What did we learn from computer chess?

[Kappatoo]

I posted the following a while ago on the Rybka forum. I'm curious what people here say: What are the most surprising/interesting/important things we learned from computer chess?
- We certainly learned many things about chess - for instance, there were surprising discoveries about certain opening lines, about endgames (from tablebases), etc. What do you think are good examples here? And were there any more general insights - for instance, about the value of material, about piece values, about how close or far humans are from perfect play, etc.?
- Did computer chess teach us things about computer science? What was it?
- Did computer chess teach us things about the human mind? If so, what?

[cdani]

For example that there are a lot more of possibilities of drawing a game. Or that a lot of previously thought inferior lines can be played, that is more or less the same.

Also new players acquire a lot better technique, specially tactical one, and much faster.

Also that the best players are much weaker than thought 30 years ago.

[melajara]

We learned that the human mind has been very overrated.

This is important for assessing the probability of a forthcoming AGI (Artificial General Intelligence), then ASI (Artificial Super Intelligence).

[jdart]

There is very little similarity between how computer programs approach the problem of finding the best move and how humans do it. So there are few inferences you can make from computer chess to human cognition.

Computer chess is all about heuristics and these have certainly become more refined over the years. There are a lot of advances in algorithms but these tend to be fairly restricted to the domain of chess or to similar turn-taking games. One exception being machine learning, which has wide applicability.

We did learn a lot about chess. A great deal of pre-computer analysis was revealed to be flawed. Tablebases have revealed a lot about endgames. Chess opening theory is now heavily influenced by computer analysis. I think it has also been shown the even very strong human players play sub-optimally with high frequency, although I am still impressed by how accurately GMs can play at times (they just can't sustain it over a whole game, typically), and how deeply they see into positions at times (but again, not consistently).

--Jon

[royb]

We learned that the human mind has been very overrated.

This is important for assessing the probability of a forthcoming AGI (Artificial General Intelligence), then ASI (Artificial Super Intelligence).

To your point about how the human mind has been overrated: especially when that mind is applied to the task of annotating chess games. It is enlightening to see how low the accuracy of GM annotations are compared to an engine's. So many books of GM game collections are surprisingly littered with errors as the engines can now show us.

That does not mean however that I find no joy in reading such books, I just know not to trust much of the analysis they present. I still find the notes they add about the personal side of things very enjoyable (who was leading, who needed to win against who in the current round, etc).

[Kappatoo]

Thanks to the input!
Some of you said, to paraphrase, that engines have shown how bad humans are / how overestimated humans used to be. Could you elaborate what you mean by this? Bad compared to what, and what was the prior estimate that humans didn't live up to? Also, is it fair to say that humans being bad / overestimated at chess is equivalent to chess being more difficult than was previously thought? Finally, can you give pre-computer-era quotes about humans' capabilities in chess that turned out to be wrong?

Concerning tablebases: What do you think are the most surprising discoveries here?
Some discoveries that surprised me were: KQ+g- and h (or a- and b-)pawn vs. KQ is often drawn; there are winning positions take take many hundred moves to convert.

Any other more general lessons about chess? I seem to recall people talking about piece values in modern engines being quite different than what is taught to human beginners. Also, my impression is that modern engines are remarkably non-materialistic.

[abulmo2]

Thanks to the input!
Some of you said, to paraphrase, that engines have shown how bad humans are / how overestimated humans used to be.

An example of how human are bad: James Mason estimated the number of positions in Chess after 4 moves (8 plies) as 318,979,584,000. Computer counted them as 84,998,978,956.

[Kappatoo]

Thanks to the input!
Some of you said, to paraphrase, that engines have shown how bad humans are / how overestimated humans used to be.

An example of how human are bad: James Mason estimated the number of positions in Chess after 4 moves (8 plies) as 318,979,584,000. Computer counted them as 84,998,978,956.

Let me rephrase:
Some of you said, to paraphrase, that engines have shown how bad humans are / how overestimated humans used to be at chess.

[jdart]

Other examples of tablebase knowledge are that two Bishops vs. Knight is actually a win, and so is Queen vs. two Bishops (with some exceptions), and BQ vs. RR, but some wins take > 50 moves (these are mentioned in Karsten and Mueller's Fundamental Chess Endings, but first published by John Nunn in Secrets of Pawnless Endings and Secrets of Minor-Piece Endings).

Re human performance, there are many, many examples, in fact any modern annotation of GM and IM play will show errors, sometimes serious ones. I also remember having a book from the 70's, How Karpov Wins by Mednis. Mednis was very admiring of Karpov's endgame play, which he was generally famous for. But the book is full of errors of analysis (seen with the aid of computers) and Karpov's play does not look as impressive when seen under the microscope of the computer.

I would add: the fact that top computer engines are now really unbeatable in a match even by the strongest players shows that human accuracy in play does not approach that of computers.

--Jon