romi : how much can learning improve romi

[ozziejoe]

wow, a plethora of engines with learning.

Any sense of how good the learning features are for prodeo and spike?

[Tony Thomas]

wow, a plethora of engines with learning.

Any sense of how good the learning features are for prodeo and spike?

Almost as good as Romi's. Note that Prodeo started doing similar learning after Romi was released, so who knows it might be the exact same thing.

[ozziejoe]

if prodeo has same algorithm as romi, and is 200 pnts better, maybe i will evaluate it.

best
J

[Michael Sherwin]

if prodeo has same algorithm as romi, and is 200 pnts better, maybe i will evaluate it.

best
J

I do not believe that Prodeo has the 'monkey see monkey do' algorithm. Ed claims to have had the other algorithm since the mid 90's, but never enabled it in the release version untill recently. I personally believe that he added the bonus/penalty very recently and then enabled the learning, because, the bonus/penalty is what made it effective. But, I do not know for sure.

In a test of mine against Spike 1.2, Romi went from 6% to about 15% after a thousand games and then fell back to about 9% after Spikes learning kicked in and two thousand games more.

[ozziejoe]

HI michael,

Well, I'm not completly sure the prodeo learning is working right. The way it is described, you should get instant moving like you do with romi (when it is on a line that has won before). It doesn't do that. The position.lrn file is updated after each game, but after 63 games it has only been updated 2 kb. . It is suppose to be saving the moves from ever game and changing weights etc, but i don't think it is.

In contrast, romi's learnign file gets larger faster

[Ovyron]

I personally believe that he added the bonus/penalty very recently and then enabled the learning, because, the bonus/penalty is what made it effective.

Ed added the bonus/penalty learning since June 2007 and considered all the older learning behavior as obsolete, switching it off.

Well, I'm not completly sure the prodeo learning is working right. The way it is described, you should get instant moving like you do with romi (when it is on a line that has won before). It doesn't do that. The position.lrn file is updated after each game, but after 63 games it has only been updated 2 kb. . It is suppose to be saving the moves from ever game and changing weights etc, but i don't think it is.

ProDeo 1.5 doesn't move instantly on learned moves, it just changes its evaluation weights so they get more chance of being played.

It already comes with learning from 6000 games, so perhaps its learning file is already big. Also, you may want to switch off ProDeo's internal book, since the book doesn't learn.

To check if ProDeo has learning ON switched correctly, play e4 against it and it should always move c5, without book.

[ozziejoe]

i've actually gone back to prodeo 1.2, because 1.5 was not reading the parameters from rebel.eng. Are you saying the extended book learning = read&write is on? Because, by default, it looks to be offf?

I also didn't want the 6000 games already in as a confound.

This is the position i am evaluating, and prodeo is out of book here

[d]rnbqkbnr/pp2pppp/8/2ppP3/2PP4/8/PP3PPP/RNBQKBNR b KQkq c3 0 4[d]

[ozziejoe]

Here is how 1.2 has been doing in this position. You think a decent learning algorithim would lead to a draw against shredder on occasion. I'll try default prodeo 1.5, and see if it does better. I doubt the 6000 game database has many (if any) games relevant to this opening


major eval of prodeo-1 2008

Pro Deo 1.2 - Chessmaster 11 4.0 - 10.0 +4/-10/=0 28.57%
Pro Deo 1.2 - Fritz 9 2.0 - 12.0 +1/-11/=2 14.29%
Pro Deo 1.2 - Shredder 10 UCI 0.0 - 13.0 +0/-13/=0 0.00%
Pro Deo 1.2 - Spike 1.2 Turin 9.0 - 4.0 +7/-2/=4 69.23%
Pro Deo 1.2 - Rybka 2.3.2a 32-bit 4.5 - 8.5 +3/-7/=3 34.62%

[Ovyron]

i've actually gone back to prodeo 1.2

ProDeo 1.2 doesn't have the kind of learning we're talking about... (It was implemented in 1.5)

[ozziejoe]

ok, i've started testing 1.5 and leaving all defaults alone (although i am not convinced it correctly reads the personality file which controls the paramters)


The webpage says this is the kind of positional learning that is on by default.:

Pro Deo will recognize games it has lost and when faced a second (third, fourth etc.) time it will adapt its playing style until it has found new and good settings for the type of opening it is playing, or is forced to play.



I am not sure this is the kind of learning we were talking about? Does this mean prodeo varies its aggresiveness or something like that, until it finds a set of parameter weights that work?

best
J