Search depht and nodes/s

[Sidiki]

Hi all programmers,
Michel, Gilbert, Taille without forget the others greater programmers, before any word, HAPPY NEW YEAR to all.
I will be happy to have the light on the tittle of my topic.
If an engine A to a same position reach 22th ply with 4M/s and another engine B reach the same depth with 8M/s. Which engine of the 2 it's more stronger.
Just to say that i don't know if it's the depth of plies reached or the nodes/s, so the speed used to reach this depht that determine the strenght of an engine.
Thank.
Sidiki

[TAILLE]

Hi all programmers,
Michel, Gilbert, Taille without forget the others greater programmers, before any word, HAPPY NEW YEAR to all.
I will be happy to have the light on the tittle of my topic.
If an engine A to a same position reach 22th ply with 4M/s and another engine B reach the same depth with 8M/s. Which engine of the 2 it's more stronger.
Just to say that i don't know if it's the depth of plies reached or the nodes/s, so the speed used to reach this depht that determine the strenght of an engine.
Thank.
Sidiki

Happy new year Sidiki and all programmers.
Of course speed allows to find the best move in less time but speed is just one minor point when judging the strength of a program.
In addition the number of position per second is the worse criteria to take into account for a simple reason: let's compare a GMI and a strong program : the GMI and the program may have the same strength but while the program sees 4M positions per second how many positions per second can see the GMI ? Obvioulsy the program compensates a poor evaluation (comparing to the GMI's one) by a huge search. For that reason I think a better evaluation, reached by reinforcement learning, will greatly diminish the number of positions per second but will lead to far better results.
The key point for a strong program is for me a good evaluation allowing to prune agressively early in the tree search. Isn't it the differrnece between a GMI and a weaker player ?
BTW I am not able to say what the number of plies means. If your program use a very agressive pruning that means that some variants may be analysed on let's say 40 plies will other will stop after 6 plies. What can you conclude ? If the evaluation is very good this agressive pruning could lead to a very strong program otherwise this program may be sometimes very strong but also very weak.

Taking into account the number of positions per second and/or the number of plies is a non sense I you do not know teh eval function and the search algorithm.

[Sidiki]

Hi all programmers,
Michel, Gilbert, Taille without forget the others greater programmers, before any word, HAPPY NEW YEAR to all.
I will be happy to have the light on the tittle of my topic.
If an engine A to a same position reach 22th ply with 4M/s and another engine B reach the same depth with 8M/s. Which engine of the 2 it's more stronger.
Just to say that i don't know if it's the depth of plies reached or the nodes/s, so the speed used to reach this depht that determine the strenght of an engine.
Thank.
Sidiki

Happy new year Sidiki and all programmers.
Of course speed allows to find the best move in less time but speed is just one minor point when judging the strength of a program.
In addition the number of position per second is the worse criteria to take into account for a simple reason: let's compare a GMI and a strong program : the GMI and the program may have the same strength but while the program sees 4M positions per second how many positions per second can see the GMI ? Obvioulsy the program compensates a poor evaluation (comparing to the GMI's one) by a huge search. For that reason I think a better evaluation, reached by reinforcement learning, will greatly diminish the number of positions per second but will lead to far better results.
The key point for a strong program is for me a good evaluation allowing to prune agressively early in the tree search. Isn't it the differrnece between a GMI and a weaker player ?
BTW I am not able to say what the number of plies means. If your program use a very agressive pruning that means that some variants may be analysed on let's say 40 plies will other will stop after 6 plies. What can you conclude ? If the evaluation is very good this agressive pruning could lead to a very strong program otherwise this program may be sometimes very strong but also very weak.

Taking into account the number of positions per second and/or the number of plies is a non sense I you do not know teh eval function and the search algorithm.

Hi Taille,
Thank very much to take time to answer. I understood now. A good eval function and/or self-learning can increase the strenght. It's the same like Alphazero which calculated less positions than Stockfisch but defeated it.
Thank.
Sidiki.

[Krzysztof Grzelak]

Happy new year Sidiki and all programmers.
Of course speed allows to find the best move in less time but speed is just one minor point when judging the strength of a program.
In addition the number of position per second is the worse criteria to take into account for a simple reason: let's compare a GMI and a strong program : the GMI and the program may have the same strength but while the program sees 4M positions per second how many positions per second can see the GMI ? Obvioulsy the program compensates a poor evaluation (comparing to the GMI's one) by a huge search. For that reason I think a better evaluation, reached by reinforcement learning, will greatly diminish the number of positions per second but will lead to far better results.
The key point for a strong program is for me a good evaluation allowing to prune agressively early in the tree search. Isn't it the differrnece between a GMI and a weaker player ?
BTW I am not able to say what the number of plies means. If your program use a very agressive pruning that means that some variants may be analysed on let's say 40 plies will other will stop after 6 plies. What can you conclude ? If the evaluation is very good this agressive pruning could lead to a very strong program otherwise this program may be sometimes very strong but also very weak.

Taking into account the number of positions per second and/or the number of plies is a non sense I you do not know teh eval function and the search algorithm.

Hi Gerard.

Only that the "learning" is the note of the future.You have to seriously consider whether this option is worth taking care of and whether it will give something to the programmer in your program. I would like to see how the learning option works with the help of a draughts program.

[TAILLE]

Happy new year Sidiki and all programmers.
Of course speed allows to find the best move in less time but speed is just one minor point when judging the strength of a program.
In addition the number of position per second is the worse criteria to take into account for a simple reason: let's compare a GMI and a strong program : the GMI and the program may have the same strength but while the program sees 4M positions per second how many positions per second can see the GMI ? Obvioulsy the program compensates a poor evaluation (comparing to the GMI's one) by a huge search. For that reason I think a better evaluation, reached by reinforcement learning, will greatly diminish the number of positions per second but will lead to far better results.
The key point for a strong program is for me a good evaluation allowing to prune agressively early in the tree search. Isn't it the differrnece between a GMI and a weaker player ?
BTW I am not able to say what the number of plies means. If your program use a very agressive pruning that means that some variants may be analysed on let's say 40 plies will other will stop after 6 plies. What can you conclude ? If the evaluation is very good this agressive pruning could lead to a very strong program otherwise this program may be sometimes very strong but also very weak.

Taking into account the number of positions per second and/or the number of plies is a non sense I you do not know teh eval function and the search algorithm.

Hi Gerard.

Only that the "learning" is the note of the future.You have to seriously consider whether this option is worth taking care of and whether it will give something to the programmer in your program. I would like to see how the learning option works with the help of a draughts program.

Hi Krzysztof,
I am not sure of my understanding. Do you consider it is not a good idea to use "learning" concepts for improving draughts program ?

[Sidiki]

Happy new year Sidiki and all programmers.
Of course speed allows to find the best move in less time but speed is just one minor point when judging the strength of a program.
In addition the number of position per second is the worse criteria to take into account for a simple reason: let's compare a GMI and a strong program : the GMI and the program may have the same strength but while the program sees 4M positions per second how many positions per second can see the GMI ? Obvioulsy the program compensates a poor evaluation (comparing to the GMI's one) by a huge search. For that reason I think a better evaluation, reached by reinforcement learning, will greatly diminish the number of positions per second but will lead to far better results.
The key point for a strong program is for me a good evaluation allowing to prune agressively early in the tree search. Isn't it the differrnece between a GMI and a weaker player ?
BTW I am not able to say what the number of plies means. If your program use a very agressive pruning that means that some variants may be analysed on let's say 40 plies will other will stop after 6 plies. What can you conclude ? If the evaluation is very good this agressive pruning could lead to a very strong program otherwise this program may be sometimes very strong but also very weak.

Taking into account the number of positions per second and/or the number of plies is a non sense I you do not know teh eval function and the search algorithm.

Hi Gerard.

Only that the "learning" is the note of the future.You have to seriously consider whether this option is worth taking care of and whether it will give something to the programmer in your program. I would like to see how the learning option works with the help of a draughts program.

Hi Krzysztof,
I am not sure of my understanding. Do you consider it is not a good idea to use "learning" concepts for improving draughts program ?

Hi Gérard,
I think that he want to say that it seem that self-learning must be à new way to improve draughts programs, and he will be curious to know how this concept is applied. So if you can show him/us an example.
Sidiki.

[TAILLE]

Hi Gerard.

Only that the "learning" is the note of the future.You have to seriously consider whether this option is worth taking care of and whether it will give something to the programmer in your program. I would like to see how the learning option works with the help of a draughts program.

Hi Krzysztof,
I am not sure of my understanding. Do you consider it is not a good idea to use "learning" concepts for improving draughts program ?

Hi Gérard,
I think that he want to say that it seem that self-learning must be à new way to improve draughts programs, and he will be curious to know how this concept is applied. So if you can show him/us an example.
Sidiki.

Hi Krzysztof, Sidiki,

OK I see. I am also very curious to know how this concept can be applied for draughts! After having read some articles on the subject (see for example Alphago) I decided to try this approach with Damy but I am still at the very beginning of the process. For the time being I am studying and testing how convergence can be achieved for the pattern values. As you see I am today completly unable to give you some results. May be some other programmers might answer your question.

[Krzysztof Grzelak]

Hi Krzysztof,
I am not sure of my understanding. Do you consider it is not a good idea to use "learning" concepts for improving draughts program ?

Hi Gerard.
It's a good idea just how to put it into practice for the draughts. As for the Alphago program, see how many he used the processors and how long the code had to be. Of course, as to the code, I can be wrong because I'm not a programmer.

[Sidiki]

Hi Krzysztof,
I am not sure of my understanding. Do you consider it is not a good idea to use "learning" concepts for improving draughts program ?

Hi Gerard.
It's a good idea just how to put it into practice for the draughts. As for the Alphago program, see how many he used the processors and how long the code had to be. Of course, as to the code, I can be wrong because I'm not a programmer.

Hi Krszyztof, Gérard
Krzysztof, actually, like Gérard wrote, himself he's on experimentation stage, so he's for now, we will wait for the first good results.
Gérard, ok we will wait. It seem effectivelly that this approach combined to patterns can give something very awesome.
Thank again for your availibility.

Sidiki