cases when X+1 beats X but not better than X

[Uri Blass]

I wonder if programmers found cases when it happened and what was the reason.

I suspect that it can happen when X+1 searches deeper at the price of tactical holes.

What happens is that games start when there is no tactics so X+1 get positional advantage against X in most cases thanks to searching deeper.

When X+1 has the advantage X has no good tactics to find so it cannot express it's relative advantage.

My latest late move reduction that was not based on history counters but only based on evaluation may be one of these cases.

Late move reduction beated no late move reduction convincingly but in tactical test suites it seems that late move reduction make the program weaker and irt is clearly possible that the late move reduction that I have is not productive in games.

Note that the first late move reduction that I tried does not include reducing checks or captures and I did not want to copy from glaurung.

I still believe that never reducing checks is not a good idea but it is probably better than what I do today.

Today the late move reduction in version 353 is simply based on the same complicated rules that I use for pruning when the remaining depth is 1.

Uri

[Dann Corbit]

There is a similar phenomenon in tuning the evaluation.

If you tune the evaluation for test suites, it will play less well in games.
If you tune the evaluation for games, it will perform less well in test suites.

[Uri Blass]

It is possible that the late move reduction that I use today is productive against other oponnets(I am not sure about it) but I clearly do not like the result in the arasan test suite and it is clear that the result in testing against no late reduction was too optimistic when I compare it against other opponents.

I read in the past claims that testing chessmaster personality against themselves can give misleading results against other opponents but I did not read logical explanation for the question why it can give misleading results.

I try to give some logical explanation.
I do not know if my explanation is correct and I did not do testing to prove that X+1 is not better than X against other opponents in my case.

Note that I do not think that optimizing for test suites is similiar because common sense tell me that optimizing for test suites is a bad idea for games when common sense does not tell me so clearly that A can beat B in games and still not be better than B against other opponents except cases of having not symmetric evaluation(for example evaluation that does not encourage trading queens in case that the program plays relatively better to other opponents when queens are on the board by giving the queen of the program higher value than the queen of the opponent).

Uri

[Dann Corbit]

In the ideal case, it will improve both game play and tactical play.

Perhaps some parameters can be experimented with to improve the overall performance

[Uri Blass]

I think that the change that is needed is more complicated than changing some parameters.

I am sure that some conditions can simply help it to solve more test positions and I plan to try them.

I plan to test not pruning big threats because I think that the main problem that cause it to be weaker in test suites are cases when a move that threats the king or threats another big piece is pruned.

Of course the idea that tord suggested already avoid pruning big threats because it does not prune in case that the move has a threat to move with the same piece again and it does not prune also in some other cases.

Uri