LMR at root of search tree - worthwhile?

[bob]

I got small improvement from it.
Caution : Root move ordering could be messed up if you use nodes count to sort the moves. I had this bug http://talkchess.com/forum/viewtopic.ph ... ht=#334554
Sorting on the cummulative nodes count from all the previous iterations (instead of just the previous one) solved this problem.

Note that the "L" in LMR means "late". I would likely not want to reduce the second move I search, which could be the previous best move.

When I get back to my office (I am out of town at the moment) I have some test runs on the cluster where I tried two different ways of dealing with LMR at the root. I played with this idea a couple of months back and think that one of the two versions was worth +50 Elo over the currently released 23.2, I'll have to collect all the cluster data and will post it either this weekend or on Monday. I have so many different sets of data it will take me some searching to figure out which version had only this change in it.

[Daniel Shawul]

When I tried to restrict LMR before nth move search at the root, I got worse results . That was actually counter-intuitive because restricting at PV nodes elsewhere always worked for me. However I did not do exhaustive test with different n valuse so it could totally be due to luck. It would be interesting to see your test results.

Daniel

[bob]

When I tried to restrict LMR before nth move search at the root, I got worse results . That was actually counter-intuitive because restricting at PV nodes elsewhere always worked for me. However I did not do exhaustive test with different n valuse so it could totally be due to luck. It would be interesting to see your test results.

Daniel

I don't think I settled on "n" but rather on the node counts. I'll look at the code and post an explanation along with the data...

[silentshark]

I got small improvement from it.
Caution : Root move ordering could be messed up if you use nodes count to sort the moves. I had this bug http://talkchess.com/forum/viewtopic.ph ... ht=#334554
Sorting on the cummulative nodes count from all the previous iterations (instead of just the previous one) solved this problem.

Note that the "L" in LMR means "late". I would likely not want to reduce the second move I search, which could be the previous best move.

When I get back to my office (I am out of town at the moment) I have some test runs on the cluster where I tried two different ways of dealing with LMR at the root. I played with this idea a couple of months back and think that one of the two versions was worth +50 Elo over the currently released 23.2, I'll have to collect all the cluster data and will post it either this weekend or on Monday. I have so many different sets of data it will take me some searching to figure out which version had only this change in it.

wow +50 ELO - worthwhile, then

although depressing for those of us whose engines struggle to get near to 20% vs crafty, when they used to be much closer!

[bob]

I got small improvement from it.
Caution : Root move ordering could be messed up if you use nodes count to sort the moves. I had this bug http://talkchess.com/forum/viewtopic.ph ... ht=#334554
Sorting on the cummulative nodes count from all the previous iterations (instead of just the previous one) solved this problem.

Note that the "L" in LMR means "late". I would likely not want to reduce the second move I search, which could be the previous best move.

When I get back to my office (I am out of town at the moment) I have some test runs on the cluster where I tried two different ways of dealing with LMR at the root. I played with this idea a couple of months back and think that one of the two versions was worth +50 Elo over the currently released 23.2, I'll have to collect all the cluster data and will post it either this weekend or on Monday. I have so many different sets of data it will take me some searching to figure out which version had only this change in it.

wow +50 ELO - worthwhile, then

although depressing for those of us whose engines struggle to get near to 20% vs crafty, when they used to be much closer!

Yes, but remember that this is not about doing LMR at the root giving +50, it is about doing it "better". version 23.2 was simply reducing some things at the root that should not have been reduced. I am still trying various options of this as it is unlikely that my first "cut" was optimal...

[silentshark]

I got small improvement from it.
Caution : Root move ordering could be messed up if you use nodes count to sort the moves. I had this bug http://talkchess.com/forum/viewtopic.ph ... ht=#334554
Sorting on the cummulative nodes count from all the previous iterations (instead of just the previous one) solved this problem.

Note that the "L" in LMR means "late". I would likely not want to reduce the second move I search, which could be the previous best move.

When I get back to my office (I am out of town at the moment) I have some test runs on the cluster where I tried two different ways of dealing with LMR at the root. I played with this idea a couple of months back and think that one of the two versions was worth +50 Elo over the currently released 23.2, I'll have to collect all the cluster data and will post it either this weekend or on Monday. I have so many different sets of data it will take me some searching to figure out which version had only this change in it.

hey bob, any chance of posting those cluster results?

[bob]

I got small improvement from it.
Caution : Root move ordering could be messed up if you use nodes count to sort the moves. I had this bug http://talkchess.com/forum/viewtopic.ph ... ht=#334554
Sorting on the cummulative nodes count from all the previous iterations (instead of just the previous one) solved this problem.

Note that the "L" in LMR means "late". I would likely not want to reduce the second move I search, which could be the previous best move.

When I get back to my office (I am out of town at the moment) I have some test runs on the cluster where I tried two different ways of dealing with LMR at the root. I played with this idea a couple of months back and think that one of the two versions was worth +50 Elo over the currently released 23.2, I'll have to collect all the cluster data and will post it either this weekend or on Monday. I have so many different sets of data it will take me some searching to figure out which version had only this change in it.

hey bob, any chance of posting those cluster results?

I am working on re-running them. If I happen to use the same version number for a second run, the first gets overwritten and the pgn is lost. I've started the runs again...