testing prunning with test positions

[adieguez]

Hi, is it effective test prunning with test positions?
I have an idea about fabricating test positions with the same engine. I have yet to test since my time is limited. Maybe you have tried it. First you use a version of your program that do less prunning, run a lot of positions and you collect the positions and best move and time, of a certain iteration, but only if the best move changed, and also is the best of the final iteration.
You use those as test positions for versions of your program that use more agressive prunning. A first problem is, that since iteration X, you will have to accept the position as solved if the program show the corresponding best move in any moment, which is a big problem because the program with more agressive prunning could change his mind very often and that will give it points. But that can be solved if we in the first procedure make exigible the best move is also the best on iteration x+1 for example(or even x+2). Well maybe this i snot usefull I don't know, but this looks like test positions you can make if you want, not just tactical.

I was also thinking (maybe I should go to bed) is there something hidden there in minimax that we don't see? for eval parameters testing what happens if we recollect a lot of positions in which we know white has a small advantadge, and when testing an eval parameter change we see if the eval is more maximized? nonsense perhaps lol, have a good day. I want a cluster like profesor Hyatt.

[Dann Corbit]

Hi, is it effective test prunning with test positions?
I have an idea about fabricating test positions with the same engine. I have yet to test since my time is limited. Maybe you have tried it. First you use a version of your program that do less prunning, run a lot of positions and you collect the positions and best move and time, of a certain iteration, but only if the best move changed, and also is the best of the final iteration.
You use those as test positions for versions of your program that use more agressive prunning. A first problem is, that since iteration X, you will have to accept the position as solved if the program show the corresponding best move in any moment, which is a big problem because the program with more agressive prunning could change his mind very often and that will give it points. But that can be solved if we in the first procedure make exigible the best move is also the best on iteration x+1 for example(or even x+2). Well maybe this i snot usefull I don't know, but this looks like test positions you can make if you want, not just tactical.

I was also thinking (maybe I should go to bed) is there something hidden there in minimax that we don't see? for eval parameters testing what happens if we recollect a lot of positions in which we know white has a small advantadge, and when testing an eval parameter change we see if the eval is more maximized? nonsense perhaps lol, have a good day. I want a cluster like profesor Hyatt.

The idea of alpha-beta is that it prunes a lot of positions and gives the exact same answer as mini-max using nodes proportional to the square root of the mini-max nodes.

Unfortunately, this is the only technique I know of that can provably produce the exact same result at the same depth with less nodes 100%.

All the other methods (that I know) sacrifice something to get the same answer in fewer nodes. Null move pruning is one of these ideas that works well in practice.

LMR is very popular now. But I guess if you add null move and LMR and search the same depth, you will often get different answers.

I think the bigger problem is not necessarily finding the same answer in less nodes for tactical problems but rather for quiet problems.

Consider that after 30 captures we are down to bare kings. So many, if not most good chess moves do not involve a capture.

How do we pare down the choices for these quiet moves? This is the more difficult and more important question.

IMO-YMMV

Will we be seeing a new "just amyan" message soon?

[michiguel]

Hi, is it effective test prunning with test positions?
I have an idea about fabricating test positions with the same engine. I have yet to test since my time is limited. Maybe you have tried it. First you use a version of your program that do less prunning, run a lot of positions and you collect the positions and best move and time, of a certain iteration, but only if the best move changed, and also is the best of the final iteration.
You use those as test positions for versions of your program that use more agressive prunning. A first problem is, that since iteration X, you will have to accept the position as solved if the program show the corresponding best move in any moment, which is a big problem because the program with more agressive prunning could change his mind very often and that will give it points. But that can be solved if we in the first procedure make exigible the best move is also the best on iteration x+1 for example(or even x+2). Well maybe this i snot usefull I don't know, but this looks like test positions you can make if you want, not just tactical.

I was also thinking (maybe I should go to bed) is there something hidden there in minimax that we don't see? for eval parameters testing what happens if we recollect a lot of positions in which we know white has a small advantadge, and when testing an eval parameter change we see if the eval is more maximized? nonsense perhaps lol, have a good day. I want a cluster like profesor Hyatt.

The idea of alpha-beta is that it prunes a lot of positions and gives the exact same answer as mini-max using nodes proportional to the square root of the mini-max nodes.

Unfortunately, this is the only technique I know of that can provably produce the exact same result at the same depth with less nodes 100%.

All the other methods (that I know) sacrifice something to get the same answer in fewer nodes. Null move pruning is one of these ideas that works well in practice.

LMR is very popular now. But I guess if you add null move and LMR and search the same depth, you will often get different answers.

I think the bigger problem is not necessarily finding the same answer in less nodes for tactical problems but rather for quiet problems.

Consider that after 30 captures we are down to bare kings. So many, if not most good chess moves do not involve a capture.

How do we pare down the choices for these quiet moves? This is the more difficult and more important question.

IMO-YMMV

Will we be seeing a new "just amyan" message soon?

Are we coming back to poetry after that release?

Miguel
PS: Welcome back Antonio.

[adieguez]

Hi, well I will try what I said and see if it is any good. Frustratingly I started playing games, 112 games each amyan version and results seems at this stage contradictory, not logical, and it's a little frustrating. A just amyan message will be only when I have something better, dunno when . Miguel and Gabor thanks for your welcomes backs.