expRate

Code, algorithms, languages, construction...
User avatar
deeds
Posts: 754
Joined: Wed Oct 20, 2021 9:24 pm
Location: France
Contact:

Re: expRate

Post by deeds » Sun Aug 27, 2023 2:07 pm

Martin_1969 wrote:
Sun Aug 27, 2023 1:24 pm
OK thanks! I will update when the engine has learned 1000 moves in my favorite opening.
It was just misleading for me when in the thread "training mode (Eman) with trouney mode (Raptora)" you already had a learning rate of 6% after a few rounds.
The Eman's experience file already learned 2 openings of the Sedat's Unique v50 opening list (expRate = 4%). After 2-3 rounds, it learned another opening (expRate = 6%). It's not hard to understand if Eman did more rounds, it would learn more openings. But as i don't want to help cheaters, i limited the rounds in order to reduce the amount of published Eman's moves.

Martin_1969
Posts: 16
Joined: Mon Jul 03, 2023 6:53 am

Re: expRate

Post by Martin_1969 » Mon Aug 28, 2023 7:17 am

I see you didn't quite understand me. In your experiment "training mode (Eman) with trouney mode (Raptora)" you played a total of 300 games, covering only 1/5 of the Unique v50 openings. BUT your engine achieved a learning rate of 6%.
I think the learning rate is exaggerated. Because after 500 games and only one opening line of my choice I reach a learning rate of 0%. You can easily check this. Play 300 games and only one opening line with a fresh learning file and you will see that after 300 games played by the engine the result is 0% learning.
I like a lot of what you write, but not everything.

Everyone else can check this for themselves

All I'm saying is that you need a lot more games than just a few hundred or thousand, and that's quite painful because there are so many possible moves that only a very small part can be covered by this system.
Bye Martin

User avatar
deeds
Posts: 754
Joined: Wed Oct 20, 2021 9:24 pm
Location: France
Contact:

Re: expRate

Post by deeds » Mon Aug 28, 2023 7:58 am

Martin_1969 wrote:
Mon Aug 28, 2023 7:17 am
...
I think the learning rate is exaggerated. Because after 500 games and only one opening line of my choice I reach a learning rate of 0%.
...
Play 300 games and only one opening line with a fresh learning file and you will see that after 300 games played by the engine the result is 0% learning.
...
I think to draw conclusions from just one opening it is exaggerated, especially when the threshold of 500 games/opening was checked by the authors of the learning engines and the learning features. Maybe even in the face of the evidence, you wouldn't accept the result if it said the opposite of your thinking.

Martin_1969 wrote:
Mon Aug 28, 2023 7:17 am
All I'm saying is that you need a lot more games than just a few hundred or thousand, and that's quite painful because there are so many possible moves that only a very small part can be covered by this system.
Bye Martin
This is the weak point of your reasoning. Engine learning is not a brute-force calculation. Every possible move doesn't have to be tried until the engine loses all games.

Martin_1969
Posts: 16
Joined: Mon Jul 03, 2023 6:53 am

Re: expRate

Post by Martin_1969 » Tue Aug 29, 2023 6:08 am

you draw conclusions from 300 games in your experiment and give a value of 6%? Which is simply wrong.

If you answer again you should respond to my primary question and explain to me how you come to a value of 6% after 300 games with an opening set of 50 lines in the thread "training mode (Eman) with trouney mode (Raptora)", which is impossible.

But now really bye and have a nice day

User avatar
deeds
Posts: 754
Joined: Wed Oct 20, 2021 9:24 pm
Location: France
Contact:

Re: expRate

Post by deeds » Tue Aug 29, 2023 6:39 am

Image

Oops !

Post Reply