Life In 19x19

Kirby, I'm not going to argue with you about what "statistical inference" means. You seem to know something about how inference works, or you wouldn't be throwing around jargon like maximum likelihood. But at the same time, you are throwing up ink in a way that has poor Splatted saying:

Splatted wrote:It's a very limited data set but those wins/losses are still statistics.

For those who care: any number can be a "statistic". A "statistical prediction" or "inference" is one that applies statistical techniques to data to find the best prediction. "You won 17 games! Here's a gold star!" is not a statistical technique, and the gold star is not a prediction.

jts wrote:Kirby, I'm not going to argue with you about what "statistical inference" means. You seem to know something about how inference works, or you wouldn't be throwing around jargon like maximum likelihood. But at the same time, you are throwing up ink in a way that has poor Splatted saying:

Splatted wrote:It's a very limited data set but those wins/losses are still statistics.

For those who care: any number can be a "statistic". A "statistical prediction" or "inference" is one that applies statistical techniques to data to find the best prediction. "You won 17 games! Here's a gold star!" is not a statistical technique, and the gold star is not a prediction.

This is just plain insulting and completely out of line. Neither you nor said XedOver said Tygem ranks aren't a statistical prediction, you said they aren't statistical because they don't make predictions.

However, the Tygem approach is not statistical in that it makes no predictions.

My point was that there are many other ways in which something can be statistical. "You won 17 games! Here's a gold star!" is a statistical technique, it's just a very basic one.

Edit: Or rather, a system that awards a gold star for 17 won games would be statistical.

It seems like we've reached the most rewarding part to disagree on - definitions

For what it's worth, I agree with Kirby and Splatted that the Tygem system is statistical, but would accept that some people may consider sophistication as part of their definitions. I do disagree with jts's initial comment that the Tygem system does not attempt to make predictions though, as I imagine that would be implicit in any ranking system.

jts, statistical analysis doesn't need to be complicated. There are complicated techniques and models that sometimes work well, yes, but it's not a requirement.

To give an example, you can take an elementary language processing course, and one of the first techniques used to determine a maximum likelihood estimate of a word, given a set of training text, is to simply count up the words and take the average to determine the probability of seeing a given word. You can obtain a more advanced model, but it's still true that a basic measurement is a statistical method for estimating uncertainty. Various ranking systems like KGS's, Tygem's, etc., are all statistical methods for determining rank.

But the main point I want to make here is that, just because a given model is more mathematically complicated, it doesn't necessarily make it a better predictor. Overfitting in machine learning is a good example of this.

When math gets complicated, sometimes people forget that the given statistical method is still subject to its own base model and assumptions. Stats are useful, but the only thing that can fairly measure how good a model is track record.

So it comes down to, when you play on KGS or you play on Tygem, in which case do you feel you are having a better playing experience?

This varies from person to person, so it is not accurate to argue that a given model is "more statistical" or "superior". The model simply makes different assumptions, and KGS's happens to use more math.

Sorry, I apologize. Use "statistical inference" to mean whatever you want. All numbers are statistics, and anytime anyone says anything about numbers, that's statistics too. A t-test is just when you g-give a test to a f-fellow with a stutter; and a parameter is about six anahalf feet. The fact that there are multiple threads about how unpredictable Tygem ranks are is surely part of the mystery of Korean play-style, like Kirby says, and has nothing to do with how that server determines ranks.

Kirby wrote:But the main point I want to make here is that, just because a given model is more mathematically complicated, it doesn't necessarily make it a better predictor.

but it can be tested and actual comparisons made.

perhaps you'd like to read Remi Coulom's paper
http://remi.coulom.free.fr/WHR/

jts wrote:Sorry, I apologize. Use "statistical inference" to mean whatever you want. All numbers are statistics, and anytime anyone says anything about numbers, that's statistics too. A t-test is just when you g-give a test to a f-fellow with a stutter; and a parameter is about six anahalf feet. The fact that there are multiple threads about how unpredictable Tygem ranks are is surely part of the mystery of Korean play-style, like Kirby says, and has nothing to do with how that server determines ranks.

I'm not sure if that deserves a response, but the unpredictability of Tygem ranks is primarily due to sandbaggers and the fact that you can register anywhere from 18k to 5d.

jts wrote:Sorry, I apologize. Use "statistical inference" to mean whatever you want. All numbers are statistics, and anytime anyone says anything about numbers, that's statistics too. A t-test is just when you g-give a test to a f-fellow with a stutter; and a parameter is about six anahalf feet. The fact that there are multiple threads about how unpredictable Tygem ranks are is surely part of the mystery of Korean play-style, like Kirby says, and has nothing to do with how that server determines ranks.

If you feel that the Tygem ranks are unpredictable, and if you feel that they do not adequately suit you with a proper opponent, then it is indeed an indication that the ranking system does not fulfill its purpose from your perspective. As I stated in the earlier post, there is no perfect ranking system, and they all have at least some set of assumptions and model upon which the ranking system is based. I also stated that the only real measure you can have of the system is how well you feel that the rank pairs you up appropriately with opponents. Precise measurements are only from the context of a given model with some set of assumptions.

And I do not intend to use "statistical inference" to mean whatever I want. I have provided what I hoped would be an objective definition - I selected it from Wikipedia, but I'd be open to using other sources.

I've also indicated examples, such as in natural language processing, where simple measures of statistics are used. Taking the average word count for a given word from a set of text, for example, seems like a simple and basic measurement to use for predicting the likelihood of a given word in new text, but it's still used, and it's still a statistic.

I have done my best to give an objective definition of statistical inference, and I'm pointing out that this definition can meet a wide variety of cases, whether the analysis be simple or complex. Based on your responses, it appears that you prefer the KGS ranking system to Tygem's. Perhaps you feel that when you play on KGS, you are more consistently paired against opponents that are close to your skill-level. This is perfectly fine, and it's a fair opinion to have.

But to say that KGS's system is "statistical" and that Tygem's is not is not a fair statement based on the definition of statistical inference that I've posted, and you have not indicated an objective reason to the contrary.

Please do not accuse me of defining "statistical inference" to mean whatever I want, as I have tried my best to give an objective definition. And between the two of us, I don't think that I am the one that has failed to provide a definition statistical inference from an external source.

Maybe you don't like Wikipedia's definition. Then, please link us to another one. Perhaps we can find a definition for which Tygem's ranking system is not "statistical".

xed_over wrote:

Kirby wrote:But the main point I want to make here is that, just because a given model is more mathematically complicated, it doesn't necessarily make it a better predictor.

but it can be tested and actual comparisons made.

perhaps you'd like to read Remi Coulom's paper
http://remi.coulom.free.fr/WHR/

Certainly, I absolutely agree, and I actually have read the paper in the past. From the paper:

The method to compare algorithms consisted in measuring their prediction rates over a database of games...

In a more general sense, ranking systems can be evaluated based on their accuracy given an objective sample set of data. On a personal level, this corresponds to my earlier explanation of how I feel a ranking system should be tested - if you feel that the system suits you up with appropriate opponents, then that's what matters.

This could be done based on your own feeling, or you could take a more systematic approach. For example, you could play 100 games on Tygem and 100 games on KGS, and then compare your win rate to the rank that you're given by the ranking system. You could analyze this in a number of ways.

But for better or for worse, empirical results are the indicator here, and not the idea that "this system is more statistical than another system".

And I personally believe that it's possible that if you and I were to both do an empirical analysis to determine the best ranking system, we could achieve different results based on the sample data we have (eg. from the games we each play). For example, I might have the bad luck of playing against a sandbagger on KGS and you might have the bad luck of playing against a sandbagger on Tygem.

In any case, I totally agree. We can make attempts to compare the systems based on sample data, but those attempts will still be subject to what we're measuring in the study, the data we sample from, etc. And the KGS ranking system certainly isn't better than Tygem simply for being "more statistical".

Just to get back to the question...

There is definitely a stylistic difference between the two servers. Fights that involve enormous groups are common on Tygem, rare on KGS. For me anyway. I'm 6k on KGS and hovering between 8 and 9k on Tygem. I have recently started winning more games on KGS, directly as a result of playing more on Tygem I think. Playing on the 2 servers also allows me to reduce playing on KGS, which means my rank is more sensitive to results.

Probably your'e still in sandbagg zone.

Register as a 1d on tygem, you'll be able to beat about 50%
I'm a 3k on kgs and 1d on tygem and mostly I'm able to keep my rank.

Edit: I do play on the chinese servers, not sure if that makes any difference.

Cheers,
Otenki

jts wrote:All numbers are statistics, and anytime anyone says anything about numbers, that's statistics too. A t-test is just when you g-give a test to a f-fellow with a stutter; and a parameter is about six anahalf feet.

Posts like this make me wish for a thumbs down vote. Obnoxious on several levels.

What's the best thing you can say about a statistical analysis?

DJLLAP wrote:

Who has similar or contrasting experiences they would like to share?

Yes. Six years ago i was 7 kyu on kgs with ease, but couldnt get above 22 kyu on wbaduk. Was somewhat disappointing. The reason I got better, was learning to punish their overplays right. Often by outright ignoring them or do some simple shape move he has to answer and takin sente.

I made the experience that on wbaduk the higher the rank the stronger the skill. I didnt lose against a 9k for a long time, and have never won against a 4k. So there must be some correlation. Its just there the rank has to be taken with a + or - 1 or 2. Except for the sandbaggers obviously. But often i cant really say if its an sandbagger or if i just play like junk.

On the gamestyle: I have the feeling on wbaduk it's more about reading than about opening and shape. Those guys raeally can't surrender, and you have to play all out and defend a 40 point margin against all their overplay attacks and really count well that semeai. As it's a bit humilating to just defend the safe way against overplays i play at the hard edge and so often lose won games. Maybe he just fills in the semeai at the end, that i think i win surely counting to 5 liberties, and only playing on the last possible moment. And then there is some neat shortage of liberties tesuji that i oversaw and gives him one liberty more and the tide turns 100 points. Sometimes i play safe, but like i said: I's boring.

otenki wrote:Probably your'e still in sandbagg zone.

Register as a 1d on tygem, you'll be able to beat about 50%
I'm a 3k on kgs and 1d on tygem and mostly I'm able to keep my rank.

Edit: I do play on the chinese servers, not sure if that makes any difference.

Cheers,
Otenki

I don't find many sandbaggers at the 10k level on tygem. Occasionally I play someone who appears to be miles ahead of me, but that happens on KGS as well. I chalk it up as a loss, review it, and try again. On the whole I appear to be evenly matched at this level, much to my continued frustration!