Recently I came across this chess book - http://www.secretofchess.com/ and although I couldn't find any extensive review, it seems to me pretty unique in its approach. The author is very strong in stockfish, he presents a lot of patterns and details their positional evaluations - Pattern A = +-number - that kind of thing. You can see some examples if you look inside the book.

Do you think that the Children of AlphaGo will be able to introduce this kind of approach for students of Go?

_________________
North Lecale

Thanks for the heads up. I'll buy it if it's in the London chess shop. But just for curiosity value. This sort of approach has been tried several times in chess and despite the hype never seems to work for humans. Typically, the authors, like the one here who spends 16 hours a day on chess and is no more than a candidate master, fail to see the irony that they themselves can't even become any stronger with their fancy methods and theories.



No.

Several of us on this forum have repeated the mantra that what is needed is hard, hard work. But I realised the other day that most people (including me) just don't understand what is meant by hard work. A guitar teacher told me that most of his pupils think their goal is to keep learning a piece until they get it right. Absolutely not, he says. The goal is to learn it until you never get it wrong. That's when you can become a pro.

For human go players, pattern matching is an excuse to delay neceassary reading. Almost exactly the same two shapes can have fundamentally different behaviour.

The excerpts from the book show the presentation of the author's ideas, which seem plausible. However, nowhere do those pages show how he has tested those ideas. Yes, he has devoted his life to the study of chess for several years, and has convinced himself of his ideas by playing against strong chess programs. Compare that with Niemzowitsch, who developed his system and tested it in play against strong opposition. So Niemzowitsch was able to demonstrate the value of his ideas with game records. Surely the author could demonstrate the superior value of rook, two knights, bishop, and four pawns vs. queen, rook, and four pawns, if not with his own play, with stockfish vs. stockfish play, or with the results of thousands of computer self-play games. Instead, he cautions the reader against playing his illustrative positions out with computers. That is a big red flag.

I do not think that "learning by pattern matching" has much to do with the author's approach. He has developed an evaluation function that he believes is superior. If he hasn't properly tested it, other people can. True, he claims that it is possible to find the best play without reading, but that would be true of a perfect evaluation function. His is obviously not perfect.

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

Visualize whirled peas.

Everything with love. Stay safe.

As far as I can tell he is pretty much telling people how to compute the Stockfish (or Stockfish-like) position evaluation function by hand (his tables look exactly like the tables found in chess engine code). People do perform a lot of pattern matching when playing chess and Go, but not like that.

This quote speaks for itself:

Congratulations, sir.

I do think that pattern matching is a hugely important part of human expertise in both chess and Go, although of course it is just one component. The canonical example is the study where masters were much better than amateurs at memorizing normal chess positions when given a few seconds to study them, but not any better at memorizing random positions. I could pick practically any game on this forum played by a kyu player and reviewed by a dan player, and find a fair number of important situations where the player was at sea or made an "obviously" bad move, while the reviewer immediately knew what the likely best move was.

Of course, this all has to be backed up by reading.

I am not sure if this is the question, but do I think that AlphaGo and other strong programs will lead to better evaluation functions for go that humans can understand? Yes. Do not underestimate human creativity.

I also think that programs as strong as AlphaGo or stronger will help humans to develop go knowledge in a scientific manner. For one thing, we can get ideas from the play of those programs. For instance, AlphaGo does not pincer as much as humans, either in self play or in play against humans. That may be a quirk of AlphaGo, but what if the next five programs to reach AlphaGo's final level also pincer infrequently? That would be a big clue that humans pincer too often. Also, as I indicated above, strong computer programs can be used to test human ideas, both by self play and by play against other programs. One reason for the relatively slow progress of top level human go is the relatively small number of top level games by which to test ideas. Computers can produce top level games much faster. Amateurs are better placed than most pros to develop go scientifically, because if a pro gets a good idea which tests well, he will benefit by keeping it a secret from other pros. Amateurs do not have the same incentive. Science is a public enterprise.

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

Visualize whirled peas.

Everything with love. Stay safe.

What, IYO, is a better computer-generated human-applicable evaluation function? What is evaluated? By which parameters is evaluation expressed? Why would that be better than existing human go theory, such as my theory of positional judgement?

If your question is to me, I did not mean a computer generated evaluation function, but one scientifically derived by humans.



Existing human evaluation functions suck, that's why. If they didn't, Monte Carlo evaluation would not have been a great leap forward. Perhaps your theory of positional judgement is a significant advance, I don't know. But unless proven otherwise through testing, my guess is that it sucks, too. Even if it is better than other human theories.

Monte Carlo evaluation is not as good as that of human pros, but it is better than what human pros can articulate.

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

Visualize whirled peas.

Everything with love. Stay safe.

I agree with the trend of your remarks about current human theories of evaluation, but I'm not sure I understand this, Bill.

We can make a clear distinction between two aspects of evaluation. One is knowing which side is ahead, and the other is knowing what the next move should be.

Monte Carlo only (at least directly) tells us who is ahead. My guess is that it's generally better than humans in the endgame (even if an exception may be made for the last few points). It may or may not match humans earlier in the game - that still seems to be one of the secrets to be unearthed from AlphaGo.

As regards knowing what the next move should be, humans may actually be better than machines on average, but they also make more fatal mistakes (that seems to be the chess experience).

I apply my positional judgement theory in my games and so win quite a few more games that otherwise I would have lost or whose outcome would have remained uncertain. With my theory in comparison to other theory, many more positions are available for judgement without first having to read for creating a more stable position and accuracy of judgement has increased dramatically on average.

Judgement of expected scores can be used to derive judgement of best moves so the two are dependent.

I agree with what you said, but not with the implied conclusion.

I think the only interesting or "game" part of go, I mean the part that is suitable for intelligent beings, is the extent where reading (search/minimaxing) can be substituted by other methods. Minimaxing is a simple task that does not need intelligence. I see tasks of pure minimaxing more like programming challenges, and I think most players who pick go over chess do so for this non-reading part.

I think you overestimate MC in itself. A program does not need an evaluation function better than humans. Even a slightly weaker evaluation can make a killer player (since it will be applied at later nodes) when paired with the computers' superior search (which OTOH relies heavily on the other kind of evaluation mentioned). And search will always remain the computers' biggest strength - and in the long run, superior to all other (strategic) approaches. But since humans cannot aim to calculate THE real solution, it's still interesting to compete with other methods.

I agree, this is another example of the search <> knowledge opposition mentioned above. Again a slightly weaker evaluation can be sufficient when paired with better search.