Prima wrote:Perhaps RH can combine both forms of learning: the learn format suggested by you and the other type of learning he originally intended. Who knows?...
Combination would mean Houdini stores everything into the permanent learn file, even things it doesn't really need, for positions you're not even going to visit. And the learn file would grow very huge very fast, with my suggested method a 16MB file can store all the learning for years before an entry needs to be overwritten, if you combine the methods then you'd need to have a file in there the size of your RAM, that doubles every time you visit a new (non-transposing) position or new game (for the times where the hash is 100% full), so in a few weeks the learn file can be several gigabytes large, full of information you're not really going to use (unless you're going to try to solve chess or something).
I think the concepts can remain separate just fine, the method I suggested focuses on making the fewest reads/writes into the learn file as possible, and into having a small "master file" for all analysis. RH's method focuses on continuing analysis from where it left off, and in having a file the size of RAM for each game you analyze (because you don't need positions of the first game using space of the hash of a second game).
They really have different purposes and no reason for combination, though the user has to choose which one to use (people leaving the engine analyzing to very high depths would prefer his method, people highly interacting with the position and analyzing as many as possible would like to stick with what I suggested).
