or equivalently 1. K10 K9 2. N10 M9 3. N8 (the hammer 🔨 analysed as a sure loss for P1)
1. K10 L9 2. K14 J9 3. M12 (or H12) I've played this as P2 enough that I'd be shocked if it's not a sure P2 win, though I haven't seen it analysed as such definitively.
1. K10 M13 2. K13 - if not a sure loss, I'm confident there are better second moves for P1 against this move. This is frequently the case; the program will randomly pick a move from its opening book - sometimes it will pick the favored move, but frequently it picks a move or two which can be beaten for its first few moves.
Retired from TB Pente, but still playing live games & exploring variants like D, poof and boat
