the power of romi learning

[ozziejoe]

I've had romi playing a particular position in the caro cann against a range of engines, to evaluate its learning. You can see that it started out losing all the time against rybka and movei. It also struggled initially agains the other opponents (rated considerably higher)

Now Romi beating everybody regularly. Huge improvement against opponents ranked much higher. It is now playing even with rybka.

I will change the time control to force romi's learning file to be generally applicable
results below

best
J



1: Romichess 280.0/506 ·····················································································
2: Rybkav2.3.2a.w32 64.0/84 1=1111==111111111111=11111101110=1=1=11111111111111111111111=1010=0101010=010=0=0010
3: Movei00_8_438 58.0/84 111111=1=11=111111111111111111111011=11====11111==11101000000010=01010101010101010=0
4: Slow 32.5/84 11111===0111=01010=011101=00=0110=01010111110=0100000=0=0000000=000000000=0000010100
5: Ruffian_210 26.5/85 11=10=0100=101=1=1=10101==0==1=0=10==1==0001=0000000=000000000=0000000=0=0=000=000000
6: List512 26.0/84 =011110001010100010101010101==00000=0000000101010=00000000000000010=010=0101010=0100?
7: Yace Paderborn 19.0/85 01011=0==10=0=010100==0=000==100000000000000=00=0=010000=000000000000=0==00001000=0==

[Tony Thomas]

Dont fall in love with her, she is mine.

[Ovyron]

Dont fall in love with her, she is mine.

That's why everyone can have their copy

[Bill Rogers]

Tony
Is Romi learning like other engines in that it only improves its book?
If not can you explain how or what it does?
Thanks
Bill

[David Dahlem]

I have the original build of RomiChes P3k included in the download, and builds by Jim Ablett and Denis Mendoza. Is there any significant difference in these builds?

Regards
Dave

[Tony Thomas]

I have the original build of RomiChes P3k included in the download, and builds by Jim Ablett and Denis Mendoza. Is there any significant difference in these builds?

Regards
Dave

Only speed differenece, I am not sure by how much.

[Dr.Wael Deeb]

I have the original build of RomiChes P3k included in the download, and builds by Jim Ablett and Denis Mendoza. Is there any significant difference in these builds?

Regards
Dave

RomiChess P3k JA Intel Compiler 10 p.g.o
32bit + 64bit + auto-paralellized 2-thread versions.

[Michael Sherwin]

Tony
Is Romi learning like other engines in that it only improves its book?
If not can you explain how or what it does?
Thanks
Bill

Hi Bill,

Romi does not start off with a book (unless you merge an initial pgn file). Romi builds a book from scratch by 'remembering' every game she plays to ply 160. Romi then plays what works. If for example, Rybka beats Romi, then Romi will play back Rybka's line. When Romi looses with Rybka's line she will see how to play it from the other side.

Romi also incorperates Pavlov's dog experiments (modified for computers) by giving each winning sides moves a small bonus (that accumulates) and the loosing sides moves a small penalty (that also accumulates). The moves are stored in a tree and moved to the Hash file before the search. This causes Romi to choose different moves to avoid loosing or reinforces moves that win.

Romi's learning as a whole is very different than other engines. Some engines now have incorperated the bonus/penalty part into their learning system.

[David Dahlem]

Tony
Is Romi learning like other engines in that it only improves its book?
If not can you explain how or what it does?
Thanks
Bill

Hi Bill,

Romi does not start off with a book (unless you merge an initial pgn file).

Hi Michael

How to merge an initial pgn file? I see nothing about how to do this in the readme, or when typing 'help' in the console window.

Regards
Dave


Romi builds a book from scratch by 'remembering' every game she plays to ply 160. Romi then plays what works. If for example, Rybka beats Romi, then Romi will play back Rybka's line. When Romi looses with Rybka's line she will see how to play it from the other side.

Romi also incorperates Pavlov's dog experiments (modified for computers) by giving each winning sides moves a small bonus (that accumulates) and the loosing sides moves a small penalty (that also accumulates). The moves are stored in a tree and moved to the Hash file before the search. This causes Romi to choose different moves to avoid loosing or reinforces moves that win.

Romi's learning as a whole is very different than other engines. Some engines now have incorperated the bonus/penalty part into their learning system.

[Michael Sherwin]

Hi Michael

How to merge an initial pgn file? I see nothing about how to do this in the readme, or when typing 'help' in the console window.

Regards
Dave

Hi Dave,

Just place a stripped pgn (i.e. no comments) in the same directory as RomiChess. Then if, for example the pgn is '2600+.pgn, just type 'merge 2600+.pgn' and then 'quit' to flush and close the file system.