personal experience regarding "self learning"

[deeds]

original post

Btw, to be honest, i have a completely different opinion and personal experience regarding "self learning", if you want to discuss it just open new thread and we will talk about it.
Regards.

Do you understand there is a difference between bestmoves and most efficient moves ?

Did you train an engine on one or more openings ? (I mean at least 500 games/opening)

Do you know why an engine does not always play the moves stored in its experience file ?

What is the "bonus" effect while learning ?

Because otherwise it will be a masquerade this exchange of personal experience...

[deeds]

Why we need min. 500 games in order to learn an opening ?

What is the main difference between an opening book and an experience file ?

What openings to learn ?

How many of the most effective moves are obtained with 2000 games ?

[deeds]

very interesting friend. especially for young people who are hungry to learn. CONGRATULATIONS.

[deeds]

before answering these questions, I await the first impressions of skynet...

[deeds]

before answering these questions, I await the first impressions of skynet...

You are not the only one Chris, who is waiting.for him. But I have made in the meantime two Android tournaments and wait for him till I'll post the results. But I'm also interested in the upcoming discussion here although, or better, because I don't know much about learning files and features.. but that's your thread DeeDs

[deeds]

ok let"s go ! (sorry skynet, hope you're fine and your family too)

Do you understand there is a difference between bestmoves and most efficient moves ?

bestmove : For most of us, this is the engine's favorite move. So it's intimately linked to his evaluation function (nowadays it's often his neural network) and the time we give him. This results in a score that corrects itself as its depth increases. The problem is that at the beginning or even in the middle of the game, this depth is not sufficient to reach the end of the game. For example, after 1. g4 d5, when an engine evaluates 2. e3 at -1.76/40, it evaluates that this move would be the best for the next 40 positions but the games starting with 1.g4 d5 contain on average 160 positions (value based on 85k games at various tc/hardware)

most efficient move : those are moves that statistically lead to more wins / fewer losses because they take into account the entire game, and are not just a succession of moves preferred by the engine for the next xx positions. As engines get stronger, their bestmoves will become more and more identical to the most effective moves. In the meantime, thanks to the learning, the engines can discover moves that are more effective than their bestmoves. If we take the previous example, after 1.g4 d5, 2.c4 or 2.Bg2 are the most effective :

[deeds]

Did you train an engine on one or more openings ? (I mean at least 500 games/opening)

At this moment i'm finishing the 9th learning session of Eman ([External Link Removed for Guests]). Each session contained about 30 openings. For each opening, Eman played 2000 games against 1 to 4 different opponents. Since 8th learning session, i set a TC 2m+2s (instead of a TC 1m+1s like my previous learning sessions). Most of the time these opponents also had a learning feature so no engine got carried away nor lost several times in the same way. I must admit that learning uses 50-80% of my cluster 24/7 (about 7-8 threads per engine).

[deeds]

1 Do you understand there is a difference between bestmoves and most efficient moves ?
--------------------------------------

2 Do you know why an engine does not always play the moves stored in its experience file ?

3 What is the "bonus" effect while learning ?

4 Did you train an engine on one or more openings ? (I mean at least 500 games/opening)

Because otherwise it will be a masquerade this exchange of personal experience...

Hello Cris, we are fine, hope you too.
1) Well, aren't the more efficient moves/considered the best ones? Otherwise, what is the point of a move if it is not effective? For example, there are more than trillions of positions in which there is only one single move, any other will lead to disaster, what do you think this move is - the best or the most efficient, or maybe the only one? If you have a different opinion about this feel free to share it.
2) I think there are a couple of reasons why the engine does this, for example, a move that is in the database, (that is in the experience file), which was saved as bad (not effective) so the engine starts looking for an alternative move. Or the reason could be a search depth tree, for example, the engine used a had more time to think and found a more suitable move on bigger depth, etc., etc.
3) Training bonus? I personally didn't find any, except that i received an bigger electricity bill that supposed, and this is definitely not a bonus.. Anyway, some time ago i was testing SF-MZ-230522 (also ShashChess) and since i didn't saw any difference in performance i stopped to spend my time with learning..
4) This is the main question.. I think that many will agree with me, because I don't see the point in wasting a year's supply of Africa's electricity just to make the engine to learn something. Seriously, in chess only after the first 4 moves there are 170,000 possible combinations, if we take into account that the engine has to play each position with white and black, we get 170,000x2 = 340,000 possible combinations, and now if we multiply this amount by the number of games that the engine has to play (about 500 each?) for the engine to learn something. And also taking into account the fact that with different time control (as well as the number of cores) the engine will reach different depths, in which there will be at least 2 possible moves, and also taking into account the updating of patches where the position estimate will be slightly changed - then training of the engine will never end.

Also i wanted to ask you about your tests Eman vs SF 15, do you believe that your test(s) was fair? I mean Eman was using experience file while SF was using nothing. Personally i see all that learning process useless, Lc0 is one of examples, zillion of played games and it still much behind the SF.
This is my humble opinion, if you want don't count it at all.

[deeds]

ok let"s go ! (sorry skynet, hope you're fine and your family too)

Do you understand there is a difference between bestmoves and most efficient moves ?

bestmove : For most of us, this is the engine's favorite move. So it's intimately linked to his evaluation function (nowadays it's often his neural network) and the time we give him. This results in a score that corrects itself as its depth increases. The problem is that at the beginning or even in the middle of the game, this depth is not sufficient to reach the end of the game. For example, after 1. g4 d5, when an engine evaluates 2. e3 at -1.76/40, it evaluates that this move would be the best for the next 40 positions but the games starting with 1.g4 d5 contain on average 160 positions (value based on 85k games at various tc/hardware)

most efficient move : those are moves that statistically lead to more wins / fewer losses because they take into account the entire game, and are not just a succession of moves preferred by the engine for the next xx positions. As engines get stronger, their bestmoves will become more and more identical to the most effective moves. In the meantime, thanks to the learning, the engines can discover moves that are more effective than their bestmoves. If we take the previous example, after 1.g4 d5, 2.c4 or 2.Bg2 are the most effective :

Lol, we definitely have a different understanding for sure. Firstly, engine doesn't have favorite move, it can play whatever, 1.c4 1.d4 1.e4 1.Nf3 etc.
Otherwise, we would have the same outcome, eg all the games would ended up with eg Ruy Lopez. As for effective moves that are based on statistics, you are right, but on the other hand, people (and even engines) do not think so. For example, if statistically, the move 1.e4 is the most effective, then those who constantly play the moves 1.d4 or 1.c4 will definitely not agree with the statistics.

[deeds]

At this moment i'm finishing the 9th learning session of Eman ([External Link Removed for Guests]). Each session contained about 30 openings. For each opening, Eman played 2000 games against 1 to 4 different opponents. Since 8th learning session, i set a TC 2m+2s (instead of a TC 1m+1s like my previous learning sessions). Most of the time these opponents also had a learning feature so no engine got carried away nor lost several times in the same way. I must admit that learning uses 50-80% of my cluster 24/7 (about 7-8 threads per engine).

Just one simple question, how much time you spent so far for the learning (i'll not ask about the burned kWh)?