Crafty - depth 30

[bob]

I don't get the "90 days".

I ran your position for about a minute and finished 22 plies.


Are you saying it took 22.7 90 days to reach depth 27?

Exactly, what I am saying. on a p3/900 though. I didn't want to interrupt the cpu time.

That would be 22.7, one cpu, default for hash and such???
I'll let my box run a while to see how long it takes to reach 27.

Single CPU for all tests... for the "newer" 3ghz computer I used hash96m & hashp32m with changes...

OK. On my box here, so far, one hour to finish depth 26 for the current version.

I can test when you want. I've tried a _ton_ of LMR changes that made the depth go faster, but every last one either offered no improvement, or hurt, when I played real games.

If a change hurt then it may be better to change to the opposite direction that means to make the depth go slower.

Uri

All of my tests are done like that. For every change I make, I try to locate the point beyond which the score starts to drop off, so that I can verify I am at a maximum point on that performance issue.

[bob]

OK. On my box here, so far, one hour to finish depth 26 for the current version.

I can test when you want. I've tried a _ton_ of LMR changes that made the depth go faster, but every last one either offered no improvement, or hurt, when I played real games.

You are probably right. I am just trying to think of a way to improve Crafty's play with an increase in depth. It would only make sense if you look deeper, you should have the advantage... it's how smart the move ordering and position evaluation that comes into play IMO...

It's a bit of an illusion, though. The more selective you make it, the deeper it will be able to see on *some* lines of play. But also, the more stuff it will miss on other lines of play that it chose not to search as deep.

That's the catch-22. If you can look deeper at better moves, you win, if you look deeper and miss the better moves you lose.

The number of ideas that make perfect sense, but then fail in testing is astounding... I've spent a couple of months on this LMR stuff and still have not found a thing better than what we have in 23.0.. I have found a couple of eval tweaks that have given us another 10-12 Elo. But no search changes so far have worked.