EPTMetalopolis prone to close positions, why? - Page 6
| Forum Index > SC2 General |
I apologize to everyone in this thread for taking the OP seriously. My mod senses are definitely off today. -- Chill | ||
motbob
United States12546 Posts
| ||
|
Krissam
Denmark189 Posts
On February 16 2011 04:52 magha wrote: Taking sample sizes of 100 would prove that nobody in history has ever won a lottery. but the lottery has a fuckton more outcomes than metal spawns User was warned for this post | ||
|
Isomer
United States186 Posts
S is the "sample space," or the outcomes you can have. So, that's close-ground, close-air, cross. S={G, A, C} An "event" is an outcome or set of outcomes of a random phenomenon. In this case, it's the 72 close, 17 air, 11 cross. We can call OP's event OP= {CCCC...AAA...GGGG} There are obviously millions of possible events given 100 outcomes and a sample space of 3 equally likely outcomes. To find out how likely getting 72 close positions is, one might simply use the following formula: (total games)! / (number of positions^number of games) (close positions!) (total games-close positions!) This gives us: 100! / 3^100 (72!)(28!) = 9 x 10^-24 So, it's highly unlikely that if these three outcomes are equally probable, then there will be an event with 72 of one outcome. Not to reduce the sort of "legitimacy" of this discussion, but I don't know that we should necessarily put that much stock in the experience of one person. It is unverifiable that OP got 72 - he can claim it, but we all know of the existence of trolls. If we want this to be legitimate, maybe 10-15 other people can load up a few games to see what they get. | ||
|
hypercube
Hungary2735 Posts
On February 16 2011 05:40 TBO wrote: My bad should read have read the thread more careful. However my statement is still partly true, at least if he has not been the only person doing the test. But just the only person posting it because he got this extreme results. And the results isn't too extreme btw, cumulated probability for close spots being 72 when probability for close spot being 1/3 in a sample size of 100 is 0,1066. Can you check your math? I got a z value of >8 (EV=33, SD=4,71), haven't looked up the actual probability but it must be really, really low. Someone posted 1:10^16 which sounds about right. | ||
|
DiaBoLuS
Germany1638 Posts
want to see real proof. | ||
|
philcorp
Canada32 Posts
On February 16 2011 05:43 motbob wrote: The sample size is plenty big enough to show a statistically significant conclusion. I'll run the relevant t-test later. It really isnt. Go and ask a full lecture hall full of first year undergraduates to go and flip a coin 100 times. By the central limit theorem you expect a gaussian distribution about a mean of 50%. If your class is big enough you will get people who flipped a coin 80 or even 100 heads in a row. Funnily enough, if you actually carry out this experiment the students will fake the data and the distribution ends up being much more sharply peaked than it should. User was warned for this post | ||
|
MoonBear
Straight outta Johto18973 Posts
How to replays save positioning data? It'd be really time consuming to try and load up the same map constantly. Would it be possible to data-mine replays? It may simply be the case that the algorithm for allocating spawn positions is skewed towards close positions. Since Blizzard can force no close positions on Shakuras Plateau, it stands to reason that it might be part of the map data itself. I don't know how to use the Map Editor. But perhaps someone could do some investigation to see if there are hard values encoded into the map? | ||
|
fiolek616
Poland2 Posts
31 reps tottal 7 close air 10 close ground 14 cross | ||
|
Keitzer
United States2509 Posts
or try to make a computer program that picks positions (which i think is impossible since you need blizzard's code, which is what we're questioning) | ||
|
BoxedLunch
United States387 Posts
edit: my results 3 cross postion 4 close ground 3 close air | ||
|
aristarchus
United States652 Posts
On February 16 2011 05:47 hypercube wrote: Can you check your math? I got a z value of >8 (EV=33, SD=4,71), haven't looked up the actual probability but it must be really, really low. Someone posted 1:10^19 which sounds about right. The calculator at http://stattrek.com/Tables/Binomial.aspx gave me a p value of about 10^-15. (That's the probability of getting 72 or more close positions on 100 tries, and I believe it's done with precise binomial distribution calculations rather than approximations. Definitely the number should be some tiny decimal of that sort. | ||
|
FeyFey
Germany10114 Posts
I played random a while and got 70% terran. made a break and 2 weeks later i got 70% of the time zerg. (custom games not ladder if you want to argue bnet would set the random races, could be the start posis as well !) You should have also noticed that you get some maps alot on some days. (has other issues but randomizes plays the biggest role imo) Have a few good examples about computer randomness. Wesnoth: 70% hit chance for me 40% hit chance for the opponent. (was abusing save and load hehehe ) did about 20 save and loads until i hit more often then the opponent. Ragnarok Online: upgrading items, there was even a rumor a specific class would have better success, but that was only due to some people testing if the randomizer for the success chance was above or below normal, with cheap npc items. If it was favorable they upgraded the good stuff and made lot of profit that way. (guess the person with the comic wanted to show that, which was a good job !) anyway get down your calc and randomize it from including 1 till 4. You could end up with 70% on 1 number after 10k tryes. Dependend on how your randomizers mood is today. | ||
|
Seronei
Sweden991 Posts
| ||
|
Deadeight
United Kingdom1629 Posts
On February 16 2011 05:42 sob3k wrote: We just need to talk to the SC2GEARS dude (Dakotafanning?) and ask him if its possible to whip up a position analyzer bit in the next release. Compared to what he's already done it should be simple and once run on even one regular players replay library you should have solved the issue. If someone who was good at this stuff could make something like that it would make such a difference. We could get through a huge number of games. P.S. To people testing vs A.I., is it quicker to scout with one of your spawning workers to check or quicker to instantly surrender and load the replay? I've stuck to scouting so far (can read TL whilst it travels and stuff). | ||
|
hypercube
Hungary2735 Posts
On February 16 2011 05:48 philcorp wrote: It really isnt. Go and ask a full lecture hall full of first year undergraduates to go and flip a coin 100 times. By the central limit theorem you expect a gaussian distribution about a mean of 50%. If your class is big enough you will get people who flipped a coin 80 or even 100 heads in a row. Funnily enough, if you actually carry out this experiment the students will fake the data and the distribution ends up being much more sharply peaked than it should. FFS, just run any statistical test, instead of relying on your (flawed) intuition. If you took statistics in college or university you have the tools, so just use them. | ||
|
Chriamon
United States886 Posts
On February 16 2011 05:00 eLiE wrote: If my research class taught me anything, it's that you need a significantly high value to assume that you're getting close positions based on something other than chance. The book said 90% and up, and it's common to go as high as 95%. I'll give the book example for clarity. If you flip a coin and it lands on one side 9 times out of 10, you can assume that the coin is likely rigged. Any less, and it's more likely that the coin landed the way it did due to chance. EDIT: frozenserpent beat me to it, but a higher sample size always improves generalizability What if I flip a coin once and it lands heads, That is 100%, much higher than your 90% requirement. This arguement is rediculous of course, but it just shows that your arbitrary '90%' requirement is completely pointless. What if blizz programmed it to be 70% chance for close spawns? Thats less than 90%, but it is obviously not "due to chance" EDIT: Also, to be on topic, wouldn't there be something in the map file itself were the map actually skewed towards close spawns? perhaps someone should open the file in the map editor and check it out. | ||
|
Lobo2me
Norway1213 Posts
P(X=x)=(1/3)^X * (2/3)^(100-X) * (100 X) where (100 X) is the binomial coefficient. That's the chance of getting X amount of close spawn, 100-X amount of non close spawns and 100 choose X permutations. Do a sum formula for that from 72 to 100 and I got 3,213 * 10^15, which is about the same chance of happening as getting head 48 times in a row. | ||
|
Ihle
Norway36 Posts
On February 16 2011 05:48 philcorp wrote: It really isnt. Go and ask a full lecture hall full of first year undergraduates to go and flip a coin 100 times. By the central limit theorem you expect a gaussian distribution about a mean of 50%. If your class is big enough you will get people who flipped a coin 80 or even 100 heads in a row. Funnily enough, if you actually carry out this experiment the students will fake the data and the distribution ends up being much more sharply peaked than it should. Do you really mean that you have met people who have gotten 80 or 100 heads in a row, then you are very gullible, the likelihood that this will happen is easy to calculate: 0.5^80 = 8.27*10^-25, if a friend of you did that he is one lucky guy! People suck at statistics. | ||
|
Zelniq
United States7166 Posts
his most recent match history (http://eu.battle.net/sc2/en/profile/717496/1/HUARGH/matches) doesn't show any metalopolis games. sure seems he's full of it. also I really doubt someone actually hosted a game, waited 10 seconds for it to start, scouted close positions (sent OL to close-air), or instead left and loaded replay, rinse and repeat 100 times. edit: someone pointed out that singleplayer games wouldn't show up in match history. true. but replays would autosave, if he wanted to prove he could post a rep pack. another person points out that if he put himself as an obs watching 2 ai's on metal then it would be very quick to see spawn locations, but then it could just be a bug with how the computer generates 2 AI spawns, not related to how players spawn | ||
|
space_yes
United States548 Posts
On February 16 2011 05:48 MoonBear wrote: I did a simple Binary Test on the data p00n provided. I considered the probability that close positions are more likely against the probability of close air and cross-map. Even allowing for a probability of such an event occurring as high as 0.65, your p value (0.038) would still suggest rejection of the hypothesis for a one-tail test and instead suggest that the true probability is much higher at the α=5% level. A probability of 0.7 gives a p value of 0.11 and so seems much more likely. How to replays save positioning data? It'd be really time consuming to try and load up the same map constantly. Would it be possible to data-mine replays? It may simply be the case that the algorithm for allocating spawn positions is skewed towards close positions. Since Blizzard can force no close positions on Shakuras Plateau, it stands to reason that it might be part of the map data itself. I don't know how to use the Map Editor. But perhaps someone could do some investigation to see if there are hard values encoded into the map? Yes. You could potentially check thousands of games this way. | ||
| ||
