lech wrote:Thanks for help.
Stockfish can be improved. Each position may be considered in 4 ways (combinations of: white<->black and King<->Queen). Results are different e.g. 138, 2, 49, 85 seconds to find solution for the same (in fact) position. It gives huge possibilities to improve Stockfish.
It needs a team work.
Can you give an example position where mirroring (which is what I understood from your post) leads to such different solution times?
So what are you comparing exactly? What was the termination condition for these four searches? Why does the output for the second search terminate after 10 iterations already? What I can see is that in all shown cases the solution (best move) was found much earlier than in the last iteration for which you reported its output.
The time used for finding the best move of iteration 10 is roughly the same in all four cases. However, you are right that there are differences, also for later iterations, and also in the evaluation that is displayed.
One possible explanation could be the order of move generation, e.g. left to right for pawns, although I would not expect a huge influence by that. Other explanations should indeed be given by the Stockfish team. But I don't have the impression that there is some major problem.
lech wrote:Thanks for help.
Stockfish can be improved. Each position may be considered in 4 ways (combinations of: white<->black and King<->Queen). Results are different e.g. 138, 2, 49, 85 seconds to find solution for the same (in fact) position. It gives huge possibilities to improve Stockfish.
It needs a team work.
The order of generated moves differs when you mirror / flip the position, so the search cannot be exactly the same for all the positions.
For instance suppose you generate 5 pawn moves at a given position, the search can choose to follow the first 2 ones and prune the last 3.
Now if you change the order it can happen that the 3 pruned moves are different from the previous case and you may end up pruning a good key pawn push that could lead quickly to a very good position so that in this unlucky case you have to search much more to find the correct answer.
Of course on average the lucky cases should match the unlucky ones, but on single cherry-picked positions you _could_ experience this behaviours.
lech wrote:Thanks for help.
Stockfish can be improved. Each position may be considered in 4 ways (combinations of: white<->black and King<->Queen). Results are different e.g. 138, 2, 49, 85 seconds to find solution for the same (in fact) position. It gives huge possibilities to improve Stockfish.
It needs a team work.
The order of generated moves differs when you mirror / flip the position, so the search cannot be exactly the same for all the positions.
For instance suppose you generate 5 pawn moves at a given position, the search can choose to follow the first 2 ones and prune the last 3.
Now if you change the order it can happen that the 3 pruned moves are different from the previous case and you may end up pruning a good key pawn push that could lead quickly to a very good position so that in this unlucky case you have to search much more to find the correct answer.
Of course on average the lucky cases should match the unlucky ones, but on single cherry-picked positions you _could_ experience this behaviours.
Marco, thanks for your explain. I tested many positions (middle-game too). Since differences can be that big (time and quality), I thought it is worth to look at this. Many of us use engines only to analyze.
