GeneM wrote on 07/05/14 at 21:05:19:
dfan wrote on 07/04/14 at 13:47:48:
Marc Benford wrote on 07/04/14 at 05:37:41:
the true evaluation of the starting position
My point is that there is no such thing. But I won't belabor it any longer, since the debate is hopeless.
Therefore there are probably many more legal chess positions for which there is no such thing as an "evaluation".
This begs the question DFan:
What make some chess positions subject to meaningful evaluation, whereas other chess positions are not subject to meaningful evaluation?.
I'd like to explain dfan's point, the thoughts behind which were also in the background of my earlier, satirical post. The existence of computer programs that spout centipawn evaluations and can beat you, and which furthermore are used to find tactical errors or opening weapons, makes some people believe in an objective or absolute true centipawn evaluation. In fact, such a thing is a contradiction in terms.
To see this, consider perfect play in the form of a tablebase, for example the seven-man tablebases just compiled and available by subscription at ChessOK. A computer (without a tablebase) may evaluate a given tablebase position as +.15, or whatever, when the tablebase has backsolved everything to either theoretical win for White, a theoretical draw, or a theoretical win for Black. These are not numeric outcomes. Even a theoretical draw is not expressible numerically: 0.00 only has a numerical meaning where other numbers are possible. Furthermore, in that context it merely expresses the expectation that White will gather half the tournament points, but says nothing of whether this will occur by drawing.
For the same reason, the king's being on the board has no point value. Other material values (for a knight or queen or pawn, etc.) are used for estimating the relative effectiveness of various material imbalances for the purpose of eventually mating, or preventing the mating, of the king. One cannot use the value of a lost king for this purpose, because it is illegal to have a lost king. Not even an infinite value will do, first because infinity is not a number and second because the object of the game is not to accumulate material advantage. And if your computer uses +1000 as a symbol for mate, that does not make mate exactly 1000 times as useful as a pawn. Mate and a pawn are things of entirely different types that are not comparable. Do not throw out your brain when looking at a computer! Game outcomes are not numeric.
So what
is the computer saying when it says White is up by .15? Isn't it saying that White has a positional advantage equivalent to .15 hundredths of a pawn? --
The computer is expressing something that only has meaning relative to imperfect play.
A person or a computer that does not play perfectly may play a given pawn-up position 100 times and win 79% of the points (if it's Karpov playing other grandmasters) or 54% of the points (if it's a child playing other children); on the other hand, a tablebase playing another tablebase in that same position will either win all the time, draw all the time, or lose all the time. For Karpov or the child, there is no such thing as a position with a 100% probability of winning, or of drawing or of losing; there are no certainties. For a group of tablebases or theoretically perfect players, there are nothing
but certainties. Every position will produce 100% draws, 100% wins, or 100% losses.
Now one might construct a table linking a given numerical advantage to a given average number of tournament or match points garnered, but such a table would only be valid for a given class of imperfect player. Its contents would have to be revised for another who plays differently. This alone is enough to tell you that numeric evaluations are not absolute truths. It is also why different engines' evaluations are indeed in different units--each engine is expressing the estimated outcome of its own treatment of the position.
Furthermore, the very idea of material value is only meaningful for imperfect players, and these meanings are equally dependent on which imperfect players are concerned. One can observe that Black has a queen for two knights, but whether this advantage is less or greater (in expected percentage of tournament points gathered) than that of three pawns depends on the players concerned and on how they handle pieces and pawns. For a tablebase, or for "theoretical" purposes, on the other hand, material advantages do not exist--only the position exists. All one can say objectively is that one side has a queen while the other has two knights and that the position is, for example, 100% drawn.
When MarcBenford jokes that the theoretical maximum ply are sufficient, but still expects a numeric evaluation there, and says that Fritz's units are closer to "true units" than Houdini's, he is contradicting himself. If an evaluation has units, it is not true, and if it is true, it does not have units.
By the way, it's impossible to remove material without creating positional ripples that are not material--and computers' numeric valuations of positional factors vary heavily, with Stockfish, for example often giving positional factors higher material equivalents than Rybka. So much for the experiment of deleting the c-pawn.
--Finally, I would suggest that some posters would get a friendlier response if they did not give orders, or presume to give permissions, to the rest of us.