|
Ryze gained a lot of power from tp meta and lane bully nerfs. Without Renekton, pre-nerfs Darius, Irelia or other ignite bruisers to keep him in check he was allowed to get to Ryze status.
I still think that kit was the most manageable of the ideas Riot has tried since.
|
On July 31 2016 15:19 Goumindong wrote:If you're thinking about dodging then it's better to just use one team. Which should also let you increase the complexity of the calculation(I.E. Include more interaction variables) and should negate your deficiency issue. Additionally with two teams you should be able to code a single champion as 0,1,-1 for "not present", "blue", red" on a single variable. Which should also iirc fix your rank deficiency issue. The enemy team won't matter in the long run to dodging except insomuch as your tool modifies what is played so just going to the one team is probably ideal even if it's less accuracy in the model Edit: it's been quite a while since I've done any work in that area so I could be way off base. Also not sure if your models have input requirements with regards to a trinary structure. (Red side/blue side effect should be the same so it should work fine) . Show nested quote +On July 31 2016 08:44 GrandInquisitor wrote:
If I remember correctly, original Ryze was OK for a long time until various meta changes and people waking up led to the realization that he was a total terror top lane. Once that happened they had to Olaf him while they tried to figure out how to rework him. He was simultaneously a lane bully but also a hyperscaler, which was completely obnoxious.
Original ryze was on/off strong. He left the meta when banshees (and something else iirc) lost mana. He came back with tear and roa buffs. (Mainly). Which is what caused him to be reworked.
Come on GMD I expected you to try harder. This is seriously disappointing.
1. I am not going to restrict myself to only one team: there is absolutely no reason to throw away half of my data. In particular, I am not going to throw away half of my data due to rank deficiency because rank deficiency can be easily solved via regularization. I'll leave this to you as an exercise to explain why regularization solves the rank deficiency problem.
2. It's already coded as 0, 1, -1 for "not present", "blue", and "red" - this is exactly why there is rank deficiency. I'll leave this again as an exercise for you.
|
On July 31 2016 19:18 JonnyLaw wrote:
Ryze gained a lot of power from tp meta and lane bully nerfs. Without Renekton, pre-nerfs Darius, Irelia or other ignite bruisers to keep him in check he was allowed to get to Ryze status.
I still think that kit was the most manageable of the ideas Riot has tried since.
lol old ryze was counter to all champion u listed
|
On July 31 2016 23:24 Sufficiency wrote:Show nested quote +On July 31 2016 15:19 Goumindong wrote:If you're thinking about dodging then it's better to just use one team. Which should also let you increase the complexity of the calculation(I.E. Include more interaction variables) and should negate your deficiency issue. Additionally with two teams you should be able to code a single champion as 0,1,-1 for "not present", "blue", red" on a single variable. Which should also iirc fix your rank deficiency issue. The enemy team won't matter in the long run to dodging except insomuch as your tool modifies what is played so just going to the one team is probably ideal even if it's less accuracy in the model Edit: it's been quite a while since I've done any work in that area so I could be way off base. Also not sure if your models have input requirements with regards to a trinary structure. (Red side/blue side effect should be the same so it should work fine) . On July 31 2016 08:44 GrandInquisitor wrote:
If I remember correctly, original Ryze was OK for a long time until various meta changes and people waking up led to the realization that he was a total terror top lane. Once that happened they had to Olaf him while they tried to figure out how to rework him. He was simultaneously a lane bully but also a hyperscaler, which was completely obnoxious.
Original ryze was on/off strong. He left the meta when banshees (and something else iirc) lost mana. He came back with tear and roa buffs. (Mainly). Which is what caused him to be reworked. Come on GMD I expected you to try harder. This is seriously disappointing. 1. I am not going to restrict myself to only one team: there is absolutely no reason to throw away half of my data. In particular, I am not going to throw away half of my data due to rank deficiency because rank deficiency can be easily solved via regularization. I'll leave this to you as an exercise to explain why regularization solves the rank deficiency problem. 2. It's already coded as 0, 1, -1 for "not present", "blue", and "red" - this is exactly why there is rank deficiency. I'll leave this again as an exercise for you.
I agree with him that a single team search is a better tool; would be nice to see how your team stacks vs an average aram team. If you could generate a single team win rate vs median that would be awesome, actually.
Although perhaps that would be a different tool.
|
^^ That was my point wrt the "single team". Its true that the full team is a better predictor, obviously. But also simply observing the end of the game and recording it is a better predictor, obviously. But neither are useful and we want information we can use, not information we can't. Also not sure why you would have to throw out half your data[half your variables yes, but not half your data]*.
WRT Rank deficiency: Its been a while but you should only be rank deficient when you've got not enough observations for your dimensionality, or you're singular. You should not be singular on the 0,1,-1 input unless i am seriously missing some obvious co-linearity. And you should have enough data points. I suppose you could just have an ill conditioned matrix but not seeing why you would have one by just by knowing the input numbers.
*You could potentially double your data because you get two observations from each game. This could causes issues but i am blanking on what they would be.
|
If you observe both teams half your data points are directly dependent on the other half of your data. Ie if team red wins team blue losses. If you count both teams as data point you are considering yourself to have twice as much independent data as you actually do.
|
He left the meta when banshees (and something else iirc) lost mana.
Banshees lost mana in patch 3.10, Ryze got his major update in 5.8. I recall him still being meta between that. Was there another mana item he built besides Archangel, RoA, FH and Banshees? I kinda remember him having no real sixth item to build. Maybe you're thinking of the FH mana nerf, although that was quite some time before.
|
On August 01 2016 04:29 General_Winter wrote:
If you observe both teams half your data points are directly dependent on the other half of your data. Ie if team red wins team blue losses. If you count both teams as data point you are considering yourself to have twice as much independent data as you actually do.
If you double the data kindof. I felt like it should make x'e no longer zero, but wasn't sure off the top of my head. And didn't(and am not going to) check. If sufficiency wants to use the data he can figure it out.
|
On August 01 2016 02:02 iCanada wrote:Show nested quote +On July 31 2016 23:24 Sufficiency wrote:On July 31 2016 15:19 Goumindong wrote:If you're thinking about dodging then it's better to just use one team. Which should also let you increase the complexity of the calculation(I.E. Include more interaction variables) and should negate your deficiency issue. Additionally with two teams you should be able to code a single champion as 0,1,-1 for "not present", "blue", red" on a single variable. Which should also iirc fix your rank deficiency issue. The enemy team won't matter in the long run to dodging except insomuch as your tool modifies what is played so just going to the one team is probably ideal even if it's less accuracy in the model Edit: it's been quite a while since I've done any work in that area so I could be way off base. Also not sure if your models have input requirements with regards to a trinary structure. (Red side/blue side effect should be the same so it should work fine) . On July 31 2016 08:44 GrandInquisitor wrote:
If I remember correctly, original Ryze was OK for a long time until various meta changes and people waking up led to the realization that he was a total terror top lane. Once that happened they had to Olaf him while they tried to figure out how to rework him. He was simultaneously a lane bully but also a hyperscaler, which was completely obnoxious.
Original ryze was on/off strong. He left the meta when banshees (and something else iirc) lost mana. He came back with tear and roa buffs. (Mainly). Which is what caused him to be reworked. Come on GMD I expected you to try harder. This is seriously disappointing. 1. I am not going to restrict myself to only one team: there is absolutely no reason to throw away half of my data. In particular, I am not going to throw away half of my data due to rank deficiency because rank deficiency can be easily solved via regularization. I'll leave this to you as an exercise to explain why regularization solves the rank deficiency problem. 2. It's already coded as 0, 1, -1 for "not present", "blue", and "red" - this is exactly why there is rank deficiency. I'll leave this again as an exercise for you. I agree with him that a single team search is a better tool; would be nice to see how your team stacks vs an average aram team. If you could generate a single team win rate vs median that would be awesome, actually. Although perhaps that would be a different tool.
An average ARAM team is an interesting idea. In my mind it might be annoying to do because of the free champions changing week to week... So your average ARAM team also changes.
Perhaps a more interesting idea is to do this in the setting of ranked solo queue instead: I give you the team comp of the opposing team, you draft a team against it, then my model will evaluate your team comp....
|
If you were very sophisticated you could combine them. Given an enemy pick you would have an expected enemy team based on that pick and so a comp of your own to compare/create. This way you could potentially identify bad drafting options ahead of time.
I.E. I know the enemy picked Snydra first, how does my proposed team compare to the hypothetical syndra team? Do you have enough data to deal with the interaction terms to make that meaningful?
|
pick malphite or zac or leona something and bop the syndra
|
Hey, they picked an immobile champion.
Let's pick champions that are good versus that.
What champions are good versus that?
Well, what about champions with gapclose+CC combos.
Hey, that works!
It's not rocket science.
|
On August 01 2016 05:55 Sufficiency wrote:Show nested quote +On August 01 2016 02:02 iCanada wrote:On July 31 2016 23:24 Sufficiency wrote:On July 31 2016 15:19 Goumindong wrote:If you're thinking about dodging then it's better to just use one team. Which should also let you increase the complexity of the calculation(I.E. Include more interaction variables) and should negate your deficiency issue. Additionally with two teams you should be able to code a single champion as 0,1,-1 for "not present", "blue", red" on a single variable. Which should also iirc fix your rank deficiency issue. The enemy team won't matter in the long run to dodging except insomuch as your tool modifies what is played so just going to the one team is probably ideal even if it's less accuracy in the model Edit: it's been quite a while since I've done any work in that area so I could be way off base. Also not sure if your models have input requirements with regards to a trinary structure. (Red side/blue side effect should be the same so it should work fine) . On July 31 2016 08:44 GrandInquisitor wrote:
If I remember correctly, original Ryze was OK for a long time until various meta changes and people waking up led to the realization that he was a total terror top lane. Once that happened they had to Olaf him while they tried to figure out how to rework him. He was simultaneously a lane bully but also a hyperscaler, which was completely obnoxious.
Original ryze was on/off strong. He left the meta when banshees (and something else iirc) lost mana. He came back with tear and roa buffs. (Mainly). Which is what caused him to be reworked. Come on GMD I expected you to try harder. This is seriously disappointing. 1. I am not going to restrict myself to only one team: there is absolutely no reason to throw away half of my data. In particular, I am not going to throw away half of my data due to rank deficiency because rank deficiency can be easily solved via regularization. I'll leave this to you as an exercise to explain why regularization solves the rank deficiency problem. 2. It's already coded as 0, 1, -1 for "not present", "blue", and "red" - this is exactly why there is rank deficiency. I'll leave this again as an exercise for you. I agree with him that a single team search is a better tool; would be nice to see how your team stacks vs an average aram team. If you could generate a single team win rate vs median that would be awesome, actually. Although perhaps that would be a different tool. An average ARAM team is an interesting idea. In my mind it might be annoying to do because of the free champions changing week to week... So your average ARAM team also changes. Perhaps a more interesting idea is to do this in the setting of ranked solo queue instead: I give you the team comp of the opposing team, you draft a team against it, then my model will evaluate your team comp....
They'd both be pretty cool, actually.
I read your Reddit thread, lol. "Your model predicts the winning game 66% of the time in an environment where an algorythm places player teams together who are equally likely to win. Wow that is trash." Wut, no that is pretty sweet.
|
On August 02 2016 07:09 iCanada wrote:Show nested quote +On August 01 2016 05:55 Sufficiency wrote:On August 01 2016 02:02 iCanada wrote:On July 31 2016 23:24 Sufficiency wrote:On July 31 2016 15:19 Goumindong wrote:If you're thinking about dodging then it's better to just use one team. Which should also let you increase the complexity of the calculation(I.E. Include more interaction variables) and should negate your deficiency issue. Additionally with two teams you should be able to code a single champion as 0,1,-1 for "not present", "blue", red" on a single variable. Which should also iirc fix your rank deficiency issue. The enemy team won't matter in the long run to dodging except insomuch as your tool modifies what is played so just going to the one team is probably ideal even if it's less accuracy in the model Edit: it's been quite a while since I've done any work in that area so I could be way off base. Also not sure if your models have input requirements with regards to a trinary structure. (Red side/blue side effect should be the same so it should work fine) . On July 31 2016 08:44 GrandInquisitor wrote:
If I remember correctly, original Ryze was OK for a long time until various meta changes and people waking up led to the realization that he was a total terror top lane. Once that happened they had to Olaf him while they tried to figure out how to rework him. He was simultaneously a lane bully but also a hyperscaler, which was completely obnoxious.
Original ryze was on/off strong. He left the meta when banshees (and something else iirc) lost mana. He came back with tear and roa buffs. (Mainly). Which is what caused him to be reworked. Come on GMD I expected you to try harder. This is seriously disappointing. 1. I am not going to restrict myself to only one team: there is absolutely no reason to throw away half of my data. In particular, I am not going to throw away half of my data due to rank deficiency because rank deficiency can be easily solved via regularization. I'll leave this to you as an exercise to explain why regularization solves the rank deficiency problem. 2. It's already coded as 0, 1, -1 for "not present", "blue", and "red" - this is exactly why there is rank deficiency. I'll leave this again as an exercise for you. I agree with him that a single team search is a better tool; would be nice to see how your team stacks vs an average aram team. If you could generate a single team win rate vs median that would be awesome, actually. Although perhaps that would be a different tool. An average ARAM team is an interesting idea. In my mind it might be annoying to do because of the free champions changing week to week... So your average ARAM team also changes. Perhaps a more interesting idea is to do this in the setting of ranked solo queue instead: I give you the team comp of the opposing team, you draft a team against it, then my model will evaluate your team comp.... They'd both be pretty cool, actually. I read your Reddit thread, lol. "Your model predicts the winning game 66% of the time in an environment where an algorythm places player teams together who are equally likely to win. Wow that is trash." Wut, no that is pretty sweet.
If you could get on LoL games, a hit ratio like that would make you stupid rich.
|
On August 02 2016 07:23 ticklishmusic wrote:Show nested quote +On August 02 2016 07:09 iCanada wrote:On August 01 2016 05:55 Sufficiency wrote:On August 01 2016 02:02 iCanada wrote:On July 31 2016 23:24 Sufficiency wrote:On July 31 2016 15:19 Goumindong wrote:If you're thinking about dodging then it's better to just use one team. Which should also let you increase the complexity of the calculation(I.E. Include more interaction variables) and should negate your deficiency issue. Additionally with two teams you should be able to code a single champion as 0,1,-1 for "not present", "blue", red" on a single variable. Which should also iirc fix your rank deficiency issue. The enemy team won't matter in the long run to dodging except insomuch as your tool modifies what is played so just going to the one team is probably ideal even if it's less accuracy in the model Edit: it's been quite a while since I've done any work in that area so I could be way off base. Also not sure if your models have input requirements with regards to a trinary structure. (Red side/blue side effect should be the same so it should work fine) . On July 31 2016 08:44 GrandInquisitor wrote:
If I remember correctly, original Ryze was OK for a long time until various meta changes and people waking up led to the realization that he was a total terror top lane. Once that happened they had to Olaf him while they tried to figure out how to rework him. He was simultaneously a lane bully but also a hyperscaler, which was completely obnoxious.
Original ryze was on/off strong. He left the meta when banshees (and something else iirc) lost mana. He came back with tear and roa buffs. (Mainly). Which is what caused him to be reworked. Come on GMD I expected you to try harder. This is seriously disappointing. 1. I am not going to restrict myself to only one team: there is absolutely no reason to throw away half of my data. In particular, I am not going to throw away half of my data due to rank deficiency because rank deficiency can be easily solved via regularization. I'll leave this to you as an exercise to explain why regularization solves the rank deficiency problem. 2. It's already coded as 0, 1, -1 for "not present", "blue", and "red" - this is exactly why there is rank deficiency. I'll leave this again as an exercise for you. I agree with him that a single team search is a better tool; would be nice to see how your team stacks vs an average aram team. If you could generate a single team win rate vs median that would be awesome, actually. Although perhaps that would be a different tool. An average ARAM team is an interesting idea. In my mind it might be annoying to do because of the free champions changing week to week... So your average ARAM team also changes. Perhaps a more interesting idea is to do this in the setting of ranked solo queue instead: I give you the team comp of the opposing team, you draft a team against it, then my model will evaluate your team comp.... They'd both be pretty cool, actually. I read your Reddit thread, lol. "Your model predicts the winning game 66% of the time in an environment where an algorythm places player teams together who are equally likely to win. Wow that is trash." Wut, no that is pretty sweet. If you could get on LoL games, a hit ratio like that would make you stupid rich.
Well maybe? The key is that there isn't a 50% win chance afte champions have been randomly selected. The champions selected is information and so there isn't any reason a person could not put down those kind of numbers. It's like saltyteemo betting. It's not hard to guess right, you often come in when one team has a, more or less, "insurmountable" lead. Bots (maybe it's bots maybe not) get to like 90% accuracy. The 50% base line is before champions are selected, not after. I would wager that the majority of ARAM players hit better than 50% prediction were they to keep track.
Now I don't know the complexity of the model, whether it accounts for team composition structures, champion synergy, etc. Those are the sorts of things the human brain is quite good at sussing out and so they could easily have an advantage over an algorithm that cannot be that complex due to mathematical limits. That doesn't mean 66% isn't good but it means we don't really know if it is.
|
I think a very knowledgeable player may be able to do better, but it's sort of a null point because but we can't know either way - I can't ask such a person to rate enough games (which could be thousands of games depending on the actual performance difference) to formally demonstrate the difference.
|
|
|
I thought ARAM uses an MMR system which pushes the majority of players at 50% win rate. If that is the case dodging games or whatever seems rather pointless because your overall win rate will not change anyway.
|
|
|
On August 03 2016 19:34 Redox wrote:
I thought ARAM uses an MMR system which pushes the majority of players at 50% win rate. If that is the case dodging games or whatever seems rather pointless because your overall win rate will not change anyway.
In the long term kindof. The law of large numbers works to put the average towards 50%. But the difference between wins and losses is not an average. Even if you play all games at 50% as you play more games the difference between wins and losses can go to infinity or negative infinity or nothing at all.
And since, well you won't be at 50% until you hit your proper MMR and since dodging poor games increases that MMR. If you are about winning it's ideal to dodge bad games.
Granted you probably shouldn't care about winning ARAMs but I bet loads do.
|
|
|
|
|
|
EPT
