Source: https://www.microsoft.com/en-us/researc ... for-chess/
GitHub: https://github.com/CSSLab/maia-chess

There is a freely available chess bot now, which was trained to play like humans would, ranging from 1000 to 1900 elo. With about 50% accuracy it can predict the move a player, eg between 1000 and 1100 would play (other chess AI/bots are way worse).

Even more interesting: It can learn your personal style of playing chess and with an accuracy of about 75% can predict which move you will play.

Subsequently it can predict blunders! Scary when you think about training it on your opponents and exposing their weaknesses. Great to learn for your own games : )

_________________
My "guide" to become stronger in Go

My understanding is that this is pretty much what Crazy Stone Deep Learning does at its kyu-level settings (train a network to predict moves by players of that level and don't do any tree search).

People must be starting to cheat with these human-like bots...

But that's not what Maia does, when I understand it correctly. Maia with 1000 to 1199 Elo is only trained on millions of human games of this level and thus "organically" chooses a move similiar to this Elo range. The neural net behind it runs at full power as far as I understand.

When I understand you correctly, Crazy Stone Deep Learning makes the network artificially weaker by not permitting tree search?



Some people certainly could start cheating with this. But unless you only use the next best Elo range, a sensible anti-cheat-algorithm will likely catch you due to playing too good all of the sudden.

_________________
My "guide" to become stronger in Go

Maia does not use tree search either - there are some interesting results that they published in their paper/data showing how adding tree search reduces the prediction quality. Presumably because tree search in the way that current bot algorithms would perform it is a highly imperfect model of how humans would read a position, and so just letting the neural do all the work is more accurate - because smoothly and intelligently performing highly complex pattern recognition in data is what neural nets are good at.

So long as the task itself is not beyond the capacity of the net to model or the quality/quantity of the data to teach without overfitting, of course (note the difficulty they have for getting good Maia models for 2000+ rated players).

So yeah, it's basically just a raw neural net, like Crazystone. No tree search.

Apologies if I'm being overly fussy about the terminology, but the network and tree search are independent things; whether or not you have tree search doesn't make the network itself any weaker. (I'm also not sure what you mean by the phrase "the neural net behind it runs at full power".) If you mean "makes the engine weaker", then yes, engines without tree search are weaker than engines with tree search, but I think the main reason for omitting tree search is just to avoid the issues with grafting a superhuman analysis ability (tree search) onto a flawed-human intuition (network).

To rephrase, I think Crazy Stone Deep Learning and Maia do the same thing:

• Neural net trained on games played by players of a certain level
• No tree search

As I recall, Crazy Stone trained on human games at a certain kyu level plays better than that level by a couple of stones. That phenomenon may be explained as a wisdom of crowds effect. The point being that different players are prone to different mistakes that most players at their level do not make. That may be the case with Maia as well.

_________________
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins

Visualize whirled peas.

Everything with love. Stay safe.