I had forgotten this was on the web. Rereading it now, what strikes me is that virtually all bad pawn structures mentioned in the article carry much smaller penalties in Rybka than I implied in the article. There is a big disconnect here between database statistics and what works best in Rybka. I wonder why. It could be a difference between human and machine chess, or it could be that our testing at superfast speeds is the problem, or perhaps something else. Someday I'll figure out the answer.

Probably here : http://home.comcast.net/~danheisman/Articles/Articles.html