Author: Thorsten Czub
Date: 02:39:17 10/17/00
Go up one level in this thread
On October 16, 2000 at 10:49:39, Robert Hyatt wrote: >>It works Bob. the main problem with CSTal (that CSTal was not that strong) >>wasn`t that Chris was a lousy programmer. He knew about chess and imlemented new >>ideas, he was also risky enough to do such a weird program, BUT his programming >>abilities were limited. >Here is my take: a lot of hyperbole, cast around some pretty neat evaluation >ideas, in an effort to make something into a legend before its time. WHEN is the time right ? CSTal is on the market for a couple of years. Isn´t it there, for anyone to buy ? > The >average consumer might buy into the "search toward unclarity" stuff, but not >"this consumer". I don't see any magic in this approach. Nor do I happen to >believe that searching into positions that are unclear is a smart thing to do. ok- i cannot convince you. i am different opinion. chess programs have to learn to take initiative. by doing a plan. if they cannot do so, they will lose. >As far as Tiger goes, I have not seen Christophe claim that he searches toward >positions that are vague. Rather, it is a combination of evaluation tuning to >exaggerate king safety terms, plus (maybe) some specific search extensions for >positions where certain evaluation characteristics are noted. right. christophe had a strong program, but the program played boring. it let the opponent control the action. this way you can hope for a draw. and you can make a win if the opponent makes a big mistake. but the chances to win are small. so he changed that. but one thing is important: he took the risk that his program would produce unsound moves, he participated with such a new thing in dutch and franch championships. >The first approach is more hyperbole than anything else. Because it doesn't >matter if you search into positions that are unclear, and then lose them. cstal is capable to win. and tiger much more. you will see. >Sure, >more speed or better search would help. But that is true for _all_ of us. right. > Yet >we have to make our compromises as to what is included and what is excluded, right, and chris and christophe have made DIFFERENT compromises than you. > to >maintain what we consider to be a search fast enough to avoid tactical pitfalls. >As far as wild play, SJLim on ICC played around with the tuning options in >Crafty, which let him turn it into a quite wild player. Or should I say wild >and unsound player. and - did it play stronger in wild ?! >The only thing I was pointing out is that if a program evaluates something as >+3, when material is even, then it will evaluate that same position as even, >when it is a piece down. no. when it is a piece down, it will normally have a weaker position. that means, the statical evaluations concerning its position will register it is lost and give malus score. normally you don´t have a piece less just coming out of nothing. so WHEN it is a piece down, it will normally give a very negative score, say -5. Only if it has a positional advantage of 3 pawns, and a piece down, it will score sero ! i see no problem in this. > It had _better_ be right. And from what I have seen, >most are not. Tiger might well play quite aggressively without making this >mistake. I only noted that in some games I watched (long games, not blitz) >its play would be called "risky". Kasparov has shown that risky chess can work. >Karpov showed that safe/sound chess could _also_ work. There is certainly room >for both. right. but we are not talking about playing style here. we are talking about ideologies when doing a chess program. >Particularly until it has been shown that "risky" can beat "sound" >more than it loses. we will see - yes. >>CSTal works. it would kill its enemies like hell when it would make more >>NPS and when the bugs would be out of the program. But the idea works. >>Its not important that the score changes. the score MUST change. its only >>important to win. cstal was not always able to win, but not because it played >>weak, it lost because it was outsearched. >>Gambit-Tiger is a program you cannot OUTSEARCH that easy. And it makes 8 x more >>NPS. >And you _really_ think that the algorithms used in Tiger are related to the >algorithms in CSTal? no - i say that christophe maybe got an idea what chris meant, because he tried out some interesting ideas in gambit-tiger. and maybe he likes the way gambit-tiger plays and continues his research. > I would make a wager that they are _not_ similar in any >way. Tiger is still a fast searcher, with a relatively simple evaluation >(relatively simple -> similar to the rest of the best). Just because you >happen to like its style of play doesn't mean it is something remotely related >to CSTal, IMHO. the relation to CSTal comes from the idea behind: a strong chess program needs to do something. it cannot wait until it had seen a material-advantage or a positional advantage in the tree. it has to do something. without this idea/plan what to do, it will be overplayed by people/programs who have this idea and play accurate. >>The paradigm is shiftet Bob, in a few years ALL chess programs will have to >>play like gambit-tiger and cstal. if they don´t do, they can never catch >>gambit-tiger. >This is an easy hypothesis to test. We only have to watch and wait. exactly. there will be a competition between the old-paradigm and the new-paradigm-programs. and this competition will force the shifting. > As time >goes forward, it favors the more accurate searcher, due to hardware speed >improvements. ???? why ? `the search program has to fight against the search tree. the non-accurate programs decide using evaluation, they don´t have this massive position-explosion that stops them coming deeper. i see no reason why ACCURATE searcher profit more from hardware than slow-searchers. i see it opposite. >Show me the solid evidence that shows it works. I played a zillion games vs >CSTal (on ICC) using equal hardware. (P6/200 single cpu at the time). It >didn't work then. 2 of every 3 games ended in an endgame. and 9 of every 10 >of those ended in a loss for the speculative program. In the games that didn't >reach an endgame, CSTal wasn't winning more than it lost. We even played games >with my king safety totally disabled. Which made Crafty also play quite >aggressively. And the search speed _still_ decided the game, even though one >program claimed to understand king attacks and the other had _nothing_ in the >evaluation about king safety at all. CSTal works. it plays chess. it scores points. and i is based on weird ideas. whatever you believe about those ideas. of course it cannot score each game. it was sold 1999 and made before 1999. If you give it at least a 400 Mhz machines and 40/120 time control it will produce good chess, and not lose. If you want i can present you with data that Rudolf Klamert has autoplayed and published in Gambit-Soft-Forum. >I've always been good enough to recognize things that work and try to >incorporate them into my program when needed, but I don't pay a lot of >attention to results obtained by beta testers, since _every_ new program >always produces significantly better results against the rest, until the >program becomes available where it can be studied a bit. ok - we wait until gambit-tiger is on the market. >First I don't think Tiger has _anything_ to do with "Chris w's crazy ideas". >Nothing at all. Other than an evaluation that is tuned to be speculative >when dealing with king safety terms. If chris would not have tried to convince people with his ideas, and not have come on the market with CSTal, things would have developed different. same counts for others too. >Second, let's see what happens when the program 'goes public' where everyone >can look at the positions and evaluations to see what is going on. I remember >a prior Nimzo version that was beating everyone in endgames. It later turned >out that it was doing so by using larger than usual scores for passed pawns. >Most adjusted and this advantage is now _zero_. ok - lets wait how the opponents will have to adapt. >In my case, I will believe that kingside attacks work when I see them work. >So far I have not seen them work. Tiger is simply a good program. I am not >yet convinced that the speculation is needed, nor that it makes it better. :-))) we will wait. bob. no problem. >I rely on exact calculation whenever I can, when I play chess. I am _certain_ >that Kasparov does the same. I doubt that against other GM players he casually >says "I play this, it seems very unclear". Rather, I have watched him calculate >for many minutes to be sure that he likes the position he is playing towards. exact calculation is not the sense of the game. the sense is to win the king, not to be right. >What if there is nothing to see? That has happened. if there is nothing to see - than defend against Rc6 ! but defend within the limit of time given by the time-control. otherwise it would be unfair - if you would use more time to defend than gambit-tiger needs to find the move, it would be unfair. >maybe or maybe not. If he offers pieces, and the attack fails, they are >not really "following" but leading, themselves. wrong. if he offers a piece, and you take it YOU have to show that the idea was wrong. it doesn`t matter the move is correct or not, it matters if you lose or win. if you defend wrong, although the sac was incorrect, than this is YOUR problem. >You are certain that wins? I am not. That is what I do with those >moves. ok. eat the pieces. that is what you do with it. i have no problem with seeing your programs eat anything because YOU don´t want to find out if it works or not. eat the pieces and pawns. that makes it easy. >I disagree. obviously. There are other ways to accomplish the same >thing. good. we will see. >A quote of "the old paradigm is dead, but reports of its death have been >greatly exaggerated. :-))) ok.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.