EPTMetalopolis prone to close positions, why? - Page 9
| Forum Index > SC2 General |
I apologize to everyone in this thread for taking the OP seriously. My mod senses are definitely off today. -- Chill | ||
Chill
Calgary25993 Posts
| ||
|
Myia
173 Posts
 | ||
|
dcberkeley
Canada844 Posts
On February 16 2011 06:28 Ludwigvan wrote: I dont really know how it works, but if first the position of player 1 is created and then player 2, we have 12 possibilities and not 6. I was going to say something to this effect but it doesn't really change the probability either way. You still have 1/3 chance of spawning at the various positions. | ||
|
nkr
Sweden5451 Posts
On February 16 2011 06:34 Chill wrote: Please notice the giant banner at the top... Let's assume there's a possibility that it's true. How do we test it going forward? Every person in the thread does 30 games each, posts it here, and some hard working soul puts them together for a bigger sample size. Maybe. | ||
|
TheGrimace
United States929 Posts
-My Initial Method- So there are a few ways we can go about this. If the OP can provide 100 replays, we can confirm his results. Using sc2gears, you can see when the game was played and where both players spawned. That would likely be the quickest way to verify. -or- If people want to run another test, you can have people play 10 games and autoquit. Every replay should be within ~1 minute of each other and the game time would only be a couple of seconds each. Then those 10 replays could be sent to someone who could quickly check in sc2gears. How to check in sc2gears: 1. Open sc2gears, go to the replay analyzer and set chart type to "map view" 2. Drag and drop each replay 3. Note the spawn positions 4. Repeat The concern would be people cherry picking replays. We could have each player play a game on Incineration Zone before and after their 10 game sample. Then each submitting player would also submit their bnet profile link, and it would be very easy to take a quick look to see that only 10 games were played between the Incineration Zone games. This would mean that the 12 file created times would be sequential. I picked Incineration Zone since it's pretty unlikely anyone is playing that map normally. If we were using the Incineration Zone bookends, each player would submit 12 replays in a zip file. Any thoughts on this method? | ||
|
KillerDucky
United States498 Posts
On February 16 2011 06:34 Chill wrote: Please notice the giant banner at the top... Let's assume there's a possibility that it's true. How do we test it going forward? I propose the OP makes a prop bet with people in this thread. Serious suggestion. Money on the line will get things in focus quick. | ||
|
Logros
Netherlands9913 Posts
| ||
|
Chill
Calgary25993 Posts
On February 16 2011 06:40 TheGrimace wrote: I wrote this while updating my sc2gears, but once I got it loaded and checked a few replays, I realized it wouldn't work properly. Is there a MASSIVE replay pack anywhere? Like 1000+ games on Metalopolis we can feed into SC2Gears? Alternatively, is there any way to batch download from any popular replay upload sites? | ||
|
RoKetha
United States211 Posts
On February 16 2011 06:21 Annq wrote: Matches against AI (singleplayer) arent listed in the match history? Atleast i think so If you look at his map stats page, which you can clearly see includes custom games (and I think does not include ladder games), he's only played on Metalopolis 86 times, and the last time he played on it was January 10. It's hard to make sense of exactly where that data is being sampled from and if it won't count certain types of games, but at first glance it certainly looks like he didn't play 100 games there. I don't actually know where Zelniq got the account name from though (possibly a PM?) because I didn't see it myself in the thread. Edit: Sorry, I realized just moments after posting that you can play matches offline (as opposed to online vs AI) and that's what the quoted post meant. Those probably wouldn't show up. | ||
|
archangel967
Canada111 Posts
:D | ||
|
Ludwigvan
Germany2371 Posts
On February 16 2011 06:51 RoKetha wrote: If you look at his map stats page, which you can clearly see includes custom games (and I think does not include ladder games), he's only played on Metalopolis 86 times, and the last time he played on it was January 10. It's hard to make sense of exactly where that data is being sampled from and if it won't count certain types of games, but at first glance it certainly looks like he didn't play 100 games there. I don't actually know where Zelniq got the account name from though (possibly a PM?) because I didn't see it myself in the thread. He checked older posts of the OP to find out his Battle net name. Also it has already been said that he could have played single player games against the computer wicht dont appear on the match history. | ||
|
gerundium
Netherlands786 Posts
On February 16 2011 06:51 Chill wrote: Is there a MASSIVE replay pack anywhere? Like 1000+ games on Metalopolis we can feed into SC2Gears? Alternatively, is there any way to batch download from any popular replay upload sites? Be wary of replay packs. people select replays to go into those, which would make the test useless (for example i'd expect more frequent cross in packs due to higher potential for epic long games). We would need to find a replay pack that is completely unbiased, plain not going to happen. which means manually loading up an x number of games in a row IS the way to go here to test it. Upload sites do not work because they suffer from the same selection bias as a replay pack. There really is no other way for this to work otherwise that i can see. | ||
|
Chill
Calgary25993 Posts
On February 16 2011 06:55 gerundium wrote: Be wary of replay packs. people select replays to go into those, which would make the test useless (for example i'd expect more frequent cross in packs due to higher potential for epic long games). We would need to find a replay pack that is completely unbiased, which would mean manually loading up an x number of games in a row. Upload sites do not work because they suffer from the same selection bias as a replay pack. There really is no other way for this to work otherwise that i can see. Right, which is why I'm suggesting a tournament replay pack, such as the TLOpen or similar. | ||
|
deejY
Germany44 Posts
On February 16 2011 06:51 Chill wrote: Alternatively, is there any way to batch download from any popular replay upload sites? as close position games on meta tend to be really lame/short, and most replay sites are feeded with games that are considered worth watching, you won’t be able to conclude anything out of this | ||
|
gerundium
Netherlands786 Posts
On February 16 2011 06:56 Chill wrote: Right, which is why I'm suggesting a tournament replay pack, such as the TLOpen or similar. That could work indeed, there is no bias in those. Contact the ICCup people? They are still running the MLG sundays casting all the MLG replays. that is a crazy amount of replays right there from all those events, see if they can send that set over to you? | ||
|
Dimagus
United States1004 Posts
There is also the issue of non-random randomness. It is not simply enough that number of occurrences satisfy a statistical variability if patterns and streaks occur. Example: repeating the pattern 123412341234... 100 times will yield a perfect 25% for each of the 4 numbers, but it isn't random. Large patterns are easy to spot but small patterns and linked numbers may not be. Assume for metalopolis 1 is 9 o'clock, 2 is 12 o'clock, 3 is 6 o'clock, 4 is 3 o'clock. 1-2 and 3-4 correspond to close spots, 1-3 and 2-4 are close air, 1-4 and 2-3 are cross positions. The problem is that if 1-2 and/or 3-4 are linked then close spots will occur more frequently. Linked means that if a 1 shows up the next number is more likely to be 2 and vice versa. 72% (33% + 39%) with a 100 game size is unacceptable, so if the mods can curb the people going "come back with 10000" since they have nothing constructive to offer. User was warned for this post | ||
|
Centorian
United States95 Posts
On February 16 2011 07:02 Dimagus wrote: I will state this up front since some people think it isn't true. A sample size of 100 is absolutely, positively and definitely a statistically relevant sample size for Starcraft 2. We are not talking about 100 mathematical calculations that occur within a single minute, we are talking about 100 games of starcraft 2 which occur over days and weeks. The allowable statistical variance must apply to samples as low as 100, and not just 10,000 or million. If it doesn't then there is a flaw in the algorithm. There is also the issue of non-random randomness. It is not simply enough that number of occurrences satisfy a statistical variability if patterns and streaks occur. Example: repeating the pattern 123412341234... 100 times will yield a perfect 25% for each of the 4 numbers, but it isn't random. Large patterns are easy to spot but small patterns and linked numbers may not be. Assume for metalopolis 1 is 9 o'clock, 2 is 12 o'clock, 3 is 6 o'clock, 4 is 3 o'clock. 1-2 and 3-4 correspond to close spots, 1-3 and 2-4 are close air, 1-4 and 2-3 are cross positions. The problem is that if 1-2 and/or 3-4 are linked then close spots will occur more frequently. Linked means that if a 1 shows up the next number is more likely to be 2 and vice versa. 72% (33% + 39%) with a 100 game size is unacceptable, so if the mods can curb the people going "come back with 10000" since they have nothing constructive to offer. *facepalm* we've moved on. | ||
|
Myia
173 Posts
Player vc Computer on Metal Close by Ground - 13 Close by Air - 19 Cross Positions - 18 Certainly more variance in the positions. But mind, this was a sample half as big as the OP's supposed size... But shows what I think, that it is indeed totally random. Perhaps the OP, if he did the test, was just very very unlucky... Edit: Just to point out, I did these all on Multiplayer, so replays are available, as are stats on whatever page you want to check them out on, if anyone is interested. | ||
|
gerundium
Netherlands786 Posts
On February 16 2011 07:04 Myia wrote: Ok, so, I have just done 50 (not 100 as of yet, as wondering whether I should continue after my results so far, which are little.. well, different from the OP's...) and have got the following results. Player vc Computer on Metal Close by Ground - 13 Close by Air - 19 Cross Positions - 18 Certainly more variance in the positions. But mind, this was a sample half as big as the OP's supposed size... But shows what I think, that it is indeed totally random. Perhaps the OP, if he did the test, was just very very unlucky... He could definitely have gotten the 1:10000 shot, it its at least way more likely that we are reading about a 1000 to 1 shot on here, because posting about something like this becomes more likely as someone tests this and comes up with a statistical outlier. | ||
|
aristarchus
United States652 Posts
On February 16 2011 06:34 Chill wrote: Please notice the giant banner at the top... Let's assume there's a possibility that it's true. How do we test it going forward? If you're worried about the difference between human and computer players and so on, I'd say you needed to use a replay pack, but it needs to be one that included all games that happened - a tournament pack would be fine - not one where someone picks out their most interesting games. sc2gears should help a lot, but I don't know of a fully automated way to do it. What's really important is that it's a single decided on test where the results will be reported no matter what they are. Saying "everyone do 10 and then report" won't work, because people could give up after 5 if they're getting boring results, etc. - no way to guarantee what you're seeing is representative of the tests people actually did. You don't need 1000 games. If the bias is anywhere near the 72% that supposedly happened in this test, 100 games is more than enough to find it. (And honestly, I'm really skeptical - if >70% of metal games were close positions, people would definitely have noticed. 70% is far enough from 1/3 to be very obvious to someone casually watching youtube videos, and definitely obvious to a pro who is carefully practicing the map.) | ||
| ||
