Chessterfield learning utility available

[Matthias Lüscher]

Geschrieben von: / Posted by: Matthias Lüscher at 28 May 2000 23:36:04:
On my homepage you can now download the utility (ChessterfieldEN) needed to
perform the learning process for ChessterfieldCL. With this utility you can
train Chessterfield with your own sets of chess games.
Unfortunately, ChessterfieldEN needs lots of memory (>=128MB) and the learning
process can take hours.
Matthias Lüscher

Chessterfield Homepage

[Volker Pittlik]

Geschrieben von: / Posted by: Volker Pittlik at 01 June 2000 11:27:45:
Als Antwort auf: / As an answer to: Chessterfield learning utility available geschrieben von: / posted by: Matthias Lüscher at 28 May 2000 23:36:04:
[...]>Unfortunately, ChessterfieldEN needs lots of memory (>=128MB) and the learning process can take hours.
That's unfortunately to much for my 128MB machine. I feel I have to buy one more of this little blocks .
Volker

[Matthias Lüscher]

Geschrieben von: / Posted by: Matthias Lüscher at 01 June 2000 14:51:29:
Als Antwort auf: / As an answer to: Re: Chessterfield learning utility available geschrieben von: / posted by: Volker Pittlik at 01 June 2000 11:27:45:

[...]>Unfortunately, ChessterfieldEN needs lots of memory (>=128MB) and the learning process can take hours.
That's unfortunately to much for my 128MB machine. I feel I have to buy one more of this little blocks .
Volker

Depending on how big your training set is, 128MB should be enough. I have trained some 860 Games on my 128MB machine and I think it would also be possible to fit 2000 games with 128MB.
The problem is, that - with Windows98 - it is difficult to guess how much memory is needed.
Matthias

[Volker Pittlik]

Geschrieben von: / Posted by: Volker Pittlik at 01 June 2000 17:46:59:
Als Antwort auf: / As an answer to: Re: Chessterfield learning utility available geschrieben von: / posted by: Matthias Lüscher at 01 June 2000 14:51:29:

Depending on how big your training set is, 128MB should be enough. I have trained some 860 Games on my 128MB machine and I think it would also be possible to fit 2000 games with 128MB.
The problem is, that - with Windows98 - it is difficult to guess how much memory is needed.

Yes, you are right. In my first test I used a training set containing ~18,000 games. Just running a set containing 324 games. I'm wondering my resulting network.dat is bigger than yours: 702 KB. I going to try to make a network.dat containing 2000 games.
Don't know. I'm running NT. For a training set containing 324 games ~11 MB were used.
Volker

[Matthias Lüscher]

Geschrieben von: / Posted by: Matthias Lüscher at 02 June 2000 14:59:47:
Als Antwort auf: / As an answer to: Re: Chessterfield learning utility available geschrieben von: / posted by: Volker Pittlik at 01 June 2000 17:46:59:

Depending on how big your training set is, 128MB should be enough. I have trained some 860 Games on my 128MB machine and I think it would also be possible to fit 2000 games with 128MB.

Yes, you are right. In my first test I used a training set containing ~18,000 games. Just running a set containing 324 games. I'm wondering my resulting network.dat is bigger than yours: 702 KB.

18´000 Games is quite a big number. I guess it would take weeks to train Chessterfield with 18´000 games.
I going to try to make a network.dat containing 2000 games.
That sounds interesting. I have just used 860 games that were automatically played with ChessterfieldCL against other engines. My idea was to test, whether Chessterfield can learn from its own mistakes. If you only take grandmaster games, Chessterfield will never learn, that it is uncool to place the queen in the middle of the board just in the beginning of the game (grandmasters will probably never do this and therefore Chessterfield has to do it several times to learn the result of such a move). An interesting thing would be to make a mixture between Chessterfield against XY and grandmaster games and to investigate, wether Chessterfield can profit from the good positional knowledge of grandmasters.
Matthias

[Dann Corbit]

Geschrieben von: / Posted by: Dann Corbit at 02 June 2000 20:26:56:
Als Antwort auf: / As an answer to: Re: Chessterfield learning utility available geschrieben von: / posted by: Matthias Lüscher at 02 June 2000 14:59:47:

Depending on how big your training set is, 128MB should be enough. I have trained some 860 Games on my 128MB machine and I think it would also be possible to fit 2000 games with 128MB.

Yes, you are right. In my first test I used a training set containing ~18,000 games. Just running a set containing 324 games. I'm wondering my resulting network.dat is bigger than yours: 702 KB.

18´000 Games is quite a big number. I guess it would take weeks to train Chessterfield with 18´000 games.
I going to try to make a network.dat containing 2000 games.
That sounds interesting. I have just used 860 games that were automatically played with ChessterfieldCL against other engines. My idea was to test, whether Chessterfield can learn from its own mistakes. If you only take grandmaster games, Chessterfield will never learn, that it is uncool to place the queen in the middle of the board just in the beginning of the game (grandmasters will probably never do this and therefore Chessterfield has to do it several times to learn the result of such a move). An interesting thing would be to make a mixture between Chessterfield against XY and grandmaster games and to investigate, wether Chessterfield can profit from the good positional knowledge of grandmasters.

Can you create an iterative version?
We train with 10,000 games, then 10,000 more games, etc.
I have literally millions of games I could run though the training motor.
If you could create (perhaps) separate files and then average them or even have a second net that learns from the learn file, it might create an excellent knowledge repository.


My FTP site

[Matthias Lüscher]

Geschrieben von: / Posted by: Matthias Lüscher at 02 June 2000 21:09:05:
Als Antwort auf: / As an answer to: Re: Chessterfield learning utility available geschrieben von: / posted by: Dann Corbit at 02 June 2000 20:26:56:

Can you create an iterative version?
We train with 10,000 games, then 10,000 more games, etc.
I have literally millions of games I could run though the training motor.
If you could create (perhaps) separate files and then average them or even have a second net that learns from the learn file, it might create an excellent knowledge repository.

That would be possible. However, I think that lots of experiments are necessary to find a good iterative algorithm which outperforms the current algorithm. Another problem with machine learning is, that one never knows the best size of a training set (in general the following rule is NOT true: bigger training set = better learning).
Another problem is, that at the moment, I only have very few time to improve Chessterfield. The task, which I am working on is, that I want to make the source code more readable so that I can distribute it under the GNU license (Unfortunately this could still take some months).
If anyone doesn't want to wait, she/he can send me an e-mail and I will send you the sourcecode. Any learning strategy improvements are welcome.