google has alphago and how realistic is this facebook program darkforest?
Who are the programmers of alphago?
What i do understand is that you quickly declare a program to be strong enough to play one of the 9 dan players. What i do understand from my own go program experiments is that already 20 years ago i outsearched silly any existing go program - they just didn't know how to search. Of course the life and death analysis i hardly did do.
What i do understand is that in 2005 i talked to set up a match chessprogram versus Kramnik and very quickly the negotiation was about that some sort of appointed lumpsum kramnik wanted 100% pay out in case he lost or drew the match and 20% of the lumpsum had to get paid when software would lose the match.
Now i tried to reverse that 20% payout - yet that was non-negotiable.
In that sense it's peanuts for alphabet to buy a 9 dan player - especially if the guy says: "heh my 9 dan status i got from playing games against a human, some sort of exhibition match i don't care about". Usually that caring starts *after* they figure out there won't be a rematch.
How honest will this match be?
What i do understand is the computer shock. I had it myself as well. No go player is used to play a strong computer, like chessplayers are by now. Result in chess is that the top players are tactical way stronger nowadays in their games. Sure they make mistakes yet they blunder less pieces away.
If you do a simple blundercheck over games from start 90s, that's hundreds of elopoints weaker in world top versus todays world top - thanks to the computer.
In go they didn't go through that yet.
What i do understand is that the best chessplayers come from all over the world. If we select the best player from a couple of billions of players then you end up with way stronger players of course than if you select it from a 100 million japanese and some dozens of millions of koreans - the best chinese players play chess nowadays - that's for sure
What i do understand is that go is easy to search in a dubious manner. In chess you still need a way to explore even the most weird line as it's all about capturing his royal highness - whereas in go you can safely forward prune.
What i do understand from some stronger go players is that the average deep combination is 30 plies in stronger go (ladders not counted) whereas in stronger chess that's 10-12 plies (checks not counted).
What i do understand is that in opening as a FM, nearly IM though (got 2 IM norm results and would be pretty easily to get IM in fact if i'd play a tournament instead of just competition where you sometimes of course are in bad shape), that even todays software doesn't have a clue and that in go it actually seems tougher yet if you analyze it, that might not be the case.
You start with empty board. So that's similar to an endgame you start with. Now the real problem with go is that the branching factor is huge - yet other than that it should be way easier than the opening in chess from knowledge viewpoint seen.
If we look objectively then chessplayers still are impressive in making choices in opening/middlegame. Yet far endgame - i refer basically to endgames with less pawns or a few passed pawns - once a very weak area of chessprograms, they are really total superior there nowadays over humankind. And i say that as someone who is GM level in endgame (though not in my last game in Belgium league where i didn't win the endgame yet drew it - bit busy building 3d printer prototype for sales).
By the way Bob - a neural network only optimizes the heuristical PARAMETERS. It doesn't generate *knowledge*. We know from earlier attempts in computer go with good parameter tuning that this was very freakingly effective.
Yet that was playing go programs that already prune in the ROOT. Please realize that hard reality in go.
If you have something that randomly searches versus something that forward prunes hard in the root - then that explains the huge improvement of computer go software.
What i do realize is that because the board is larger yo usimply need a specific number of nodes a second. Some hundreds of millions preferably, to profit more from algorithmic improvements. You need to get that 30 plies.
What i do not know is how well it would scale strengthwise in go.
We'll see in March whether it's the kasparov/kramnik scenario. Yet i really wonder. Is facebook also serious or is this already a run race with some big pile of dollars changing hands and alphabet has something to brag about?
