Is there a consensus on Mentor Learning?

[thomasahle]

Omid David-Tabibi, Moshe Koppel and Nathan S. Netanyahu wrote an interesting paper[1] on tuning an evaluation function by using another engine as a 'mentor'. I'm sorry if this has already been discussed enough in this forum, I'm still new here.

I think the method is interesting because it lets us tune evaluation functions with less of the noise coming from win/loss rates. The way the paper uses the idea is a bit limited, since it requires another engine to mimic, but what about if we let the other engine search some 10-15 plies and then use that score as the evaluation value to mimic. This way an engine can even mentor itself.

Anyway, I'm trying my luck with this method at the moment. Only for piece square tables for now. I'll report the results be they good or bad.

[1] http://u.cs.biu.ac.il/~koppel/papers/ch ... -final.pdf

[Greg Strong]

I tune against a number of other engines. I haven't read the paper quoted, but I would think if you were tuning eval parameters to maximize the outcome against one specific engine, you would be weakening your engine in general by essentially tuning for the specific purpose of countering the tactics that play out against that engine. I have to believe you would get better overall results competing against a number of engines that are somewhat stronger than yours, but not vastly superior...

[thomasahle]

Sorry, maybe I was not clear enough. The idea is to not tune "against" another engine, but to use another engine to create a list of (position, evaluation) tuples, and then use machine learning to approximate this function by simpler means.

[Gerd Isenberg]

Omid David-Tabibi, Moshe Koppel and Nathan S. Netanyahu wrote an interesting paper[1] on tuning an evaluation function by using another engine as a 'mentor'. I'm sorry if this has already been discussed enough in this forum, I'm still new here.

I think the method is interesting because it lets us tune evaluation functions with less of the noise coming from win/loss rates. The way the paper uses the idea is a bit limited, since it requires another engine to mimic, but what about if we let the other engine search some 10-15 plies and then use that score as the evaluation value to mimic. This way an engine can even mentor itself.

Anyway, I'm trying my luck with this method at the moment. Only for piece square tables for now. I'll report the results be they good or bad.

[1] http://u.cs.biu.ac.il/~koppel/papers/ch ... -final.pdf

Hi Thomas,

welcome to CCC and congrats for your Sunfish - great to learn some Python. I am not an expert in eval tuning and GA, but if the mentor has eval knowledge you don't have, i.e. pawn structure, passers, backward, isolated, doubled pawns, some king safety, etc., I wonder how meaningful the learned results are with psq-tables only. Also, I guess, it makes sense to implement quiescence first. Otherwise, you may find Peter Österlund's recent description on how he tuned Texel interesting:
http://www.talkchess.com/forum/viewtopi ... 22&t=50823

Gerd

[tpetzke]

Hi,

if your evaluation consists only of psq tables you can reach almost the same result by just thinking of reasonable numbers and apply some symmetries. PSQ cover some well known very general rules only.

An engine with PSQ plays certainly better than an engine that only has material eval but is much weaker than an engine that is able to evaluate the actual position more specifically.

Good luck anyway
Thomas...

[thomasahle]

I don't agree that psq's are easy to just invent by hand. If they were, there wouldn't be factor 10 differences between the tables between engines.

None the less mentor training has nothing to do with the specific metric. It is merely a much more efficient metric to use when tuning your evaluation function.