LMR differences in long vs short games

[jacobbl]

I've been testing 2 different versions of my LMR.

1. Reduce if number of moves > 3 and OK top reduce (checks, captures...)
2. Reduce if number of moves > 6 and depth > 3 and OK top reduce (checks, captures...)

In short games (40 moves in 1 min) version 1 is better then 2, but I'm not sure if this holds for longer matches. I don't have the computer power to get statistically significant results in 40 min matches.

Does anyone have an opinion if it's better to do tougher reducing in short vs long games?

Regards Jacob

[bob]

I've been testing 2 different versions of my LMR.

1. Reduce if number of moves > 3 and OK top reduce (checks, captures...)
2. Reduce if number of moves > 6 and depth > 3 and OK top reduce (checks, captures...)

In short games (40 moves in 1 min) version 1 is better then 2, but I'm not sure if this holds for longer matches. I don't have the computer power to get statistically significant results in 40 min matches.

Does anyone have an opinion if it's better to do tougher reducing in short vs long games?

Regards Jacob

In general, the more "error" you introduce in shorter games, the more dramatic the impact. In longer games, the error can get washed away by the sheer size of the tree. In all of my testing, I have not found (yet) a clear example of where a specific "thing" helps more at long time controls than at short. I have seen examples where Program A plays better at shorter or longer time controls than it does at the other, however.

[jacobbl]

If I understand you correctly you are saying that version 1 is introducing more "error", and if that was negative it should have a larger impact on short games. Since the effect is positive it should also have a positive effect on long games, but maybe a smaller effect?

[bob]

If I understand you correctly you are saying that version 1 is introducing more "error", and if that was negative it should have a larger impact on short games. Since the effect is positive it should also have a positive effect on long games, but maybe a smaller effect?

The idea is that the more restricted the tree, the more error terms affect the root. If you saw the hashing collision study I did a couple of years back, it is truly remarkable how many errors you have to introduce into the tree search before you begin to affect the root move, much less to actually change it.

[jacobbl]

So what you are saying is that it is better to do tough reductions on long games than on short games, because errors in long games get "washed out" in a large tree? So it should be safe for me to use the tough reductions on the long games?

I once tried to add a random number to my eval function, and was quite surprised to see that even a random number up to 300 centipawns didn't make the engine play too bad.

[bob]

So what you are saying is that it is better to do tough reductions on long games than on short games, because errors in long games get "washed out" in a large tree? So it should be safe for me to use the tough reductions on the long games?

I once tried to add a random number to my eval function, and was quite surprised to see that even a random number up to 300 centipawns didn't make the engine play too bad.

That has been discussed here at length. I have labeled it "The Beal Effect" since Don Beal first reported on random eval tests in a JICCA paper years ago...