The impression I got from Lee Se-dol in the Netflix documentary was that he had a lot of his identity tied up in this (particularly being the best Go player in the world). Not a huge surprise given the time and effort required to do it, but there's probably a healthier mental framing (I think Kasparov has a healthier one).

If people give up on learning or doing something every time someone or something else can do it better then there's going to be a lot of disappointment as things continue to move forward.

The point is we don't really know what limits exist beyond our small pinhole of perception. During industrialization I'm sure people said the same thing: "What's the point of hand-made clothing if machines can do it?". Today both hand-made and factory-made clothing have their place in the market.

Throughout history, tools have disrupted humanity and we've adjusted and expanded our horizons. To me this feels like another wave, and If I could bet and could collect winnings centuries in the future, I'd say this isn't even the last one.

Some track and field records set in the '80s still stand, e.g., Marita Koch's 400 m, Jürgen Schult's discus throw, Galina Chistyakova's long jump.

Maybe he's lost motivation, but this retirement is not a surprise.

No. Consider this scheme.

I take my copy of AlphaGo and for a full year, I'll build a database of all opening positions AlphaGo is willing to explore. I'll rank these opening positions from "best" (Black-wins with most consistency) to "worst" (White-wins with most consistency).

I'll put all of this information into a 16TB database on a singular, $400 16TB Hard drive, and load it into my computer during the contest. https://www.newegg.com/p/1Z4-002P-015K6

If you dare to pick AlphaGo's best move, you will lose. Because I already know which moves AlphaGo will take, and I already checked to find all of the positions AlphaGo wants to play (but loses anyway).

The only way you can equalize the field is if you yourself ALSO build a 16TB database to consult and override AlphaGo's instincts during the game. If you see AlphaGo wanting to play "losing position #6234115", you'll tell AlphaGo to "search harder" and find another move instead.

Good luck.

Go does not have openings. It has reasonable choices to make, with an insane branching factor, with few moves making much of a difference by themselves. Therefore your database will only extend a few moves, and all of the positions that it winds up with will still be very close to even. So your database confers very little advantage.

If you have alpha-go play itself a thousand times, it is unlikely that by move 10 you will wind up with the same board position twice.

This is exactly the problem that made Go so hard for computers in the first place. Alpha-beta search is useless within practical limitations of computer hardware.

That said in a different game, such as chess, your strategy would work very well indeed. (Which is why all decent computer programs have an opening book.)

There are 381 opening moves in Go, but really only 96 because of symmetry. 96 (opening moves) x 380 responses x 379 x 378 x 377 == ~2 Trillion positions after 5 ply.

These 2-trillion positions will easily fit in a 16TB hard drive for $400. That's 8-bytes per position, so you probably can get there with more symmetries and some compression applied.

----------

You're thinking too much like a human. There's no Go-openings in the age of Human-Go. But in the age of Cyborg-Go where 16TB hard drives are allowed, we can begin to exhaustively build openings.

We even have a super-human AI that can automatically, and algorithmically, explore this opening book. We can build AWS-instances with V100 Tensor cores to use neural-nets to explore all of these positions now.

-------------

> If you have alpha-go play itself a thousand times, it is unlikely that by move 10 you will wind up with the same board position twice.

Alpha-Go doesn't seem to implement much randomness at all into the moves it plays. The source of randomness is in time-controls (AlphaGo may choose MoveX before 30 seconds of analysis, or MoveY after 30 seconds of analysis), but this is a fairly constrained number of moves.

Play alphaGo by itself a thousand times, at precisely the same time MCTS-controls (say: 1-million nodes visited in the MCTS tree), and it will probably play the same game 1000 times in a row.

This makes AlphaGo extremely prone to opening database "attacks". Which is why I am using opening books as an easy example for how to beat a particular AlphaGo network. At least, until AlphaGo updates its algorithm for more random play.

If the goal is "Beat AlphaGo" in a game, then the opening book construction is far, far simpler. Even with random elements (ex: AlphaGo picks randomly from the top 10 best positions it generates), that is far more constrained than a full 381 x 380 x 379 x ... style opening book.

No, you're thinking too much like someone who understands chess but not Go.

Suppose that we built an opening book with a trillion reasonable courses of action on it. Each one analyzed well. As you have discovered, you will only go a few ply into the game. And all of the positions that you will be directed towards will have only a small edge.

Instead put a tiny fraction of the computing power necessary to build this book into self-training. You will get a better internal model and therefore a significantly stronger computer player. (That is how alpha-go was built in the first place.) This option will produce much better results for far less effort, and again leaves no role for a human to do anything useful.

The fact that a memorized opening book is useless has nothing to do with human vs computer vs cyborg. It has to do with the characteristics of the game. In chess, it is useful to memorize openings and both computers and humans do it. In Go it is a waste of energy, and it is a waste for both computers and humans.

In my opinion, it's really strange to describe this as an expert "beating" AlphaGo, when really it's just a technique for making AlphaGo stronger than it is without a huge pre-calculated cache of moves.

10 moves into a chess game is actually pretty far. The directions the game can go are massively narrowed 10 moves in.

If the OP spent a month learning Go, I am sure that it would make sense to him as well. Work through a series like https://senseis.xmp.net/?LearnToPlayGoSeries while playing Go regularly against a variety of opponents. Before book 3 it should be obvious.

I'm not an expert by any means, but I'm well past "spend one month" learning the game.

If you like, I could have a look at several your lost games and maybe suggest how to improve. Just a 3k at kgs, but still.

This means if you enter a hypothetical "Cyborg" Go competition (computer-assistance allowed), the majority of newbies will simply be playing LeelaZero #1 plays over and over again.

You don't need to build an exhaustive opening book covering all possible moves. You only need to pick say, the top 5 moves LeelaZero ever considers. If you spend ~16-bytes per position and store 1-trillion positions on a 16TB Hard Drive, you'll be able to exhaustively map the top5 moves LeelaZero considers into 17-ply.

From there, you pick the positions that LeelaZero thinks its winning in, but in actuality is losing. You have a map towards 1-trillion positions to choose from, and your opponent (if they only pick from the top5 best moves LeelaZero ever outputs) will walk into your trap.

---------

As long as your opponent picks the top 5-moves from LeelaZero, you'll have the map towards victory. I think you're severely underestimating the abilities of a simple, dumb, opening database.

Do you believe that AlphaZero could continue to improve dramatically with another 6-months of training? Or if it can improve at another 6-months after that? At some point, the network will reach a local maximum, and it will be unable to improve beyond that.

Characterizing AlphaZero's moves through big-data analysis is innately going to become useful as self-training plateaus. Even Google wasn't able to get more than a few months of training in before the plateau.

At which point, it will be more reasonable to characterize the weaknesses of the network and build an opening book. Avoid the positions that the network was unable to learn about. Etc. etc.

Opening books, at a minimum, would grossly improve AlphaZero's play at competitive levels. Anyone with an opening book of AlphaZero's mistakes will be able to push AlphaZero into a mistaken position.

And then AlphaZero was better than AlphaGo after around a day of self-training.

Furthermore you are arguing for an opening book without considering how small an advantage an opening book would be. As I have said repeatedly, an opening book takes a tremendous amount of work to generate, will only go a few ply in, and the positions it directs you to will only be a tiny bit better.

Therefore for the foreseeable future, more training and better algorithms will produce better results than trying to create an opening book. Theoretically this could change. But that day will not be today or this year. I would be astonished if it happened during this decade. I would be surprised if it happened in a lifetime.

Your proposed approach is an excellent one for many games. But not for Go.

The plateau is real. I'm not sure if continuous self-play will lead to continuous progress for all of eternity. The system is clearly slowing down in self-learning.

AlphaZero's plateau is also well documented: https://i.imgur.com/NMNp6Kq.png

--------

I'm not trying to cast doubt upon reinforcement learning / MCTS / Neural Nets in the game of Go. It is clearly the best methodology we got today.

But anyone who has any experience with neural nets knows about the local-maxima problem. ALL neural nets reach a local maxima eventually. Once this point is reached, you have to rely upon other methodologies to improve playing strength.

Assuming Elo-growth for all time using a singular methodology is naive. We will go very, very far with Deep Learning, but are you satisfied with that limit? Other open questions remain: Go is very far away from being a solved game, even with a magical machine that plays 2000+ Elo stronger than humans.

This only works if AlphaZero never retrains on previous games.

At 5 ply, the complexity of the game hasn't started in any meaningful way. In a typical game, that's 4 corner moves, and then one of a: an approach to a corner (kakari), b: an enclosure (shimari), c: a wedge (waruichi) or d: creating a side framework (such as the Chinese fuseki or Sanrensei). There are some odd opening such as tengen, or corner-corner-corner-kakari which typically turns into a sente fight, but 99% of games will fall into the aforementioned pattern. The database you describe is about as useful as a database of amateur games, since most games, including AlphaGo's games, follow just a few basic openings that early, and even amateurs can play these first few moves "correctly".

Even if you get out to 10 ply you're still only getting partway into a single joseki sequence, often leaving three whole corners of the board which haven't even been approached, so this database still isn't very useful.

Incidentally, your numbers are also wrong. Symmetry reduces the first move to 55 possibilities, not 96, and there are 361 points on the board, not 381.

The basis of reinforcement learning algorithms is the exploratory nature of learning due to the initial application of largely random moves.

Only after some time the agent is given confidence into his learned ability and grafually moved into a more deterministic behaviour mode.

This is the exact opposite of your statement. Star Craft players have noted that the fleet of different AlphaStar instances training in ensemble observed very different behaviour due to this property of RL.

Then you wring out a tiny advantage-- a fraction of a stone. It's a small benefit compared to building other parts of understanding of the game.

Human + computer might still beat computer in Go - this was true for a few years in chess, and even now to some extent in correspondence chess - but what you describe isn't really that.

All you have done is establish a computational arms to see whose computing rig wins when you press the 'pick best move' button. You're not playing Go any more, you're playing Database Administrator.

Based on what we've seen in a couple decades of computer-aided chess, no. A good chess player using a top-rated engine to help them can pretty consistently beat the engine by itself.

There are tournaments and even a world championship in computer-aided (correspondence) chess and you don't come close to winning by just taking the program's recommended move every time.

That's only true if AlphaGo never makes a mistake or if AlphaGo will 100% always make the better or equal decision than a human + computer at any given state of the board. I know the former certainly isn't true and I assume the latter isn't true either, but I don't know enough about Go to say for sure.

Even if AlphaGo makes mistakes, and somewhere on the board a better move can be found, you would also need the human to reliably spot it.

Eg: AlphaGo makes a move. Let's say that at least 20% of AlphaGo moves can be bettered. Is this one of them? How can you tell? Most of the time, you'll mistakenly think a move can be improved and end up playing a worse one.

But, let's make AlphaGo even more fallible. Let's say that at least 50% of AlphaGo moves can be bettered. Again, is this one of those? How can you tell? And more to the point, on the times you are wrong, are you more wrong than AlphaGo is with its mistakes? Because even if you imagine you can spot a better move than AlphaGo and pick the actual better move 50% of the time, you also need your mistaken moves to be better than AlphaGo's mistaken moves or you'll still lose.

Worst of all, you can rule out a really good ability to spot AlphaGo's mistakes already. Let's say 99% of AlphaGo's moves have a better option. If you could spot them all, you'd be beating AlphaGo regularly on your own. As no human can now beat AlphaGo, this plainly isn't true.

So it's likely that:

a) No human can reliably pick a better move than AlphaGo and/or b) No human can reliably spot a move from AlphaGo can be improved, and/or c) Human mistakes are worse than AlphaGo mistakes, so even if you could fight it up to parity you'd still lose.

You can use the relative scores to decide when to overrule the AI. Eg, if move A has a 50.1% win chance with 2k branches explored, and B has a 50.2% chance with 1.9k branches explored, I would go with the opinion of an expert human, as AG thinks the moves are essentially equal.

Plus, you only have to outsmart the car AI once to 'win' - e.g. just override one 'drive into the highway barrier' or 'run over that pedestrian' AI mistake.

An accurate metaphor would be a person on a motorcycle competing against another person on a motorcycle.