Introducing Maia, a human-like neural network chess engine

I respect what this team is doing, actually, all the ingredients to make bots like human are in place it’s a matter of time. Chess.com could have done this, but there is no such demand, what a difference such bot can make when you have millions of actual human playing.

I don’t know the team, who they are, professional data scientists, chess players, talented developers or a bit of all, but they should not have used AlphaZero in this context, as Zero (if you check name evolution) means a version without human data
“By playing games against itself, AlphaGo Zero surpassed the strength of AlphaGo Lee in three days by winning 100 games to 0” Google team introduced this zero approach on game Go, but logic stands for chess as well.

Bots with different elo = they train bots on games with different elo rating, but the fact that bot trained on 1500 games has 1700+ actual elo means something in their approach is wrong (revealing common chess engine performance on top of human-like NN).

@maia9 plays exactly like humans

@ProposeBurgers I think of the neural networks as learning to approximate (interpolate) what humans will do. So if they see 10 games where black trades queens on e6 and 10 games where the same thing happens c6, the they'll also try to do it on d6. We trained 9 different models, so the three bots we have just have the same designs they were trained on totally different games.

@gandie I'm the main person working on maia chess, and I'm doing it as part of my PhD. I'm just happy my work is creating something other people can benefit from, instead of just creating papers that will be read by other researchers. Regarding the name, we used Leela/lc0 as the starting point and so far all our models are compatible with them, that's what they use to provide UCI interfaces. We are giving credit to the work that came before by saying we're like AlphaZero. There are many other previous neural network approaches to chess and there are even some that use self-play, like Giraffe.

Soon we won’t have anything to do since computers and robots will do it all. What about work? Hmmm. Neural networks for everything? Humans not necessary?

Its amazing! A new era of chess computers.

There’s only one Maia:

en.m.wikipedia.org/wiki/Maia_Chiburdanidze

How can you play against these bots? When I try challenging them to a game it always gets declined...?

This is great! Is there any chance to train the network to be predictive of a bit higher level play? Say 2100 or 2300 in lichess ranking.

I'm enjoying playing the maias. I seem to do better against them compared to similarly rated humans, they don't cheat, they don't insult when you lose and they don't accuse you of cheating when you win.

What's not to like?

@grandie, if you have the time you can read the actual research paper. I would highly recommend it. It is interesting and actually surprisingly accessible for someone who is a layman.

Why their research is relevant is beyond just chess. The main reason they chose chess to base their research on is because nothing else compares to the data offered by billions of chess games. Each chess game on Lichess has the outcome, every move, move time and sometimes even computer analysis already attached. For a researcher such data is a wet dream.

Why their research is relevant? It's obvious AI is taking greater role in so many aspects of our lives and often AI can do a better job than a human, however, while the results of the AI can often be agreeable the workings can often be obscure. In chess the phrase 'computer move' is quite common- it's a strong move but it makes no sense on the surface and only deep and exact analysis reveals its purpose. There is a clear disconnect between human decision making and AI. Their research is in this area.

Your point about training Maia on 1700s rated players ending up 1900 being flawed is wrong. As I read the paper and some of the blogs based on it, the same thought occurred to me. However, if you think about it it actually makes sense. If you learn from ALL the 1700s EVERYTHING they know, it makes absolutely sense that you would end up being stronger than any singular 1700. First of all when Maia learns from these players it doesn't specifically try to learn their strengths. It only tries to mimic and predict their moves. However, good moves are logical and make sense, bad moves are generally less so. So simply by statistics and probability over vast number of positions- it is natural that patterns of good logical moves is more likely to be reinforced rather than less logical bad moves. Yet Maia still learns to blunder in a human way if that way is common enough in human play.

The research of Maia reaches way beyond of chess but even in chess there is some real value to it I believe. If lichess would integrate the maia algorithms in your 'chess insights' you could potentially identify a lot of things about your play and even ways to progress. By training Maia un different skill levels they basically encapsulated the 'style' of certain skill and also what blunders are common to that skill level.

As part of their research they did that. They took only 4 games of a player trained Maia with those 4 games, which is a laughably small sample for AI training. And after studying those 4 games Maia was able to pick out with amazing accuracy games from a sample size of 400 games played by that player. So I think Maia could potentially be integrated in many areas of chess, like tailoring your puzzles to best suit your weakest aspects of play. Suggesting areas of study and more indepth analysis of your play.

Again if there's even a touch of nerd in you, I would recommend reading the full research paper. It's good stuff.