[ List Earliest Comments Only For Pages | Games | Rated Pages | Rated Games | Subjects of Discussion ]
Comments/Ratings for a Single Item
I think you misunderstand how engines work. Tree search is simply their way to make plan. That is by trial and error, discarding what does not work, rather than by constructive reasoning selectively producing only that which does work. But the result is exactly the same, and you should not be able to see the difference by watching it play. For Chess there are many different engines, with many different 'personalities'. Yet the search algorithm of every engine is almost exactly te same. It is the evaluation that creates the personality. But the search is necessary to distill strategic traits from transient 'noise' in the evaluation, and thus make the evaluation a meaningful guiding principle. E.g. like havinga passed Pawn on the 7th rank, (in most cases very good), which happens to be untenable and will be lost in 2 moves, making that you are simply a Pawn behind. If engines storm the opponent's King fortress with Pawns, it is because their evaluation has some non-linear term in its King Safety, which penalizes a King as the (say) square of the number of enemy Pawns close to it. They don't have the slightest idea that a checkmate might likely be the result. It is the programmer of the evaluation function that new that, and new that by the time the engine has the Pawns in place, it will likely see a way to push them that breaks down the opponent's fortress and leads to mate within its search depth. The problem with very good engines like Rybka or Fritz is that their 'plans' are completely beyond us mortals, so that we cannot recognize them as such. While the plans we would make in the same position ar in fact hopelessly futile attempts that they easily refute and blow to pieces. This is why it is much more fun to watch games between 'weaker' engines than Fritz. Btw, the 3 Alpaca vs 2 Knights match ended in 47.9% after 402 games. This is really within the margin for equality, so it seem that an Alpaca is nearly exactly 2/3 of a Knight. (i.e. 215 in Kaufman units, which has N=325.) The 2 Guanacos vs 2 Alpacas + Pawn was lost by the Guanacas with only a 37.4% score (434 games). This confirms a Guanaco is worth a lot less than an Alpaca + 0.5 Pawn, perhaps 235. I am now doing 3 Guanaco vs 2 Knights and 2 Alpacas vs 2 Llamas (all as Knight replacements).
I think I do understand how chess engines work. It's obvious that Larry better understands what I'm talking about. I gave the example of the rook maneuvre to the king side, in front of the friendly pawn chain. There is no way that any program, not even Rybka, can calculate any gain from that maneuvre. He cannot even back up the attack with another piece. So, within the horizon, it's a useless move. Nevertheless, he just positions it there and hopes that it can be of good use in the future. But in doing this he weakens the first rank, etc. But it doesn't matter that it's a bad plan. If the human doesn't play well then the computer can have good use of the rook position. A human player must take measures against it and try to find a refutation. This is the kind of chess that humans like to play, i.e. a game which isn't perfect, instead it is full of, perhaps, silly and refutable plans. But this allows room for creativity. Look at the games of Adolf Anderssen, for instance. Sometimes unscientific creative chess is called 'café chess'. It builds on the fact that a brutal pawn storm, for instance, can succeed although the positional criteria are unfavourable. After all, chessplayers do succeed in many games with bad plans and bad openings. It is because it's initiative and creativity that counts. Humans don't like the perfect and the clinically sterile. They like oil paintings of Henri Matisse and William Turner. They don't appreciate perfectly realistic sterile computer reproductions of reality. Such creations are dead. And so is every chessgame played by today's advanced computer programs. /Mats PS. How about the Vicuna and the Llama then?DS.
What you describe is purely a matter of evaluation. Such Rook manouevres (when they are pointless) is typically what you get when you award attacks on the King's Pawn shield by evaluation points. Because that makes attacking these pawns a goal in itself, and it will do it even when there is no follow up (i.e. when the plan was pointless). Without such evaluation award, it would only go for the manouevre if it did have a follow up that would be awarded in the evaluation (e.g. breaking down the Pawn shield, exposing the enemy King) AND it would reach the depth to see it. Awarding the mere attack on the Pawn shield makes that it will find the devastating King attack at much lower depth (and thus much further in advance at the same depth). But it has the problem that it also can backfire, when the plan is pointless, and would isolate a Rook that was needed badly in defense on the Queen side (e.g. to stop a passer break through). In this case it just leads to bad Chess. This is exactly the problem that I am facing in my new Xiangqi engine, HaQiKi D. To make it survive the middle-game against strong opponents, it is essential that it gives high evaluation points for Horses and Pawns appearing close to the enemy Palace. If it doesn't, it allows the opponent to amass his material for an attack, until it suddenly sees that there is a mate within the horizon in every branch, because its strategic position is so bad that there is no cure. But it also backfires badly: in end-games where it is behind, e.g. HaQiKi D has 2A+H, and the opponent has 2A+2E+2H, it might be able to draw by defending with 2A+H against 2H. (The opponent cannot use his 2A and 2E in attack.) Trading H vs H would be enough, as 2A vs H is a theoretical draw. But is loses all such games: his own H is drawn to the opponent's palace like a moth to a flame, in a completely futile attack on an impenetrable fortress, defended by 2A+2E, where it could not threaten a checkmate even against a bare King. And then of course his own King is toast, as 2A is no defense against 2H. And then it makes it even worse: in 'remedy' of attacks by the 2H on its own defending A, it start to counter-attack the opponent A or E, so that they are traded. Effectively, this indirect trade is like allowing the opponent A or E to cross the river and join in the attack, breaking don further what minimal defense there was. (After trading one A, the game is even lost even if you could trade Horses, as A vs. H is a theortical loss.) Incredibly stupid! Virtually every end-game from a slightly inferior position is lost by HaQiKi D this way. So allocating material to futile plans is not free: it is in fact losing Chess. Which means that it will disappear (in this simple Human-recognizable form) on deeper search, when the search corrects the initial misevaluation. I.e. when it can see the checkmate can no longer be avoided within the horizon after trading A, or the opponent's promotion cannot be stopped after isolating your Rook in front of your f,g,h-Pawns. If there is any fun to be had watching the engine make such futile attempts, it is only the fun of pitying the stupidity of your opponent when he does this.
No, I am discussing AI here, but you seem indoctrinated with the brute force search algorithm. If all chess programs weren't allowed to search beyond 4 ply, then programs would need to develop *intelligence* to be any good, and then computer chess would become interesting and fun again. They would begin to play cafe chess, like humans. /Mats
One way to influence a simple look-ahead is to establish goals within the game. For example(just spit-balling here), the creation of a particular Pawn structure. The engine would modify its evaluation of the play to include achieving this particular condition. Each of these 'goals' might be weighted for both priority and field condition. The engine can be easily 'tricked' by having the positive or negative conditions of these 'goals' affect its evaluation of each plys. The opposing player's achievement of these same goals can also influence the evaluation. There could even be the field position whereby the conditions and/or priority of these goals may change.
Well, unless you secretly changed topics, we are not discussing AI here at all. We were discussing Zillions. And Zillions works exactly as I decribed. Even Humans think far deeper than 4 ply; without that it is not possible to play any decent Chess, except perhaps at the level of a 6-year-old. But my main criticism still stands, and is totally independent of any mode of move productions: you admire moves that are futile, and call them 'a plan'. While in fact they are just bad Chess, by an entity that does not properly know what it is doing. You loath 'bean counters' for no apparent reason other than that they play good Chess. Well, for deriving piece values I need Chess of a reasonable quality, and the better the engine, the faster it can play to deliver that quality. If Zillions needs 30 min per game to reach the same quality as Fairy-Max has at 1 min per game, a 400-game run that I do in a day with Fairy-Max would take a month with Zillions. That is not doable, as you need mny such runs to derive the value of a single piece. The 3 Guanacas beat 2 Knights by 54.8%. Again spectacularly little better than 3 Alpacas. The small difference between these pieces continues to amaze me. I should do more tests with divergent pieces, to see if this is a general trait of non-capture slider moves. 2 Alpaca vs 2 Llamas ended at 50.6%. I forgot for whom, but that does not really matter as this is equality to far within the resolution of the test. I will stop this testing now for some time, as I have to test the opening book for my engine HaQiKi D, to get it ready in time for the Computer Olympiad in Pamplona (May 10-18).
There is another important factor than the non-capture slider moves. The Guanaco has very short capture range, and few capture squares. It is the combination of these factors that make the non-capture slider moves less useful. Had the capture range been larger, then the Guanaco would more often be capable of an forcing an exchange with a light piece. Hence, its value would approximate a light piece. It is probably the intrusive character of long capture range which increases the piece's value. If the enemy piece won't accept being exchanged, it would have to withdraw behind its own lines. Moreover, the few extra orthogonal squares that the Guanaco can leap to are easily controlled on the crowded Western board. After all, the leaping square needn't even be controlled. A piece needs only stand in the way and block the path, while it is not a capture square. So the enemy Guanaco is easily neutralized. Since you're working with Xiang Qi, I understand your interest in piece evaluation. After all, it's much more difficult to define the piece value in Xiang Qi because it changes very much with the position. Good luck in Pamplona! /Mats
You have to be careful with that kind of reasoning. If the Guanaca had more captures, it might very wel be worth more than a light piece intrinsically, and you would not even want to trade it for a light piece. There are some pieces for which your reasoning works, e.g. the Camel on an 8x8 board. this is basically a useless piece in the end-game; the side that has it will almost always lose it without compensation. The opening value seems to be purely derived by its forking power on a densely populated board allowing it to be trded for something else. (In the end-game you can often not even trad it for a Pawn...) The intrinsic value of a piece is usually determined by how well it cooperates with the King in positions with several Pawns, to protect its own Pawns and support their advance, and to gobble up the opponent's and stop their advance. This is how most Chess games end, so a piece that does well there has a really large impact on the average performance. Despite its fairly large number of posible targets, the Camel is totally inept in this respect: the targets are non-contiguous and mostly far apart, and if they are not actually off board, they are still too far away for useful manouevring, once there are ony very few 'centers of activity'. Hoppers might also best be gotten rid of before the late end-game, although a Cannon does remain dangerous even with the tiniest supporting material (e.g Cannon + Ferz makes a won end-game against a bare King). Grasshoppers definitely get useless very quickly. But the Guanaca is not particularly bad against Pawns. So I expect most of its value is intrinsic, and making it more powerful by adding moves would up its intrinsic value faster than the ease to trade it for B or N.
Should perchance anyone be interested, I have now implemented various
Guanaco variants in a 10x10 context, namely Mastodon Chess,
as for now only downloadable from the Zillions site:
Mastodon Chess (10x10)
I evaluated it to a knight's value on this big board, and had to tweak down
its value considerably.
/Mats
Guanaco variants in a 10x10 context, namely Mastodon Chess,
as for now only downloadable from the Zillions site:
Mastodon Chess (10x10)
I evaluated it to a knight's value on this big board, and had to tweak down
its value considerably.
/Mats
10 comments displayed
Permalink to the exact comments currently displayed.