Humans v Houdini chess engine (Elo 3300)

Sort:
kingofwhite
LegoPirateSenior wrote:
PrawnEatsPrawn wrote:

Humans have no chance against good software/good hardware combinations, at any time control.

PEP, in the position below, white has a completely, 100% winning move. A human player found it OTB. I am curious how long will it take for your setup to find it.

 

Can you please give more details and analysis on the position posted? Yace for example finds Qd5 right away, but it doesn't look like it's won in any way for white.

What are the next moves? If you actually did the analysis to back up what you said it will be easy to post...

After Qd5 and exd5, how do you secure that win? Rf1 doesn't seem to do anything, really... let's play it move by move here... you tell the winning move and I tell the reply.

All this does is draw...

pfren

For the record, I have beaten Houdini 2.0 a couple of days ago in a correspondence game... Tongue out

It was drawn position, because my opponent (playing white) had built a fortress, where I had all the play, but I could not find a way to break in- most probably there was none.

So, I played a move where white could play more actively and "better" (Houdini gave an evaluation difference of 0.3), but the fortress would be broken. My opponent (apparently not a strong player) trusted Houdini's recommendation, followed his recommended line, and lost rather easily.

I may post the game on my blog.

LegoPirateSenior

@kingofwhite - I think you made a typo: Qd5 does not win. 24.Qxe5 does, and after fxe5, 25.Rf1 indeed is necessary. 

I mostly relied on analysis that can be found on the web.  I've just added two relevant links to the game page: http://www.chess.com/games/view.html?id=53873 

When analyzing this position with an engine, I recall that you need to play maybe another 10 plies into the game before the engine sees a win for white. Backtracking will then maintain a winning evaluation until you run out of hash.

-Est-

the best way to find engines weakness is by playing the kingindian. it evaluates it around +0.5 since white has space advantage + the center, but as u go through the middle game it often changes it's evaluation when black attack gets stronger. ofcourse the engines dont understand some positional things and that is why they beat each other! i am 100% sure that i am much stonger than any engine if i use my reybka. it is highly unlikely but possible for a human to beat an engine in correspondence but i am sure it would never happen OTB since noone would be able to be that accurate. i am pretty sure the best player now would not be able to win Houdini down a tempo OTB.

kingofwhite

Mr. Pfren, you are saying yourself that "Houdini gave an evaluation difference of...", which means you were using it as well.

It's not the first time you misinterpret the topic: it's human vs. machine, not human+machine vs. machine. I have to assume this is voluntary because I credit you with a lot more logic than that.

With time to test for tactical errors (with a machine), the "inequality" human + machine > machine proves that (I know, it's oversimplifying) human > 0.

You have to acknowledge the true topic and let it rest. And the fact that GMs would only take on an engine for financial motivation... if they could easily beat engines they would do it to shut everyone up. If chess engines were so bad, the #1 correspondence player would not have given up his hobby due to this phenomenon. If correspondence chess was so great, how come we find all sorts of errors in older correspondence games?

Please, just please...

pfren

Mr. kingofwhite, I think you understood what I've said perfectly well: I said that Houdini sometimes, in certain positions, "commits" serious errors, simply because he's unable to grasp some positional elements which are rather apparent, even to a mediocre player like myself- and this cannot be compensated by their near-perfection to combinational play.

And yes, I do consult engines (more than one, as their evaluations can vary wildly) at LSS correspondence chess, mainly to "predict" my opponent's next move (the wide majority of them just copypaste Houdini's best lines). This has earned me quite a few easy points in the past.

For the record, my favorite engine is Critter. It trails Houdini 1.5/2.0 very slightly, but its suggestions have much more "human touch", they simply make more sense to me.

ElKitch

I read the whole topic and watched the games posted. Nice and interesting.

Dodger111
PrawnEatsPrawn wrote:

For once, pfren is well off target.

 

Humans have no chance against good software/good hardware combinations, at any time control.

All this talk that engines don't understand positional play is simply out of date nonsense.

 

Yep....the top programs have been beating the big boys for some time and winning by bigger margins as time goes by. It won't be too long before they are unbeatable.

pfren

Oh boy, some people are unable to understand simple things...

Here is the position from my game- it's white's turn to play.

Feed it to Houdini 2.0c or any other engine, and let him think for as long as you please.

After calculating "everything" he will come up with either Ne1 or Kh2, with a close to equal evaluation - but both moves are losing. Factly, Kh2 isn't losing either, as long as white follows the "inferior" plan of keeping the knight passively on g2.

The correct move is either Rc1 or Ra1, which is regarded as way inferior by Houdini, but it draws instead: Black cannot break into white's fortress, despite the fact white having effectively no white squared control at all.

himath2009

Well done, pfren, thank you for this...

Dodger111

From a 2010 article chronicling GM's versus PC chess programs:

"

With the advancement in technology, computer programs are increasingly become successful against the grand masters. In 1998, a computer called Rebel 10 defeated the then world No 2, Viswanathan Anand by a margin of five games to three. In the early 2000’s, commercial computer programs like the Fritz and the Junior were able to draw matches against the likes of Kasparov and Kramnik. In 2003, Kasparov went ahead with another two exhibition matches with the computer program Junior and later with X3D Fritz, both of which ended in draws. Subsequently, from the year 2005, computers started emerging as the undisputed champions commencing from the defeat of Michael Adams. In 2006, the Deep Fritz handed a crushing defeat to the then Classic champion Vladimir Kramnik."

wiki also covers some GM versus different PC program matches, their performance is certainly nothing to scoff at:

http://en.wikipedia.org/wiki/Human%E2%80%93computer_chess_matches

Here's some more GM versus compter match results:

 http://gambit.blogs.nytimes.com/2008/10/01/man-vs-computer-match-ends-in-victory-for-man-but-with-a-catch/

Meilan1
pfren wrote:

Oh boy, some people are unable to understand simple things...

 

Here is the position from my game- it's white's turn to play.

Feed it to Houdini 2.0c or any other engine, and let him think for as long as you please.

After calculating "everything" he will come up with either Ne1 or Kh2, with a close to equal evaluation - but both moves are losing. Factly, Kh2 isn't losing either, as long as white follows the "inferior" plan of keeping the knight passively on g2.

The correct move is either Rc1 or Ra1, which is regarded as way inferior by Houdini, but it draws instead: Black cannot break into white's fortress, despite the fact white having effectively no white squared control at all.

I set up Droidfish (stockfish on android phone) with this position comp. vs. comp., white moved Kh2 and eventually lost. I then moved Ra1 and let the comp. take over, and the game ended in a draw by 50 move rule.

Interesting example of Man vs. Machine.

Dodger111
Meilan1 wrote:
pfren wrote:

Oh boy, some people are unable to understand simple things...

 

Here is the position from my game- it's white's turn to play.

Feed it to Houdini 2.0c or any other engine, and let him think for as long as you please.

After calculating "everything" he will come up with either Ne1 or Kh2, with a close to equal evaluation - but both moves are losing. Factly, Kh2 isn't losing either, as long as white follows the "inferior" plan of keeping the knight passively on g2.

The correct move is either Rc1 or Ra1, which is regarded as way inferior by Houdini, but it draws instead: Black cannot break into white's fortress, despite the fact white having effectively no white squared control at all.

I set up Droidfish (stockfish on android phone) with this position comp. vs. comp., white moved Kh2 and eventually lost. I then moved Ra1 and let the comp. take over, and the game ended in a draw by 50 move rule.

Interesting example of Man vs. Machine.


Sure  computers do indeed lose to top notch players.

Just not that often anymore.

In ten years probably not at all.

 

What's your point?

Elubas

Pfren, if it's played out, the computer does eventually change its opinion right? It's kind of like the computer admitting "ok, never mind, you were right." Smile

Still, I swear, there are times where it almost seems like Houdini is in fact following a plan, even though it isn't. A lot of times it will recommend consistent moves -- say for queenside expansion: b4-b5, Rc1 -- it will often build on a particular side of the board like a human would. It often centralizes its pieces, even maneuvers its knights, even when it doesn't attack anything. So I just want to point out that I agree entirely with your point, but also think that sometimes, despite the computer's lack of plan, it is so strong in other areas that it can often play moves consistent with a strong human plan anyway! In other words, it might find the same sequence of moves as a grandmaster, just in a different way. One can of course mention the exceptions -- a computer can't play everything perfectly -- but in my opinion, those exceptions are rather rare, but have, sometimes, occurred in my experience.

pfren

Yes, it does change his mind "later", but this is of no use- damage already done. Going 40-ply from the diagram position is obviously going to take a couple of weeks, even in fairly modern hardware- so impractical, even for correspondence chess.

The above position happened when I, as Black, opened the h-file prematurely (Black had a pawn on h5 and white on h2, plus one more rook each), where Black should eventually come out with a winning plan. But I also committed the very same error: I trusted the engine evaluations, without realizing the fortress plan. It was only when the 5-6 "best" lines all came to the very same evaluation (very favorable to Black) when I realized that there was something wrong, white could set up an inpenetrable position, and Houdini/Critter/pfren had in reality blundered half a point... Undecided

better_than_morphy

Isnt there something called anti computer chess or something???

hankas
better_than_morphy wrote:

Isnt there something called anti computer chess or something???

There is, but I don't think it works well against today's computers though. Basically, it involves closing the position and slowly outmanouvering the computer.

Yereslov

The rating is theoretically 3300 ELO. 

The rating largerly depends on the software and the amount of time the software takes to process.

Yereslov
pfren wrote:

Yes, it does change his mind "later", but this is of no use- damage already done. Going 40-ply from the diagram position is obviously going to take a couple of weeks, even in fairly modern hardware- so impractical, even for correspondence chess.

The above position happened when I, as Black, opened the h-file prematurely (Black had a pawn on h5 and white on h2, plus one more rook each), where Black should eventually come out with a winning plan. But I also committed the very same error: I trusted the engine evaluations, without realizing the fortress plan. It was only when the 5-6 "best" lines all came to the very same evaluation (very favorable to Black) when I realized that there was something wrong, white could set up an inpenetrable position, and Houdini/Critter/pfren had in reality blundered half a point...

Which position? I must have missed it.

Yereslov
pfren wrote:
CharlieFreak wrote:

It's weird the way they always win though . . .


When? To my poor knowledge, the last time a strong corr. player met a computer, the human won 2-0, but this is some five years ago.

The reasons that such matches are not held is rather simple: The programmers are well aware of the limitations of their engines, and won't put up such a challenge before improving  on them.

ELO points mean nothing for that matter- their statistical model was tuned for different sorts of events. On ICCF championships, the highest ELO currently is 2729, which is way below your 3300 dream. But this is the highest a super modern engine under skillfull human aid has achieved.

How well Kramnik, or any other strong GM can play without computer aid at long time controls? Logic says more than that, factly they would beat the machine without too much effort. Chess is not a bunch of zeroes and ones which should be counted as fast as possible, this is the truth- whether you like it, or not.

But of course, the first thing the silicon lovers will think when the GM beats their beloved Dumbini is that the GM had computer aid...

That match was played seven years ago.

Technology has improved by a longshot since then. 

Bronstein also beat a computer in the sixties.