if not much has changed since than how come engines have become stronger over the time? just by testing and playing millions of stupid bullet games?
best computers are vastly superior to most GMs at calculating positions not understanding them. They don't follow any ideas in the opening or in the endgame. they just do math and if hardware is good they are just excellent at math.
No it happens more often than you think... just put 8 pieces including pawns and rooks anywhere and try to evaluate mate or a draw and have endgame tablebases handly. I know many positions where sf is unable to find mate in 40 moves.
Apart from that it sticks to the first line and keep calculating at great depths. I've inputed some alternate moves and once it is in the RAM and if sf finds it better it changes evaluation more than half a pawn and sometimes even drops the first line it was calculating. So they are horizon blind. they can't watch beyond a limit and keep walking on the high way even if there is a shortcut round the corner.
Heh...ok, so, yes, in answer to your rhetorical question, engines have gotten so much better because of massive testing of ideas. There's nothing surprising there; test enough ideas, and you'll find some really good ones eventually.
Tord's post was about the approach to designing an evaluation function. That approach is still the correct one, and what he said about more knowledgeable evals not being equivalent to better evals is just true. People have also refined a lot of search tricks since then, which have also helped substantially.
No one is making any claims about whether computers "understand" anything or follow "ideas".
You said that the top 10 players in the world were stronger than engines, and I said that actually SF is stronger (i.e., would convincingly win any match at long time controls).
Carlsen has admitted this, and coupled with the top players in the world barely managing draws against Fritz and Junior in the mid 2000s, is quite clear, as SF would obliterate Fritz and Junior.
On the subject of endgames, I never claimed that SF played perfectly. No one would make such a ridiculous claim.
Obviously the tablebases will play better than a non-syzygy SF. That's not the point. The point is that the engines still play better than humans, whatever shortcomings they may have.
Of course they make mistakes and are not perfect. We all know that. Just look at the most recent TCEC. Sure, there were a lot of draws between Komodo and SF in the superfinal, but there were still decisive games, so of course there are mistakes.
The problem (for the claim that the top 10 humans are stronger than engines) is that humans make a LOT more mistakes.
All you're showing is that engines aren't perfect, which no one is arguing. Humans are even farther away from perfect play, so humans lose in matches.
Again, Carlsen wasn't lying, nor are the results of playing today's SF against the engines that drew Kramnik and Kasparov (complete obliteration).
You could start with a KotH eval function that only assigns win/loss values to positions with a king in the center, instead of measuring distance and trying to weigh that in with the rest of the carefully tuned evaluation.
For positions where relatively few moves towards the center are required to force a win, I would imagine SF will find the winning path, then?
@psuter: Agreed. That would be a good enough first step, because at least forced KotH wins would be found.
Ideally there should still be bonuses for the king's being closer to the center, but they should be much smaller than in the initial implementation.
if this is the method then lot of testing is required before it can be any good as koth requires even more analysis than regular chess as we have to calculate an additional value here, regarding king's distance from center and safe path to reach it.
Well, the idea is that it shouldn't take much changing to be very good, since the highest level KotH games tend to just be regular chess without an endgame.
Running for the center in the opening and middle game tends to just be really bad, so as psuter said, just returning a plus mate score for the side with a king in the center 4 squares should be enough for some strong play at first, as it'll be strong enough tactically to find those "mates".
To make it even stronger you're right, we'd have to introduce some features like distance from center and start testing different weights. Nothing for free :)
One of the most famous problems in the field of computer science is the Halting Problem:
"In computability theory, the halting problem is the problem of determining, from a description of an arbitrary computer program and an input, whether the program will finish running or continue to run forever."
http://en.wikipedia.org/wiki/Halting_problemGiven an infinite amount of computational resources, a computer can play a perfect game. However, it's not possible to allocate infinite resources, and it's necessary for a computer to manage its resources in order to play well.
I am surprised at how challenging of a game KOTH is, since rule-wise it's virtually identical to standard chess. I had imagined that a modified SF would easily outcalculate most human players, but so far this appears not to be the case!
@Toadofsky: Well, I think in this case it's because the weights for distance from center were so very high.
It would be like modifying SF to value doubled pawns at some very, very high negative value, which would make it not so difficult to play against.
I'll probably tinker around with it a bit tonight to see what I can manage :)
I'll probably first just remove the bonuses entirely, leaving only the defined mate scores for reaching a KotH win, and then have a version with the bonuses play the version without the bonuses.
My hypothesis is that the version with the bonuses will perform much worse, just because of the size of the bonuses.
I'll let you know the results. Should be fun! :)
Cool! I did reduce king-to-the-center bonuses by 50%-80% but maybe it would be better to remove them altogether. Have fun! :)
