Well you see, for Larry as a grandmaster and senior champion, Larry does a lot of his opening preparation and study also with Komodo, he has stated since his Rybka times that he finds this correlation between human evaluations and computer evaluations very important, he and Don work hard at it in Komodo.
Stockfish on the other hand, to name another program, is developed with very little of this kind of input, and we don't have grandmasters on the team, as far as I know. Endusers have to really get used to its evaluations being a bit different from at least Rybka, Houdini and Komodo I think. But for the program itself, it only becomes a problem if its (mis)evaluations would make it give up any long term advantages. The material calculations in the short term, apart from horizon effects are never a problem, compared to most human play I mean. It is true that some positional characteristics are really very hard to judge for most programs, and maybe the commercials are working very hard on that. But as long as the program would scale all its positional evalterms the same way (in relation to each other), at least roughly, it does not really matter if it seems in general overestimating or underestimating things, as long as it does not give up good long term traits for worse ones. So for all evaluations staying under a pawn, giving up a pawn too soon because of misevaluations is not an extreme risk. And some of the other long term traits, well the program simply does not understand those anyway
.
And for the people who have studied Fruit it is not a secret: some of the remaining misevaluations will disappear, as by magic, as the game progresses, solving some of the problems literally automatically. That is because since the time of Fruit, but I think Fruit was not really the first to do this, all the eval terms have both a middlegame and endgame component. And based on some, maybe crude, but effective, rules depending mostly on the material situation, every term is then scaled between those two components as the game inevitably progresses towards less material, towards the endgames. An opening book helps with the very early stages. Of course it remains extremely important to tune everything for maximum effectiveness and minimum blunders/horizon effects, but that already does not have much to do anymore with how the grandmaster looks at a position, how all the things he has learned are balanced out on the board. The total sum of the terms is about the same but the grandmaster adds different terms to come to his conclusion than the computerprogram. In the end the silicon grandmasters are still mostly counting beans in the eyes of human topgrandmastters I think.
Hope this gives some programming perspective to your question!
Regards, Eelco
