Improving TT replacement scheme

[algerbrex]

Ah, gotchu. Yep, long-time control testing would definitely be best, but unfortunately, I don't have the hardware or time to support such extensive testing

Or you could test at shorter time controls, but with only 1MB of hash.

Very true, not sure why this didn't occur to me.

[Kotlov]

I experimented a lot with a replacement and made a general conclusion that even the most cunning replacement schemes are slightly different from the simplest scheme “write always”

[algerbrex]

I actually went back and looked at the replacement code today and realized I had a bug where I was replacing the depth preferred entry first instead of replacing the always replace entry first. Basically had things backwards. Fixed the bug, ran another test, with a 1MB hash to put more pressure on the replacement scheme, and saw a 30 Elo gain in self-play.

Not sure if it's worth experimenting with 4 buckets or not, since two seems to work quite well now that I've ironed out the bugs.