Computer Chess of the Year 2050

[S.Taylor]

Computer Chess of the Year 2050

Solved. All games end in a draw.

Well, might be, maybe weakly.

This is the question! Is it only becauase they are drawn, weakly.

[Laskos]

Computer Chess of the Year 2050

Solved. All games end in a draw.

Well, might be, maybe weakly.

This is the question! Is it only because they are drawn, weakly.

Solved weakly because in chess it's probably easier. It means to prove that the standard opening position is a draw. An engine which solves weakly chess will not lose any games from the standard starting position (if the solution is a draw), but will miss some wins against imperfect engines.

[duncan]

The diminishing returns behave empirically like 1/N^(1.3 to 1.9), where N is the number of doublings or depth. Adding the terms up in this series gives a finite result. I am not inventing here 1/N or 1/(2^N) series out of the blue.

Bob I think you said the series adds up to infinite. if you could you respond to this I would appreciate it.

[Laskos]

I believe that if draws become more than 99% of the TCEC games then practically programs with the right opening book will never lose a game simply because with the right opening book they will play drawing lines with black and not take risks in order to win.

It is possible for example that programmers find that with 1.e4 e5 their program get 100% draws when with 1.e4 c5 they only get 98% draws and in this case the practical choice in games against slightly stronger opponent(not TCEC) may be 1.e4 e5


It may be interesting to have some automatic program that build books for engines(one for a win and one for a draw)
when the target of is of course win for the stronger program and draw for the weaker program and after using this program to generate books
test engines with the relevant book.


Practically we have 4 players
1)stronger player(book win)
2)stronger player(book draw)
3)weaker player(book draw)
4)weaker player(book win)

We can make matchs
1 against 3, 1 against 4, 2 against 3 and 2 against 4
and we should have based on the results
rating(1)>rating(2)>rating(3)>rating(4) but it may be interesting to see the difference.

Yes, the hunch is very good, the choice of openings will be VERY important. Computer Chess by 2050 will integrate the opening choice theory in all relevant play, from tournaments to testing. In this thread:
http://www.talkchess.com/forum/viewtopic.php?t=61105 , Michel Van den Bergh had a very important contribution to side-and-reverse play from openings, the usual play in tournaments and tests. He showed that 5-nomial is the correct description in this case, and that anything from SPRT to p-value for number of games to stop at certain confidence goes as score difference squared over the 5-nomial variance. With this in hand, and using Bayeselo model (ELO model, draw model, unbalance (bias) model), I conclude now that by 2050 the most important openings will be skilfully selected UNBALANCED opening positions. These positions will enhance resolution by an order of magnitude compared to balanced ones (as used today). The observation is that Bayselo and other models to include Davidson draw model (empirically more suited) like Ordo, show a great enhancement of resolution for very high Drawelo (eloDraw) due to the value of the 5-nomial variance and higher difference in score with very "biased" (unbalanced) opening positions. This was not evident until now, when everyone in Computer Chess computed just the usual 3-nomial variance. With high eloDraw and similarly high eloBias (unbalance) the enhancement in score difference compared to regular, balanced openings, is shown here from Bayeselo model:



The difference in score can be as high as an order of magnitude larger for unbalanced (biased) opening positions compared to those balanced. Meanwhile, the empirical 5-nomial variance is close in these two cases. Hence, the relevance (sensitivity) is an order of magnitude higher.

This prediction of the model is confirmed experimentally. For Draw rate of order of 98% with balanced positions (eloDraw~800), we have:

Score of SF2 vs SF1: 96 - 75 - 9829 [0.501] 10000
ELO difference: 0.73
5-nomial error: 0.44
the match is pretty incoclusive, t-value is 1.66.

Introducing a ~800 ELO points bias (unbalance), we get:

Score of SF2 vs SF1: 3659 - 3027 - 3314 [0.532] 10000
ELO difference: 21.99
5-nomial error: 1.49
the match is highly conclusive, t-value is 14.8

So, the highly unbalanced positions for very high eloDraw (800 to give 98% draw rate) will shorten the number of games compared to balanced opening positions by 2 orders of magnitude. That is important, and the choice and theory of unbalanced openings will be an important part of Computer tournaments and testing in 2050.

For fun, I took an even higher unbalance, eloBias~1200. The resolution decreases, as the Bayeselo plot predicted:

Score of SF2 vs SF1: 4613 - 4474 - 913 [0.507] 10000
ELO difference: 4.83 +/- 6.48
5-nomial error: 0.78
t-value is 6.19

So it's important for the optimum to choose the unbalance comparable to eloDraw.

[Lyudmil Tsvetkov]

In 2050 we will:

- watch beautiful chess
- average game length will be over 200 moves (the stronger the opponents, and less mistakes made, the longer the game)
- will Komodo still be the strongest engine?

[jefk]

for example that programmers find that with 1.e4 e5 their program get 100% draws when with 1.e4 c5 they only get 98% draws and in this case the practical choice in games against slightly stronger opponent(not TCEC) may be 1.e4 e5.
It may be interesting to have some automatic program that build books for engines(one for a win and one for a draw)
when the target of is of course win for the stronger program and draw for the weaker program and after using this program to generate books

yes, in fact as extra factor (besides eval and statistics) a sharpness
(or unbalanced) factor could be calculated for opening positions,
and added in the advanced opening book editor (Sirius?or Aquarium 2050) .
Then specify the sharpness you want, eg. from rating difference,
combine the three factors, and do a full minimax. voila.

Whereby for eval in future also some neural network thinking
(Falcon?) or MC evaluation for slow games (depending on future
comp speedups) may play a role. But anyway ELo will converge,
chess is a draw in my opinion so small positional inaccuracies
can be allowed after a while; Elo is a statistical thing, and
you never can get beyond 100 pct accuracy. Lets
assume that top engines now play about 95 pct perfect.
And presume that 99,9 pct is sufficient for a draw.
Then it doesnt matter if you have one engine which
plays 99,999999999999999999 pct perfectly and the other
one onlly 99.91110000 perfectly. Result is a draw. And rating
probably will not go above 4000; maybe even less.

jef

PS about computer speedup, currently it seems ''Moore''
seems to be stagnating a bit because of technology limits,
which they try to overcome by adding more cores. Small
breakthroughs /stepwise improvements again will be achieved by
other technology, which , depending on costs, will be coming in
small steps i presume. Eg. more photonics, and stuff like that.
Not talking about quantum computing, this is more for
cryptography and so on.

[Vinvin]

I think I better start learning Go!

Mark

Crazyhouse !
http://www.talkchess.com/forum/viewtopi ... 39&start=0

Or shogi : https://en.wikipedia.org/wiki/Computer_shogi