An interesting search reduction!

[Michael Sherwin]

If an engine maintains a fifty move rule counter then there may be an interesting reduction that it can be used for. When remaining depth is small and fiftyCount - n > ply then reduce. The idea is that when there has not been a permanent change of position move made in a certain number of game plies plus search plies then more moves of the same are not likely to be of any use.

[Dann Corbit]

Have you tried it, and if so what gain did you see?

It seems to me that it might be a win for some positions but a loss for others.

[Michael Sherwin]

Have you tried it, and if so what gain did you see?

It seems to me that it might be a win for some positions but a loss for others.

I am trying it now! I have only tested a couple of variations of it to get a feel for it. And it seems very interesting. The 'feel' is this. In almost every game there are several points where neither side has made any moves that have permanently changed the position and in those cases RomiChess depth of search starts to go up substantially and RomiChess is then able to find interesting ways to change the position to her benifit.

It is still too early to start claiming an improvement. However, it looks very promising!

[BubbaTough]

This is a good (as in interesting) idea. I would guess most of the time it neither helps nor hurts, but when it is active (such as during piece shuffling) it helps. Since piece shuffling is pretty annoying, this change may have more esthetic benefits than strength benefits.

questions:
a) how are you testing its effectiveness? Since it does not come up that much, I doubt play-testing would find an ELO difference very quickly. Maybe a library of starting positions after x non-pawn non-capture moves has occurred as starting positions?

b) how many ply of non-pawn non-capture moves are you using before reductions start?

c) do you use other interactive eval stuff? Such as drawing score closer to 0 when number of non-pawn non-capture moves becomes high?

-Sam

[Michael Sherwin]

This is a good (as in interesting) idea. I would guess most of the time it neither helps nor hurts, but when it is active (such as during piece shuffling) it helps. Since piece shuffling is pretty annoying, this change may have more esthetic benefits than strength benefits.

questions:
a) how are you testing its effectiveness? Since it does not come up that much, I doubt play-testing would find an ELO difference very quickly. Maybe a library of starting positions after x non-pawn non-capture moves has occurred as starting positions?

b) how many ply of non-pawn non-capture moves are you using before reductions start?

c) do you use other interactive eval stuff? Such as drawing score closer to 0 when number of non-pawn non-capture moves becomes high?

-Sam

a) just playing some games and watching them to get the 'feel' of it. No rigoris testing started yet.

b) trying various amounts of presearch sequences plus whatever the search adds. Hunting for a formula.

c) no.

[Bill Rogers]

I found something interesting with my own chess program and that is the chose move that is finally made so far has always been in the top three moves of my move generator. I will do more testing but if this holds true even though I might generate say 40 moves I only have to test the top three and that can be a tremendous saving in time.
Bill

[Dann Corbit]

I tried that (because I made the same observation). It works great except when it is a disaster. There are times when the only saving move is not one of them.

However, a sliding reduction for moves outside the "chosen band" can be a good idea.

Try it, you may like it.