romi : how much can learning improve romi

[ozziejoe]

I am pitting romi against the engines below (well ruffian 2.1 instead of ruffian 1.05).



Romi has trained on this opening position in 1 min 2 sec increment games: 1)e4, c6; 2) d4, d5; 3)e5, c5; 4) c4!? (about 400 games) This interesting position has been giving white trouble in the human database, but has not been played alot.


The tourni now involves 2 min with 2 sec increment. Romi keeps learning. Do you think romi's learning will allow it to beat all these stronger engines ? I'll post results as they occur



ed on January 25, 2008 with Bayeselo based on 237'120 games
Rank Name Rating Score Average
Opponent Draws Games LOS
ELO + −
1 Naum 2.0 32-bit 2797 +8 −8 48.9% +2.9 32.1% 5893
100.0%
2 Movei 00.8.438 2703 +16 −16 39.9% +68.8 30.8% 1329
74.7%
3 List 5.12 2696 +11 −11 48.8% +8.6 29.3% 3022
92.5%
4 SOS 5.1 2685 +12 −12 46.7% +23.0 29.0% 2633
75.8%
5 Ruffian 1.0.5 2679 +12 −12 49.3% +4.0 31.9% 2414
90.2%
6 Aristarch 4.50 2668 +12 −12 45.8% +31.9 27.4% 2771
100.0%
7 Yace 0.99.87 2612 +17 −17 50.5% −4.7 30.8% 1248
100.0%
8 RomiChess P3kLD 2468 +20 −20 38.7% +83.1 24.3% 931

[ozziejoe]

ps it might be fun for someone to run a similar test with another interesting opening. You see the learning value maiximized when you focus on one opening, rather than trying to increase strength in all openings

[Ovyron]

It would be interesting to have it play against ProDeo 1.6 BETA with learning ON, as ProDeo would be learning how to beat Romi while she tries to learn how to beat ProDeo

[ozziejoe]

how reliable is this version of prodeo. I have had a few problems with it in the past.

Also, does prodeo use the exact same learning function (if i win, repeat what i did; if you win, repeat what you did)

J

[ozziejoe]

can someone send me a link to prodeo 1.6 beta? I'd love to take a look at its learning

best
J

[Tony Thomas]

Isnt Prodeo's learning awfully similar to the learning of Romi?

[ozziejoe]

maybe, but prodeo looks to be over 200 pnts stronger?

[ozziejoe]

First round, and romi struggles of course. Lets see if it can improve


1: Romichess 4.5/14 ··
2: List512 2.0/2 11
2: Ruffian_210 2.0/2 11
2: Naum 2.0/2 11
5: Yace Paderborn 1.0/2 01
5: SOS 5.1 for Arena 1.0/2 10
5: Movei00_8_438 1.0/2 01
8: Aristarch 4.50 0.5/2 0=

[Tony Thomas]

First round, and romi struggles of course. Lets see if it can improve


1: Romichess 4.5/14 ··
2: List512 2.0/2 11
2: Ruffian_210 2.0/2 11
2: Naum 2.0/2 11
5: Yace Paderborn 1.0/2 01
5: SOS 5.1 for Arena 1.0/2 10
5: Movei00_8_438 1.0/2 01
8: Aristarch 4.50 0.5/2 0=

If you really want a quick improvement, you should run single matches. For example, after about 200 matches against List, Romi should start winning. Prodeo is stronger, but since it has a similar learning mechanism, I doubt that Romi will be able to close the gap. Other engines with goo learning(that I can think of) are Spike and Frenzee.

[Ovyron]

how reliable is this version of prodeo. I have had a few problems with it in the past.

Also, does prodeo use the exact same learning function (if i win, repeat what i did; if you win, repeat what you did)

J

Yes, ProDeo does pretty much the same. Anyway, I'd recommend you to just use ProDeo 1.5, as under certain conditions it's stronger than 1.6b and you have learning ON activated by default (to get 1.6 Beta you need to join a private forum and wait for permission etc.)

Download 1.5 from http://www.top-5000.nl/prodeo.htm

(And, sorry, I didn't know that ProDeo was so strong, so maybe Romi needs more "training" before facing it.)