Page 1 of 1
Plain alphabeta speed
Posted: Tue Dec 23, 2014 1:29 am
by tetra
So, I just finished polishing my move generator and now that I have a very fast generator I moved to the search completely..
Common engines like stockfish crunch 20 depth in a matter of 1-2 seconds in my computer but my plain alpha beta with no improvements just can get to 7-8 depth in +/- 10 seconds.. Is this normal?
I also tested negascout without even understanding it and the speed improvements were small.. it is faster but not even 10 depth can reach in considerable time.. Are those normal results?
// tetrapack
Re: Plain alphabeta speed
Posted: Tue Dec 23, 2014 9:39 am
by H.G.Muller
That even sounds fast. Do you already have a Quiescence Search?
It also depends on what evaluation you have. E.g. with material only (and no Piece-Square Tables) alpha-beta becomes very efficient, because every move that is not a tactical blunder scores the same, which means that you almost always will pick the best move first, whichever one you pick.
Also important is whether you have a transposition table or not. How long does it take to solve Fine #70, for instance?
Re: Plain alphabeta speed
Posted: Tue Dec 23, 2014 1:21 pm
by tetra
H.G.Muller wrote:That even sounds fast. Do you already have a Quiescence Search?
It also depends on what evaluation you have. E.g. with material only (and no Piece-Square Tables) alpha-beta becomes very efficient, because every move that is not a tactical blunder scores the same, which means that you almost always will pick the best move first, whichever one you pick.
Also important is whether you have a transposition table or not. How long does it take to solve Fine #70, for instance?
Yes I do have a quiescence search and yes I do understand why with only material evaluation alpha-beta is fast. Because of that I made my evaluation function taking into account both material and mobility.. This are the score and speed I'am getting for the starting position, and like I said just plain alpha-beta with no tt, heuristiscs, etc:
Code: Select all
>>> think 9
depth 01 = 0,00 in 0,000013942s
depth 02 = 0,01 in 0,000118899s
depth 03 = 0,06 in 0,000517903s
depth 04 = -0,03 in 0,004138761s
depth 05 = 0,07 in 0,026121426s
depth 06 = -0,04 in 0,551037494s
depth 07 = 0,08 in 3,642081666s
depth 08 = -0,05 in 82,813356898s
>>>
If you say this is normal then I'am happy
But is it really?
What you mean "solve Fine #70"?? that would take too much time..
Re: Plain alphabeta speed
Posted: Tue Dec 23, 2014 4:44 pm
by H.G.Muller
I mentioned the Fine #70 position because it is a Pawn ending, for which fixed-depth is a good approach. But without TT it would be hopeless.
For comparison: micro-Max 1.6, which is also fixed-depth without TT, takes 5 sec to reach depth 7 in the opening position:
Code: Select all
dep score nodes time (not shown: tbhits knps seldep)
7 +0.14 13.2M 0:05.46 e2e3
6 0.00 1.64M 0:00.75 b1c3
5 +0.15 143800 0:00.06 b1c3
4 0.00 15572 0:00.00 d2d4
3 +0.15 1784 0:00.00 c2c4
2 0.00 142 0:00.00 c2c4
1 +0.15 22 0:00.00 c2c4
0 0.00 1 0:00.00 a8a8
So 10 ply doesn;t seem bad at all. It also depends a lot on what move-ordering devices you use. (E.g. PV move, killer, history, IID.)
Re: Plain alphabeta speed
Posted: Sun Feb 01, 2015 12:22 pm
by grandsmaster
I think it's actually quite fast for a plain-alpha beta, assuming you're talking about the initial position. More importantly, does the search return a move you're expecting?