Do we need learning file for engines?

[JoAnnP38]

I would like all GUIs to support the UCI_Opponent option so that engines could make minor adjustments based on their opponent. If it could learn what adjustments to make that would be even better!

[Enderjed]

I would like all GUIs to support the UCI_Opponent option so that engines could make minor adjustments based on their opponent. If it could learn what adjustments to make that would be even better!

I'm pretty sure at the very least, BanksiaGUI supports UCI_Opponent.

[Uri Blass]

I vote with yes.

I think that it may be interesting to make tournament when engines learn with learning files to find out if some engines has a better learning algorithm.

Suppose you give engines A and B to play 1000 positions like the superfinal TCEC positions at 1+1 time control and you get something like 1200-800 for A.

Suppose you repeat the match again and again when the engines use learning files.
In case B has the better learning algorithm then we may see something like this:

A 1200 B 800 in match 1
A 1100 B 900 in match 2
A 1000 B 1000 in match 3
A 900 B 1100 in match 4.

If we do not test in this way we will never know that B has the better learning algorithm and people will not know about B's advantage relative to A.

[Eduard]

Stefan Pohl tested Eman like this a long time ago. There was no progress with learning.

[Uri Blass]

Stefan Pohl tested Eman like this a long time ago. There was no progress with learning.

Unfortunately I know no tournaments to encourage progress with learning except ssdf rating list.

Unfortuantely people who make rating lists for engine based on position that is not the opening position do not play the same position many times in matches between engines.

There are 2 type of possible learning.
1)repeat opening lines that you won that I know engines used in the ssdf games.
2)Try to beat the opponent in the same way the opponent beat you in the previous game with opposite colors.
If the opponent does not lose because of choosing a different move then try this different move in the next game so the opponent cannot beat you twice in the same way.

I do not know if there are engines in the ssdf rating list that tried 2.

[Graham Banks]

Using engines that learn for rating lists would make direct comparisons less meaningful, as everybody's copy of an engine would have different learning at any given point.

[marsell]

Personally I don't see any real benefit in learning engines yet. For learning to be effective you would have to play thousands of games for each opening. A good opening book is easier to realise.

[abgursu]

Personally I don't see any real benefit in learning engines yet. For learning to be effective you would have to play thousands of games for each opening. A good opening book is easier to realise.

A good opening book and experience is better, since most books are repeating each other nowadays and usually end up in similar positions. Using a learning file after book is good to improve variations. But everyone having different learnings is the downside of it.

[Uri Blass]

Using engines that learn for rating lists would make direct comparisons less meaningful, as everybody's copy of an engine would have different learning at any given point.

I am not against rating lists without learning but I would like also to have rating list with learning.

For comparison between engines you can decide that you have fixed number of games between every 2 engines that play in the list(let say 50 games) and after every match of 50 games between 2 engines you delete the learning files so practically engines that play always learned from the same games between them(I do not know one rating list that does it).

[sarona]

I vote with yes.
....
Suppose you repeat the match again and again when the engines use learning files.
In case B has the better learning algorithm then we may see something like this:

A 1200 B 800 in match 1
A 1100 B 900 in match 2
A 1000 B 1000 in match 3
A 900 B 1100 in match 4.

If we do not test in this way we will never know that B has the better learning algorithm and people will not know about B's advantage relative to A.

I, too, voted yes and agree with what you have written above.