Sooo, anyone cares for experiment?
ā¦and if you do, go ahead and install GrapheneOS on it too! Heaps of privacy and security improvements, super simple install process, easily 99% app compatibility. Get on it Pixel owners 
1 Like
I make mistakes with either stronger or weaker opponents so I learn from both. I just use the auto matcher and it works pretty well.
And honestly at 18-19k in rank, you can easily lose via castarophic mistake to either something 2-3 stones above or below you. When I play teaching games against weaker players, the big jump comes around when you reach 12-13k and then again at 7-8k. Much less mistakes when you reach those cut-offs.
4 Likes
Iām a bit confused too, but perhaps the issue is with the very large sample that would be required to assess low probabilities?
To take an extreme example, letās say A and B are beginners, and if they play games against opponents close to their levels it quickly becomes apparent that B is stronger than A. But they only play against quite strong players, and while A has perhaps 1% chance to win, and B has 3% (3x more), such difference would only become statistically visible over hundreds of games. In practice A and B just both lost all their games, and so the ranking is inaccurate.
Or maybe Iām saying complete nonsense, that works too.
1 Like
Itās more subtle than that.
Glicko is an important factor.
Suppose I continuously play someone 2 ranks stronger than me, and my rank is correctly estimated.
I will start by losing ~75% (*) of games against them. Each time I lose, my rank will go down slightly (because of glicko). Then I will be under ranked. When I win (~25% of the time) my rank leaps up a lot (because of glicko) and puts me back in about the right place ā¦ so on some kind of long term average I am correctly rated, but the tendency is for me to be āat or below my true rankā.
If you do this same argument for always playing weaker players, you will find that the tendency is to be āat or above my true rankā.
Because of glicko, it averages out pretty well, but at any given time, if you are playing āupā you will experience periods of being wacked down, and if you are playing ādownā you will experience periods of artificial inflation.
(*) or whatever the percentage is, itās āsomething like thatā.
3 Likes
Thereāre non-ranking games, you can challenge higher players as many times as you canā¦
And if you want to get higher rank, I suggest to play with AI. Because AI can be adjusted easily, AI donāt eat, drink or sleep, AI move stones fastā¦ AI is the answer for these who hope to play Go betterā¦
I think you could also win before you lose when playing stronger players (WLLL is just as likely as LLLW), so it doesnāt seem like playing āupā will necessarily bring your rank down (at least according to the assumptions of Glicko)
3 Likes
Well the problem was that my account is quite new so I quickly played a lot of ranked games (almost exclusively against much stronger opponents) so I could get a ranking since there are several features (primarily tournaments and ladders) that are not available if you donāt have a ranking. I guess the rank will sort it self out with time although I should probably try to play a better mix of ranks in the future.
1 Like
I think rating systems can handle some unbalance in games against stronger opponents and games against weaker opponents. Only if your overall winrate is quite extreme (like less than 10% or more than 90%), your rating could be deflated or inflated.
3 Likes
Iām not sure on the assumptions of Glicko: I havenāt read them.
I would have thought that it assumes that you play a random mix of ranks. Thatād be the āintuitive assumptionā.
I donāt really see how you can design a system that becomes sure about a personās rating if by playing down that person never loses, for example.
You need some losses to work out the maximum possible rating the person could be.
It seems to follow that an even number of losses and wins provides the best estimate of rating.
So much for theory which I am just guessing at.
My own experience is that I can drive my rank in either direction, in the short term, by chosing to play the opposite direction. And that this doesnāt ālastā. IE I havenāt experienced being able to continually drive myself up by playing down. As it should be.
1 Like
Glicko (and Elo) should be able to handle rank mismatches. The thought is that you sometimes win against stronger people albeit with small probability.
For instance, if Iām 7k, but I only ever play 5ks, you can still have a decent idea of my rank because Iāll mostly lose, but I play well enough to win on occasion. Obviously this breaks down as you reach the extremes as @gennan pointed out.
But I think itās less about the win:loss ratio or the stronger:weaker ratio and more about how close your opponents are. I could play both amybot and katago all day and never find out my rank, but if I consistently play people 1 stone above me, I will know my rank with relatively high confidence.
Edit: so basically, I think we agree that you need a sizeable amount of losses, but not that you need to play a mix of stronger and weaker to achieve that.
5 Likes
Itād be great to run the maths on that.
What rank does glicko settle down to chosing the strategy of āconsistently play people 1 stone above meā, given a ātrue rankā of say 10k (starting at 9.5k precisely)?
What rank does glicko settle down to chosing the strategy of āconsistently play people 1 stone below meā, given a ātrue rankā of say 10k (starting at 9.5k precisely)?
These are questions that have concrete answers in a for loop with some maths 
Remember that you have to take into account that while the rating changes, the person will be continuing to chose an opponent of 1 stone higher. This brings a hard-to-intuit effect ā¦ we arenāt chosing 1-stone-worth-of-glicko-rating-points, we are chosing opponens with a random selection of ratings in the glicko range that is 1 stone above our rank ā¦ which also remains constant while our rating moves in the 10k range.
Iām chasing server glitches, if someone else could do the math 
3 Likes
Best of luck smashing server bugs!
2 Likes
If this is the case, Iād say that you tend to do better than expected against stronger players and worse than expected against weaker players. Perhaps you play a bit more carefully against stronger players and a bit more carelessly against weaker players?
I win against weaker players and lose against stronger ones.
In the short term, I donāt lose often enough against weak players to wack my rating down again.
I know that if I kept doing it, that would happen from time to time, so there would be a sawtooth.
Oh sorry, so itās the other way around than what I said. You tend to do worse than expected against stronger players and better than expected against weaker players.
This could be explained in different ways: It can be a trait of your player personality/style, or it can be that the predictions of the system in your rating range are off (maybe due to the rank-rating conversion formula around your level), or it can be a combination of those.
Some sawtooth effect can be expected if you play against players who are significantly weaker/stronger (involving a very high/low winning probability). Similar to how your money balance looks when you participate in some lottery.
I knew it. Go is just a game with luck.
2 Likes
When you particiate in a FAIR lottery. So never!
Btw is there some math modelisation of sawteeth? A name of an application?
I think sawtooth is actually wrong for this, itās more like shark-fin.
This is a way of saying āwhat you experienced is very likely due to personal factorsā.
I canāt deny that, but on the other hand purely intuition is telling me that I bet this is common.
Averages are one thing, short term outcomes that we experience are another.
This is about āthe experience of not getting bad impacts on my rankā (to quote the OP).
My feeling is that if you play down when your rank took a hit and play up when youāre getting ahead, maintaining about 50/50 up and down, your rank will be best served, and so will the community (who are all also trying to get games up and down).