Sidiki, the main weakness in Damage (think others could add to this), is the situation in the early game phase with a king for many opposite pieces. Often this is not good for the king, but there are examples where in the end this create a huge advantage.
Another weakness is when both have a king with an imbalance in normal pieces, this is another situation which is hard to evaluate. As there might be not many learning examples, and most likely both sides will introduce many errors, as the right moves are beyond the search horizon.
Bert
Damage 15.3
Re: Damage 15.3
Nice Bert,
I see , with kings no easy to evaluate. So the importance to have a endgame can solve the problem ?
I understand that make a solid evaluation function for these cases need a huge job to be perfect. A deep knowledge in draughts or the game concerned is't required to fixe this point? Because even the grand Masters have problems with Kings into both sides.
We are glad to the work that you done you and the others. Thank
Sidiki
I see , with kings no easy to evaluate. So the importance to have a endgame can solve the problem ?
I understand that make a solid evaluation function for these cases need a huge job to be perfect. A deep knowledge in draughts or the game concerned is't required to fixe this point? Because even the grand Masters have problems with Kings into both sides.
We are glad to the work that you done you and the others. Thank
Sidiki
Re: Damage 15.3
We encourage you and we are waiting the tests until the release, to enjoy Damage.
Re: Damage 15.3
Hi Bert,BertTuyt wrote: ↑Mon Feb 17, 2020 15:02Sidiki, the main weakness in Damage (think others could add to this), is the situation in the early game phase with a king for many opposite pieces. Often this is not good for the king, but there are examples where in the end this create a huge advantage.
Another weakness is when both have a king with an imbalance in normal pieces, this is another situation which is hard to evaluate. As there might be not many learning examples, and most likely both sides will introduce many errors, as the right moves are beyond the search horizon.
Bert
There is 2 kind of programs, the first need to be runed on a powerful computer to reach his top level, another one is focused on the evaluation. Damage, the new is of what side ?
Thank.
Sidiki
Re: Damage 15.3
Sidiki, Damage (more or less) behaves in a similar way as Scan and Kingsrow.
Most likely also comparable with Maximus, but I never tested that program.
As with the similar structure of the evaluation the programs are now really fast, and in addition to aggressive pruning all reach significant depth.
You don't need powerful hardware for that these days.
Also keep in mind that moderns processors have multiple cores (even the base and/or entry models have 6 or more), and turbo frequencies exceeding 4 GHZ. Even on a laptop.
This yields (at least for Damage) speeds around (maximum) 70 MNodes/second. Kingsrow on my machine even runs at 80 MNodes/second.
With 32 core Threadrippers you enter then the Champions League, but if this speed is really needed is doubtful, as Kingsrow and Scan (most likely) already play close to perfection.
Bert
Most likely also comparable with Maximus, but I never tested that program.
As with the similar structure of the evaluation the programs are now really fast, and in addition to aggressive pruning all reach significant depth.
You don't need powerful hardware for that these days.
Also keep in mind that moderns processors have multiple cores (even the base and/or entry models have 6 or more), and turbo frequencies exceeding 4 GHZ. Even on a laptop.
This yields (at least for Damage) speeds around (maximum) 70 MNodes/second. Kingsrow on my machine even runs at 80 MNodes/second.
With 32 core Threadrippers you enter then the Champions League, but if this speed is really needed is doubtful, as Kingsrow and Scan (most likely) already play close to perfection.
Bert
Re: Damage 15.3
Thank Bert,
I know now that Damage reached a high level. 70Mo/s is enough to be like your already said, to the same league that kingsrow and Scan. This is for our Joy.
I suppose that the engine will run under Hub. So is't possible to do a 32 bits engine version too, we will run it on Android with exagear Windows emulator.
Thank again
Sidiki
I know now that Damage reached a high level. 70Mo/s is enough to be like your already said, to the same league that kingsrow and Scan. This is for our Joy.
I suppose that the engine will run under Hub. So is't possible to do a 32 bits engine version too, we will run it on Android with exagear Windows emulator.
Thank again
Sidiki
Re: Damage 15.3
Sidiki, the first Damage engine will still be based upon the GUIDE protocol, later i will add a HUB version.
Bert
Bert
Re: Damage 15.3
OK, Bert
Understood, this is already nice, the dxp mode will give us many challenges against Kingsrow, Scan, Dragon and Maximus.
Thank, we are very excited by discover it.
Sidiki
Re: Damage 15.3
Sidiki, I will issue this evening an overview of the tests I did, related to strength of the Damage program , and number of games for learning used.
Bert
Bert
Re: Damage 15.3
Very nice Bert,
As i saw it you was very busy these last days. So if the result it's for this evening, we are happy and hopefully to see the good result.
The fact it's that we learn many things from engines that help us again to perfect our game. Having Damage, Dragon, Kingsrow, Maximus and Scan as trainers by learning to these differents playing styles help us.
It's my case, it's why i'm always excited and enjoyed to see a new strong engine.
Thank to you, Ed, Fabien, Michel, Jaap and the others that make it possible. God bless you all.
Sidiki.
As i saw it you was very busy these last days. So if the result it's for this evening, we are happy and hopefully to see the good result.
The fact it's that we learn many things from engines that help us again to perfect our game. Having Damage, Dragon, Kingsrow, Maximus and Scan as trainers by learning to these differents playing styles help us.
It's my case, it's why i'm always excited and enjoyed to see a new strong engine.
Thank to you, Ed, Fabien, Michel, Jaap and the others that make it possible. God bless you all.
Sidiki.
Re: Damage 15.3
Bert,BertTuyt wrote: ↑Thu Feb 20, 2020 16:56Sidiki, Damage (more or less) behaves in a similar way as Scan and Kingsrow.
Most likely also comparable with Maximus, but I never tested that program.
As with the similar structure of the evaluation the programs are now really fast, and in addition to aggressive pruning all reach significant depth.
You don't need powerful hardware for that these days.
Also keep in mind that moderns processors have multiple cores (even the base and/or entry models have 6 or more), and turbo frequencies exceeding 4 GHZ. Even on a laptop.
This yields (at least for Damage) speeds around (maximum) 70 MNodes/second. Kingsrow on my machine even runs at 80 MNodes/second.
With 32 core Threadrippers you enter then the Champions League, but if this speed is really needed is doubtful, as Kingsrow and Scan (most likely) already play close to perfection.
Bert
It seem that Montecarlo's algorithm it's more used these days instead of alpha-bêta. Is't for speed or ? Because i think that the first one it's optimal.
Thank
Sidiki
Re: Damage 15.3
Sidiki, it is not for speed, it is a different approach.
Often used in combination with NN (Neural Network), where the NN is embedded in a GPU.
Good examples in Go (AlphaZero) and chess (Leela).
Bert
Often used in combination with NN (Neural Network), where the NN is embedded in a GPU.
Good examples in Go (AlphaZero) and chess (Leela).
Bert
Re: Damage 15.3
OK Bert , i understand now that it's linked to NN, thank again.
So, my Last and not least question for you
: It question of Damage, what's the advantage to use self learning or learning option ? Because of Alphazero famous games, it's writed that he learned chess in 4 hours of computing. In draughts how can we reach it. Maximus for example used in his latest version, it's strongh but don't enough to defeat Scan or Kingsrow.
You mentionned learning into Damage, what's for you the limit or the benefit or danger to use this latest technology?
Sorry if I ask many question, it's for understand very well some concepts.
Thank
Sidiki.
So, my Last and not least question for you
: It question of Damage, what's the advantage to use self learning or learning option ? Because of Alphazero famous games, it's writed that he learned chess in 4 hours of computing. In draughts how can we reach it. Maximus for example used in his latest version, it's strongh but don't enough to defeat Scan or Kingsrow. You mentionned learning into Damage, what's for you the limit or the benefit or danger to use this latest technology?
Sorry if I ask many question, it's for understand very well some concepts.
Thank
Sidiki.
Re: Damage 15.3
Sidiki, herewith some answers, and others might have a different opinion.
* The advantage of learning is that you can write now very fast, and very good evaluation functions, without the need to have draughts domain knowledge.
* The 4 hours are also related to the hardware Google has, will provide a more detailed answer later today.
* The strength of the Engine is a combination of search, evaluation, opening book, endgame DBs, time management, processing power.....
In the case of Maximus it could be that the evaluation is equal (or even better), but due to a less effective search and the somewhat slower Java, the overall strength is below Scan.
If we all would use a 32 core (4 GHZ) Threadripper and 20/min game, I guess most top programs would draw all the time.
* The limit and danger is the draw black hole. With machine learning one can write a strong draughts program without domain knowledge, and when you add the power of a multi-core CPU, than all programs draw in the end. All known search algorithms and enhancements are very well documented and available on the web (Scan, Stockfish). Ed provided his 8p DB to the community.
So people who start today, and who study and use all ideas and insights all on the internet, can have a jump start (see all the clones in chess).
You only might have differences when you limit the match to very short time controls, and/or limit the number of cores used. When regular time is used and all the CPU power is available you need to play zillions of games to see a win-lose.
For chess this is different as we still see ELO increase on a yearly basis.
I believe that with draughts we are close to perfection, and I dont believe that we see a program in the future to beat Kingsrow/Scan with regular time (2-5 min/game) with a score like 10 win -0 lose -148 draw, with the normal 158 DXP match. But I also thought a long time ago that Kingsrow was the end-point, end i was wrong too.
If this is the case (draw black hole) then people might lose their interest in Computer Draughts.
A way out is to change the rules, like going to killer or breakthrough Draughts.
Or go away from alpha-beta search , and only do tournaments with MCTS and NN, which is interesting!
Hope this, partly, answers your questions.
Bert
* The advantage of learning is that you can write now very fast, and very good evaluation functions, without the need to have draughts domain knowledge.
* The 4 hours are also related to the hardware Google has, will provide a more detailed answer later today.
* The strength of the Engine is a combination of search, evaluation, opening book, endgame DBs, time management, processing power.....
In the case of Maximus it could be that the evaluation is equal (or even better), but due to a less effective search and the somewhat slower Java, the overall strength is below Scan.
If we all would use a 32 core (4 GHZ) Threadripper and 20/min game, I guess most top programs would draw all the time.
* The limit and danger is the draw black hole. With machine learning one can write a strong draughts program without domain knowledge, and when you add the power of a multi-core CPU, than all programs draw in the end. All known search algorithms and enhancements are very well documented and available on the web (Scan, Stockfish). Ed provided his 8p DB to the community.
So people who start today, and who study and use all ideas and insights all on the internet, can have a jump start (see all the clones in chess).
You only might have differences when you limit the match to very short time controls, and/or limit the number of cores used. When regular time is used and all the CPU power is available you need to play zillions of games to see a win-lose.
For chess this is different as we still see ELO increase on a yearly basis.
I believe that with draughts we are close to perfection, and I dont believe that we see a program in the future to beat Kingsrow/Scan with regular time (2-5 min/game) with a score like 10 win -0 lose -148 draw, with the normal 158 DXP match. But I also thought a long time ago that Kingsrow was the end-point, end i was wrong too.
If this is the case (draw black hole) then people might lose their interest in Computer Draughts.
A way out is to change the rules, like going to killer or breakthrough Draughts.
Or go away from alpha-beta search , and only do tournaments with MCTS and NN, which is interesting!
Hope this, partly, answers your questions.
Bert
Re: Damage 15.3
Thank a lot Bert,
You answered more that i hoped.
It's clear that with the programs that exist in draughts, the strenght is closer of perfection. It's also true that draw are frequently saw into 158 dxp games at the top level.
*Like you, in the past, for me Truus was the end point, after Buggy, and Damy, and Kingsrow, and Scan, and now the "3" Scan, Damage, Kingsrow.
Just to say that, only if we don't find any new programming method, like you said, some 10-0-148 result will never appear again.
*Concerning killer rules, i think that we will apply it to draughts engines, because this perfection of draughts itsn't yet reached even by GMI.
So i think that we will prepare a tournament for all engines that have killer implementation to have a fun.
Thank again for taking time to answer to the question, i understood many things.
Sidiki.
You answered more that i hoped.
It's clear that with the programs that exist in draughts, the strenght is closer of perfection. It's also true that draw are frequently saw into 158 dxp games at the top level.
*Like you, in the past, for me Truus was the end point, after Buggy, and Damy, and Kingsrow, and Scan, and now the "3" Scan, Damage, Kingsrow.
Just to say that, only if we don't find any new programming method, like you said, some 10-0-148 result will never appear again.
*Concerning killer rules, i think that we will apply it to draughts engines, because this perfection of draughts itsn't yet reached even by GMI.
So i think that we will prepare a tournament for all engines that have killer implementation to have a fun.
Thank again for taking time to answer to the question, i understood many things.
Sidiki.
