Although this is labelled "off topic" I think there is a strong go connection apart from AlphaGo.
I mentioned in a previous thread a 30-second evaluation method presented by Mizokami Tomochka. Although limited to about the first 40-50 moves it really seems to work.
But what is more fascinating is that, like Google Brain AI, it seems it shouldn't work yet does. It shouldn't work because it starts with a chaotic, somewhat randomly chosen slice of unlabelled data - but recognises if not a "cat" at least a good move that should be played. By chaotic, random and unlabelled, I mean that he starts from a half portion of the board and simply counts the stones on each side. There is no assessment of weak groups, thickness or any other such classical elements.
If this method was programmed as is into a computer the output would often be farcical. In real life, Mizokami's method can apply human controls. For example, if there is a group about to be killed you can override the numerical assessment and save it. But in the fuseki and early middle game, such controls don't have to be applied very often. Also, his method benefits from the fact that there is often more than one good move early in the game. The upshot is that his method can be applied more often than not.
His method can perhaps be described as "AI plus common sense." As I have explained by quoting chess writer John Watson in another thread, there are strong indications that modern pros in both go and chess are already using this sort of approach at even deeper levels, supported to a large degree by huge databases. Mizokami's method is certainly a useful and interesting way of getting a toehold into this new way of thinking.
|
(emphasis added))