Different player

[joko]

Hi,
is it possible to play as different "player" ?
I try to explain what I mean:
If I play with different levels all the results have an influence in my elo-rating.
If it would be possible to choose: "Player xy normal" or "player za agressive" or "player ab defender" I have to each type an elo and I can test myself which style is the best for me.
For hiarcs I can choose normal-agressive...that option is that thing what I need for me, only to test my style without changing the elo-value
Jörg

[bob]

Hi,
is it possible to play as different "player" ?
I try to explain what I mean:
If I play with different levels all the results have an influence in my elo-rating.
If it would be possible to choose: "Player xy normal" or "player za agressive" or "player ab defender" I have to each type an elo and I can test myself which style is the best for me.
For hiarcs I can choose normal-agressive...that option is that thing what I need for me, only to test my style without changing the elo-value
Jörg

It is absolutely possible. I've now run several million test games over the past year +, and have found that a change will improve results against A, hurt against B, and be insignificant against C. We've batted the idea around of multiple "personalities" targeted to specific opponents, as an interesting conversation point. It is definitely doable. Whether we will do it or not is a different matter however..

[Uri Blass]

Hi,
is it possible to play as different "player" ?
I try to explain what I mean:
If I play with different levels all the results have an influence in my elo-rating.
If it would be possible to choose: "Player xy normal" or "player za agressive" or "player ab defender" I have to each type an elo and I can test myself which style is the best for me.
For hiarcs I can choose normal-agressive...that option is that thing what I need for me, only to test my style without changing the elo-value
Jörg

It is absolutely possible. I've now run several million test games over the past year +, and have found that a change will improve results against A, hurt against B, and be insignificant against C. We've batted the idea around of multiple "personalities" targeted to specific opponents, as an interesting conversation point. It is definitely doable. Whether we will do it or not is a different matter however..

In theory it is possible
practically I saw no proof for it.

It is possible to tune for specific position and maybe even for small set of positions(silver suite) but when looking at CEGT results or CCRL FRC results(that are based on random choice of one position out of 960 positions) I see that performance seem to be indpendent on the opponent
and it seems that difference in performance can be explained by statistical noise.

Here is the link for 40/4 FRC list when programs played match of 100 games.

http://computerchess.org.uk/ccrl/404FRC ... t_all.html

You can see the result of every program and the difference between rating and performance by clicking on the name of the program.

If performance is dependent in the opponent then I could expect to see cases when programs have performance that is more than 100 elo better than their rating

Looking at performances the maximal difference between rating and performance is only 83 elo and there are 213 different matches of 100 games so difference of 83 elo between rating and performance can clearly be explained by statistical error.

I found only 5 matches out of 213 matches when the difference between rating and performance was more than 70 elo.

HiarcsX54-Ufim8.02 -83 elo
Rybka 64 bit-Loop10.32f +79 elo
Fruit 2.2.1-Movei383 +74 elo
Hermann0.2-Hamsters 0.4 -73 elo
Hermann 1.7-Ayito 0.2.994 -72 elo


Uri

[bob]

Hi,
is it possible to play as different "player" ?
I try to explain what I mean:
If I play with different levels all the results have an influence in my elo-rating.
If it would be possible to choose: "Player xy normal" or "player za agressive" or "player ab defender" I have to each type an elo and I can test myself which style is the best for me.
For hiarcs I can choose normal-agressive...that option is that thing what I need for me, only to test my style without changing the elo-value
Jörg

It is absolutely possible. I've now run several million test games over the past year +, and have found that a change will improve results against A, hurt against B, and be insignificant against C. We've batted the idea around of multiple "personalities" targeted to specific opponents, as an interesting conversation point. It is definitely doable. Whether we will do it or not is a different matter however..

In theory it is possible
practically I saw no proof for it.

It is possible to tune for specific position and maybe even for small set of positions(silver suite) but when looking at CEGT results or CCRL FRC results(that are based on random choice of one position out of 960 positions) I see that performance seem to be indpendent on the opponent
and it seems that difference in performance can be explained by statistical noise.

Here is the link for 40/4 FRC list when programs played match of 100 games.

http://computerchess.org.uk/ccrl/404FRC ... t_all.html

You can see the result of every program and the difference between rating and performance by clicking on the name of the program.

If performance is dependent in the opponent then I could expect to see cases when programs have performance that is more than 100 elo better than their rating

Looking at performances the maximal difference between rating and performance is only 83 elo and there are 213 different matches of 100 games so difference of 83 elo between rating and performance can clearly be explained by statistical error.

I found only 5 matches out of 213 matches when the difference between rating and performance was more than 70 elo.

HiarcsX54-Ufim8.02 -83 elo
Rybka 64 bit-Loop10.32f +79 elo
Fruit 2.2.1-Movei383 +74 elo
Hermann0.2-Hamsters 0.4 -73 elo
Hermann 1.7-Ayito 0.2.994 -72 elo


Uri

I have seen the proof. We have _lots_ of test results where a change makes a significant difference against different opponents. One change plays better against fruit, worse against glaurung. And vice-versa for another change. We have a weighting scheme to choose whether to accept or reject a change, but if we wanted we could certainly have opponent-specific versions that would play much better against a specific opponent. We are more interested in general-purpose solutions however...