World Draughts Forum

It is currently Sun Dec 17, 2017 11:13

All times are UTC+01:00




Post new topic  Reply to topic  [ 186 posts ]  Go to page Previous 19 10 11 12 13
Author Message
PostPosted: Mon Oct 09, 2017 14:07 
Offline

Joined: Wed Apr 14, 2004 16:04
Posts: 1559
TAILLE wrote:
Rein Halbersma wrote:

Hi Gerard,

Is it possible for you to modify your routine so as to print the capture positions after the last generation pass? And then drop them for the final compression stage? Or would this be a too invasive change of your generator?

Rein


I don't quite understand your question.

Basically my routine looks like:

for (db = first db; db < last db; db = following db)
{
for (p = first position of the db; p < last position; p = following position)
{
if (p == capture position) continue;
else {generate successors; evaluate p; save result in memory db}
}
compress memory db and store on disk;
}

Of course I added some complexity in order to have a multithread routine


Ah OK, thanks I understand. From what I learnt from Ed's description, he generates the full results for all positions, including captures, and then after generation he sets the capture positions to the most common value (e.g. a draw) and then he compresses the databases. With that method, you can still generate capture statistics just before compression.


Top
   
PostPosted: Mon Oct 09, 2017 21:34 
Offline

Joined: Sun Dec 28, 2003 20:24
Posts: 235
TAILLE wrote:
Hi Michel,
Good news to see your interesting by this beautitul draughts variant.
As with Bert I have the same difficulty with you : I did not generate the capture positions. Bert managed to add statitistics on non capture position in order to compare our figures but I do not know if it is easy for your implementation to do the same.
FYI my figures are
1x1 NCW1= 1 028 NCL1= 837 NCW2= - NCL2= -
2x1 NCW1= 30 993 NCL1= 6 865 NCW2= 11 268 NCL2= 26 595
2x2 NCW1= 424 883 NCL1= 285 789 NCW2= - NCL2= -
3x1 NCW1= 446 168 NCL1= 55 534 NCW2= 113 391 NCL2= 388 438
3x2 NCW1= 6 778 061 NCL1= 1 948 633 NCW2= 3 775 435 NCL2= 4 952 702
3x3 NCW1= 63 546 643 NCL1= 35 985 469 NCW2= - NCL2= -
4x1 NCW1= 4 518 137 NCL1= 363 194 NCW2= 927 678 NCL2= 3 954 768
4x2 NCW1= 67 423 737 NCL1= 11 426 249 NCW2= 28 548 951 NCL2= 50 312 655
4x3 NCW1= 652 076 102 NCL1= 184 976 683 NCW2= 438 218 089 NCL2= 398 830 286
4x4 NCW1= 4 486 519 068 NCL1= 2 080 557 566 NCW2= - NCL2= -


I did a non-capture count of my databases, and the numbers match :-)

How big a database did you build?

Michel


Top
   
PostPosted: Mon Oct 09, 2017 22:25 
Offline

Joined: Thu Apr 26, 2007 18:51
Posts: 807
Location: FRANCE
MichelG wrote:
TAILLE wrote:
Hi Michel,
Good news to see your interesting by this beautitul draughts variant.
As with Bert I have the same difficulty with you : I did not generate the capture positions. Bert managed to add statitistics on non capture position in order to compare our figures but I do not know if it is easy for your implementation to do the same.
FYI my figures are
1x1 NCW1= 1 028 NCL1= 837 NCW2= - NCL2= -
2x1 NCW1= 30 993 NCL1= 6 865 NCW2= 11 268 NCL2= 26 595
2x2 NCW1= 424 883 NCL1= 285 789 NCW2= - NCL2= -
3x1 NCW1= 446 168 NCL1= 55 534 NCW2= 113 391 NCL2= 388 438
3x2 NCW1= 6 778 061 NCL1= 1 948 633 NCW2= 3 775 435 NCL2= 4 952 702
3x3 NCW1= 63 546 643 NCL1= 35 985 469 NCW2= - NCL2= -
4x1 NCW1= 4 518 137 NCL1= 363 194 NCW2= 927 678 NCL2= 3 954 768
4x2 NCW1= 67 423 737 NCL1= 11 426 249 NCW2= 28 548 951 NCL2= 50 312 655
4x3 NCW1= 652 076 102 NCL1= 184 976 683 NCW2= 438 218 089 NCL2= 398 830 286
4x4 NCW1= 4 486 519 068 NCL1= 2 080 557 566 NCW2= - NCL2= -


I did a non-capture count of my databases, and the numbers match :-)

How big a database did you build?

Michel


Hi Michel,

I generated also 5x1, 5x2, 5x3, 5x4, 5x5, 6x1, 6x2, 6x3 and 6x4.
My goal was to reach 6x6 but during the 6x5 generation I noted my db reached 300Gb and the 6x5 db was just half generated. I stopped it because I conclude I will not be able to generate the 6x6 db due to the space needed on the disk. In addition the time needed becomes also quite large.

Now I make a pause on the generation process and I work on the eval function.

Later I expect to find a new compression mechanism in order to solve disk space issue.

_________________
Gérard


Top
   
PostPosted: Fri Dec 15, 2017 08:47 
Offline

Joined: Sun Dec 28, 2003 20:24
Posts: 235
A small update from my part.

I have added the breakthrough rules to dragon.

First i played with the original evaluation function (from international draughts) against scan 3.0. This scores about 10%.

Currently i am learning a new set of evaluation patterns for breakthrough. The initial results are encouraging; after about 80 hours of learning, dragon wins about 33%. (using 15-field patterns)

Before increasing the learning time, there is some more research to do. For example selecting the right positions to feed the learner, and finding the best hyperparameters.

Dragon now uses 7 piece endgame databases, but it seems that these increase strength only by a relatively small amount.

Michel


Top
   
PostPosted: Fri Dec 15, 2017 09:50 
Offline

Joined: Thu Jun 20, 2013 17:16
Posts: 561
Real name: Krzysztof Grzelak
I'm sorry to ask you Michel, I understand that you have added updates on your homepage program.


Top
   
PostPosted: Fri Dec 15, 2017 19:32 
Offline

Joined: Thu Apr 26, 2007 18:51
Posts: 807
Location: FRANCE
MichelG wrote:
A small update from my part.

I have added the breakthrough rules to dragon.

First i played with the original evaluation function (from international draughts) against scan 3.0. This scores about 10%.

Currently i am learning a new set of evaluation patterns for breakthrough. The initial results are encouraging; after about 80 hours of learning, dragon wins about 33%. (using 15-field patterns)

Before increasing the learning time, there is some more research to do. For example selecting the right positions to feed the learner, and finding the best hyperparameters.

Dragon now uses 7 piece endgame databases, but it seems that these increase strength only by a relatively small amount.

Michel


Hi Michel,

I continue to work on breakthrough draugths. I am trying to understand reinforcement learning, I confess it is quite hard for me but I am really motivated though my progress are rather slow!

Could I ask you one or two questions Michel? I tried my first reinforcement learning on a very simple base : what is the value of a material advantage ? In my classical evaluation I decided that the value of one man was +100 and the value of a win was equal to the value of 20 men i.e. +2000. With my first experiments with reinforcement learning and keeping the value +2000 for a win the programm told me after a reinforcement learning that the value of an advantage of one man is a value near from +1000, the value of an advantage of 2 men is a value near form 1600 etc. and the value of en advantage of 10 men is very near from +2000.
Clearly the values I estimated represent the probability to win and obvioulsly it could not be a linear function. Do you also have a estimation function representing the probablity of winning (this sounds a great simplication due to the fact that the draw does not exist) and do you have also such non-linear results?

Gérard

_________________
Gérard


Top
   
Display posts from previous:  Sort by  
Post new topic  Reply to topic  [ 186 posts ]  Go to page Previous 19 10 11 12 13

All times are UTC+01:00


Who is online

Users browsing this forum: No registered users and 6 guests


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB® Forum Software © phpBB Limited