EPTGeneral Discussion - Page 4473
| Forum Index > Dota 2 General |
IRC chatter should remain in IRC - http://webchat.quakenet.org/?channels=tl.dota2 Posts that relate to topics with their own thread, such as in News, Tournaments or Strategy should go in those threads. | ||
|
Steveling
Greece10806 Posts
| ||
|
devilesk
United States140 Posts
On December 06 2013 04:38 hariooo wrote: People will tear you apart on this forum if you build a Battlefury on a Bounty Hunter but if you tell someone that there is a mathematical reason why some of their basic complaints with DSR are invalid, suddenly everyone gets extremely sensitive. Are my complaints about DSR invalid? No one seems to actually want to seriously discuss DSR. | ||
|
njt7
Sweden769 Posts
| ||
|
PassiveAce
United States18076 Posts
There really isnt that much to talk about. Its a rating system, but since it isnt made by valve and works on imperfect input data its ratings are not 100% accurate. Trying to figure out how much less then 100% accurate it is sounds like a horribly boring discussion that doesnt really have anywhere to go. | ||
|
WindWolf
Sweden11767 Posts
| ||
|
writer22816
United States5775 Posts
| ||
|
hariooo
Canada2830 Posts
On December 06 2013 04:54 devilesk wrote: Are my complaints about DSR invalid? No one seems to actually want to seriously discuss DSR. You mean this? On December 06 2013 04:03 devilesk wrote: If all his ML algorithm is doing is trying to classify matches as pro or not pro and the ratings are just derived from uncertainty because it can't perfectly say whether a match is pro or not, then how is what DSR is measuring any different than Normal, High, and Very High brackets? If anything, it seems strictly worse, because rather than partitioning games into 3 brackets, it partitions them into just two. I haven't read anything anywhere about there being 2 brackets split by pro or not pro so I don't even know what that is about. The site is down so I can't even look for what you're talking about. | ||
|
Plansix
United States60190 Posts
Also making up numbers is fun! | ||
|
Klonere
Ireland4123 Posts
On December 06 2013 05:14 writer22816 wrote: the only thing good about DSR is that it lets you laugh at your friends visas are the worst thing in esports if only we had backing from respectable companies like riot and tencent | ||
|
Excludos
Norway8279 Posts
On December 06 2013 05:23 Plansix wrote: I know for a fact that the DRS system is 56% less accurate than the system Valve would make. At best they can get to 78% accurate if they try, but that is the best we can hope for. Also making up numbers is fun! I've been running the numbers, and there is a 75% chance that DSR is 47.7% more useless than what Valve could have made which is 82.4% correct in solo que in the event where you have a 36% chance to win with your best hero if the other people on your team each have 25% or more. | ||
|
Sn0_Man
Tebellong44238 Posts
You should probably just believe me because I'm telling you this. When I completely rewrite the algorithm tomorrow to change your rating from 200 to 890 you should just assume that I know exactly what I'm doing and I'm making it more accurate. The next day, when I decide your rating is ACTUALLY 72, thats still fine. Just keep believing me. After all, I can predict the outcome of matches with 98% accuracy. Anybody who disputes me obviously understands nothing about statistics, you should go insult them on online forums. After all, I have mathematical proof I'm right. Or at least, I say I do and obviously I've never lied to random people on the internet, that would be silly. | ||
|
devilesk
United States140 Posts
On December 06 2013 04:59 PassiveAce wrote: if you want to discuss dsr just reread the last 20 pages. There really isnt that much to talk about. Its a rating system, but since it isnt made by valve and works on imperfect input data its ratings are not 100% accurate. Trying to figure out how much less then 100% accurate it is sounds like a horribly boring discussion that doesnt really have anywhere to go. I don't want to just discuss how rating systems suck in general or how this is inferior to Elo or how the data is imperfect. I want to discuss his actual methods and what his methods are actually measuring, because he recently just gave us a little more explanation about what he does. From my understanding, his model tries to classify matches as either pro or not pro. What is determined to be pro is set by whatever he trains his machine learning algorithm with (something like a set of pro games). Once he has determined a baseline for what is considered a pro game, he can then go through matches and say whether a game is pro bracket or not. Since it's not a perfect system, he can only give a probability for how confident it is in classifying whether a match is pro or not. And from these probabilities, DSR is derived. My criticism is that it's still ultimately just classifying matches as either in bracket A (pro) or bracket B. It's just two brackets. That's the metric that his model is trying to approximate, and if his model could perfectly say whether a match is pro or not, then DSR wouldn't even exist. With a perfect model, DSR would be eliminated and you'd just be left with what % of your matches are in a pro bracket. To me, a metric that tells you what % of your matches are in a pro bracket seems the same as if I told you what % of your games are played in very high bracket. So if just knowing what % of your matches are in pro bracket isn't good, why would DSR, an approximation of that, be any better? | ||
|
Comeh
United States18919 Posts
On December 06 2013 05:25 Sn0_Man wrote: Its OK harioo, I wrote a quick algorithm that has 98% chance of correctly predicting the outcome of a dota match. You should probably just believe me because I'm telling you this. When I completely rewrite the algorithm tomorrow to change your rating from 200 to 890 you should just assume that I know exactly what I'm doing and I'm making it more accurate. The next day, when I decide your rating is ACTUALLY 72, thats still fine. Just keep believing me. After all, I can predict the outcome of matches with 98% accuracy. Anybody who disputes me obviously understands nothing about statistics, you should go insult them on online forums. After all, I have mathematical proof I'm right. Or at least, I say I do and obviously I've never lied to random people on the internet, that would be silly. Sample size of like 30 games. When there is like 50 variables you need to control for. Seems legit. | ||
|
crms
United States11933 Posts
http://i.imgur.com/VzbcciG.png Trolling friends about DSR has been more fun this week than actually playing DOTA. | ||
|
-Celestial-
United Kingdom3867 Posts
Please, carry on.  | ||
|
devilesk
United States140 Posts
On December 06 2013 05:15 hariooo wrote: You mean this? I haven't read anything anywhere about there being 2 brackets split by pro or not pro so I don't even know what that is about. The site is down so I can't even look for what you're talking about. I got all that from reading his reddit posts. http://www.reddit.com/r/DotA2/comments/1s4ezx/mydotaskillcom_and_why_it_should_be_ignored/cdtubbc?context=3 And in another post he pretty much told me to look up "SVM" which I did on wikipedia, http://en.wikipedia.org/wiki/Support_vector_machine Others who seem more knowledgeable on the subject also mention things like that. If I had to draw a comparison DSR uses the classification of matches as pro or not pro to create a rating similar to how Elo uses W/L to create a rating. However, my criticism is that DSR isn't even based on that, but on the uncertainty in his attempt to classify matches. | ||
|
crms
United States11933 Posts
| ||
|
Klonere
Ireland4123 Posts
On December 06 2013 05:29 crms wrote: I don't care how accurate, inaccurate, bullshit or truthful DSR ends up being, the entire site and all the drama was worth it just for this: http://i.imgur.com/VzbcciG.png Trolling friends about DSR has been more fun this week than actually playing DOTA. its good for this as well | ||
|
Steveling
Greece10806 Posts
| ||
|
hariooo
Canada2830 Posts
On December 06 2013 05:33 Klonere wrote: its good for this as well https://twitter.com/HASHTAGGSL/status/408218149771698176 Bulldog queues pubs with Sheever a lot. This particular example isn't too surprising. | ||
| ||
