|
A Case Study of ZvP Macro on Shakuras Plateau Across Leagues
Abstract/Introduction
+ Show Spoiler +The purpose of this study is to examine, in-depth, the differences in macro across leagues (only in the early game). Specifically, in the initial data, it is to investigate how the rate of worker production correlates with player skill. Read: rate, not final worker count. It is known that a master player is better than a diamond player. I'm not trying to prove that one league is better than another, because that is redundant and pointless. There is another thread here http://www.teamliquid.net/forum/viewmessage.php?topic_id=266019 that goes into a spending quotient and compares macro across leagues at an extremely zoomed out level. It's very cool stuff, but unfortunately Starcraft 2 is a game where the possible scenarios grow to exponential proportions, similar to chess. In the ETA-Concept be Felo, he quotes: "Chess is infinite: There are 400 different positions after each player makes one move apiece. There are 72,084 positions after two moves apiece. There are 9+ million positions after three moves apiece. There are 288+ billion different possible positions after four moves apiece. There are more 40-move games on Level-1 than the number of electrons in our universe. There are more game-trees of Chess than the number of galaxies (100+ billion), and more openings, defences, gambits, etc. than the number of quarks in our universe!" - Chesmayne How many variations can there be in terms of worker production in Starcraft 2 by the 10 minute mark? The average deviation in whatthefat's analysis is astronomical. He also only looked at the end result of the games. I'm interested in something different. I wanted to investigate macro in terms of a the progression in the first 10 minutes of a Starcraft 2 game. I feel it's interesting to look at zerg instead of terran and protoss due to the larva mechanics. In terran and protoss, constant worker production = best possible macro (in terms of worker production), which isn't that hard to do (keep tapping cc/nexus hotkey). However, as zerg, you have to consciously decide when the drone, how many to make, etc. As of the intial posting, I've only invested maybe two hours into this. This will be an ongoing project where I will be adding data. The topics investigated are: Workers built/minute Unspent resources/minute Food/minute - Oddly enough, it appears that looking at food is useless, since the difference is too little. Every single player (so that's 19 players) across all the leagues had food ranging from 60-80 at the 10 minute mark. Aside from Stephano. The higher level players had better tech.
Methods
+ Show Spoiler +I needed to set parameters for the games that I will be investigating. Since matchups and the maps they take place in require so much variation in play style, I settled on ZvP on Shakuras Plateau. Investigating all games from the zerg perspective, I chose this map because it's one of the easier maps for zerg to get a third (and thus encourages more macro). In selecting the games, I made sure that the zerg was not allining by checking the build order and making sure no roaches or exessive amount of zerglings were built in the first 6 minutes of the game. I also checked to see what the protoss was doing. Ideally, I would have a fast 3 hatch vs FFE, but matches like those are virtually impossible to find on replay websites in leagues lower than diamond. When looking for games from bronze to plat, I attempted to find games where the zerg player tried to macro to the best of his ability.
In summation, in each game:
- The zerg didn't build much army and had no tech for the first 6 minutes. - The zerg expanded quickly (relatively to what's expected for their league). - A majority of the zerg player's food was drones for the first 9 minutes. - The zerg did not face early game aggression, such as 2gate proxy zealot or 4gates. - All replays are from the NA server (except for Stephano).
For the workers built per minute, I looked at 3 games from each league. I know 3 is a small sample size, but considering how narrow my parameters are, I felt it was acceptable. I counted the workers every 30 seconds, averaged them, and compiled the data into a graph.
In order to find the unspent resources per minute, I will use the same approach by averaging unspent resources every 30 seconds and fit a linear regression to the data. I will take the gas and multiply it by 1.25 (since 4 gas is mined instead of 5 minerals), and add that to the mineral count. Therefore, the graph will only display unspent minerals, but will factor in stockpiled gas. It's not perfect, but call me lazy.
Sources of Error
There are many sources of error. I will list them out here:
- Gas timings. People who opt for speed earlier will naturally have less drones.
- Protoss aggression. Luckily, in almost all the games the players were untouched for the first 8 minutes (even in master). There was some stargate and zealot aggression, but none of the games consisted of 5-8 gate allins or double stargate allins.
- Diversity of skill within a league. Expanding upon this, it could be possible that I happened upon 3 platinum ZvP's on Shakuras Plateau that didn't involve early aggression, where all three zergs sucked about equally as much (check out how close the data points are to the linear fit). There will most likely also be the most deviation in bronze and master, with the least in silver/gold/plat.
- Aggressive decision to stop droning after a certain point (mostly in diamond and masters).
Results and Discussion
Part 1
+ Show Spoiler +This will be updated as I get more data. Here is the first set of interesting stuff: Workers alive over time The first thing you guys notice after inspection is probably that platinum players have the lowest drone production of all the leagues. This looks like I screwed up bad, but upon reflection and closer study of the replays, it makes sense. The platinum replays were by far the most consistent. I can upload the replays if you guys don't believe me. Here is what I believe the progression of Bronze -> Diamond is. You do 1/2 base allins until plat, then you start learning how to play. The people who struggle to "macro better" in bronze-gold are floundering around, not knowing when to drone, when to build army, etc. In platinum, they are constantly scared of aggression and under drone. Master and diamond players have similar worker production. However, in the replays, it is clear that master players are more clever and have better mechanics. Therefore, it appears that the greatest difference between masters and diamond isn't macro, but in fact multitasking ability and game understanding. For master and diamond players, I believe it when you say you have 60+ workers at 10:30. Esp since you can build them in bulk quickly. But this is only to the 10 min mark. Similarly speaking, the difference between bronze and gold isn't so much worker production, but from what I can see, spending the money and having a somewhat reasonable build order. A large fraction of the players I studied in these leagues built many workers blindly without scouting what the opponent was doing, probably after having people on teamliquid telling them to macro better, or Day9 telling them to lose by overdroning. I believe that the drone production outlined in my data above overestimates the average drone production for zergs in bronze and silver. It's also likely that bronze/silver/gold players make drones blindly, while platinum players are more cautious. Again, notice the difference here is around 6 drones before the 10 min mark. I'm sure by the end of the game, platinum players generally have more total workers built. Stephano is obviously in a league of his own. In the end though, at least for zerg in ZvP on Shakuras Plateau, it seems the number of workers you can get away with early game is directly proportional to your skill and how gosu you are. The conclusion I came to is that it's not the number of workers built that matters, but the rate at which workers are built, especially in the early game. This most likely only pertains to zerg as terran/toss worker build rate is capped by their cc/nexus count.Feel free to conduct your own analysis of my data. I assure you that I did everything in my power to make the data as accurate and precise as possible despite the small sample size (I painstakingly studied 20+ replays! many were omitted).
Extra
If anybody would like to help me out on my quest for knowledge, I would really appreciate it if you posted replays of yourself playing a macro-focused game of PvT on Antga Shipyard! Obviously leave your league too. =)
Also, yes, the sample size is small, so the results here aren't law. I'm honest that I only have 3 replays per league. If you guys are curious about what I'm talking about here, please then submit ZvP replays on Shakuras Plateau that are macro-oriented, and I will try to compile more holistic data. The limiting factor here is that it's hard to find replays within the constraints on replay websites, especially in lower leagues. Ideally I could have maybe 10 games for each league.
This is very much a work in progress. Also, the results of this isn't "plat players suck", if this is what any of you were walking away with.
|
This is very interesting, thanks for your work!
Out of curiosity, why Antiga? That map is... interesting for pvt.
edit: oh I know! It's because protoss never gets a fourth!
|
as much as your results tell us things we already know...
~3 per league is not a good enough sample size
|
Stephano having 40~ drones in between 6 - 7 minutes. That is the mark of a pro.
|
On March 31 2012 15:10 sam!zdat wrote: This is very interesting, thanks for your work!
Out of curiosity, why Antiga? That map is... interesting for pvt.
edit: oh I know! It's because protoss never gets a fourth! its shakuras, and its zvp
|
You're going to have to explain yourself how a master player has 45 drones at the 10 minute mark. I am a midhigh master zerg and at this point assuming 4gate pressure into third and I don't lose my third hatch, I would have 90 drones at this point then sinking 15 into spines. I can beat a 3hatching zerg with a 10 minute 1 base 4 gate push if they had 45 drones.
I have had games where my drone production was 20% higher than stephanos on that graph when a shitty toss goes for a fast third base too. These stats don't take into account much.
|
good job, every zerg will be droning like crazy now trying to copy stephano and die to any pressure/allin :D
|
10 min 20 drones on an average for platinum? really? 2/3rd of even bronze? I dont think Platinum players are that scared of droning. Maybe the replays of platinum have some special cases. It is hard to believe that a platinum player would be scared to make more than 20 drones.
|
Hold on a minute...
What exactly is your data set, and where'd you get it from?
You apparently selected games to filter out games that didn't involve "macro" but your criteria for filtering seems to be up to your discretion, guided by some principles you outline (although they aren't very firm)
How many games did you review?
How'd you review across leauges?
The thing that gives me the most pause is your source of "error" is "diversity across leauge". I thought that's exactly what you were trying to measure. That's not a source of error, that's the signal you're trying to distinguish from the noise.
I think this kind of study is useful, but I'm less than confident in your statistical or methodological skills. Maybe if you could give a useful explanation of your method it'd go some ways towards alleviating that fear.
Update: None of these concerns btw begin to address systemic issues in your design... limiting to one map, to one matchup, etc. But at least you're up front with those parts. I'm more concerned about the parts we don't know.
|
You need 20 minimum for each league to have an acceptable sample size
|
On March 31 2012 15:55 statikg wrote: You need 20 minimum for each league to have an acceptable sample size
That statement makes no sense unless you have broader context about what's being measured. That information's really not present here. So while I appreciate you're trying to increase the robustness of the statistics, that statement kind of points out your unfamiliarity with basic statistics.
|
On March 31 2012 15:55 statikg wrote: You need 20 minimum for each league to have an acceptable sample size Not really as long as the data pans out and is as thorough as possible.
|
On March 31 2012 15:52 celeryman wrote: Hold on a minute...
What exactly is your data set, and where'd you get it from?
You apparently selected games to filter out games that didn't involve "macro" but your criteria for filtering seems to be up to your discretion, guided by some principles you outline (although they aren't very firm)
How many games did you review?
How'd you review across leauges?
The thing that gives me the most pause is your source of "error" is "diversity across leauge". I thought that's exactly what you were trying to measure. That's not a source of error, that's the signal you're trying to distinguish from the noise.
I think this kind of study is useful, but I'm less than confident in your statistical or methodological skills. Maybe if you could give a useful explanation of your method it'd go some ways towards alleviating that fear.
Update: None of these concerns btw begin to address systemic issues in your design... limiting to one map, to one matchup, etc. But at least you're up front with those parts. I'm more concerned about the parts we don't know.
Did you just skip everything I typed and looked at the pretty picture?
I filtered out games that didn't involve early pressure and instead encouraged and allowed for macro.
How'd I review across leagues? I averaged the worker count every 30 seconds for all players within a league.
My source of error is diversity within a league.
The reason for such constricting parameters (one matchup, one map, no early game aggression, majority drones, etc.) is to make the data as accurate and consistent as possible.
I'm creating a manageable model by setting variables as constants. Think partial derivatives.
I'm not confident in your analytic or reading skills.
On March 31 2012 15:22 Jombozeus wrote:I have had games where my drone production was 20% higher than stephanos on that graph when a shitty toss goes for a fast third base too. These stats don't take into account much.
Your macro is better than Stephano's? Cool story bro. 20% better? You get 72 drones by 8 minutes?
On March 31 2012 15:10 sam!zdat wrote: This is very interesting, thanks for your work!
Out of curiosity, why Antiga? That map is... interesting for pvt.
edit: oh I know! It's because protoss never gets a fourth!
Only looking into early game macro, so 4th isn't a question.
Granted, one of the masters players went for fast tech after hitting 40-ish drones and dropped roaches into toss' main. It worked splendidly. So perhaps the master players stopped droning on purpose after a certain point. But this is a study on early game macro. Focus more on the slope of the linear regression and less on the Y-intercept I believe if Stephano was doing the same roach drop, he'd do it with more drones.
|
Ah, I missed that you were only looking at first 10 minutes.
Any reason you want to look at zvp on shakuras and pvt on antiga? Those are matchups that get whined about on those maps, dunno if you have considered whether than might make a difference in the findings.
@johnny read the bottom of op
|
Here... maybe this is more straightforward. Answer these questions please, because they're necessary to evaluating a study like this, not because i'm trying to be confrontational with you. (be slightly less sarcastic too)
1) Data source: How did you find this data? Replays? If so where from. Did you look at sc2 gears data? Did you look at score screens? Build listings from games?
2) How many games in your data set? How are those divided between leagues?
3) "I averaged the worker count every 30 seconds for all players within a league." Does this mean you summed all the workers at each 30 second increment, in every game you reviewed, then divided by the number of games, for each leauge?
Again "diversity", as you call it, aka randomness, is not an error. It's something that random sampling will correct for. "Error" is either your native error from statistical sources (which is why i ask about sample size), or it's systemic, from things like using 1 matchup, or 1 map, or pulling all your games from games people chose to upload, etc.
Also I don't think partial derivatives are the analogy you should be going for.
Restricting your data to "macro" games is fine, but I'd be more comfortable if your criteria was strictly designated, not partially based on your gut feel about whether the game was "macro" oriented or not.
Hopefully others with some math/science background will comment here and point out that these are not extreme demands of an empirical study.
I'm pretty sure my analytical, and reading skills are at or above the grammar skills in your first sentence too.
|
I would love to go over 3, but you can't believe how hard it is to find replays of ZvP on Shakuras Plateau in each league where there are no early-game allins or fast aggression from either player (like, FFE -> 5gate robo and Z doing a fast 3rd).
|
On March 31 2012 16:35 celeryman wrote: Here... maybe this is more straighforward. Answer these questions please, because they're necessary to evaluating a study like this, not because i'm trying to be confrontational with you. (be slightly less sarcastic too)
1) Data soruce: How did you find this data? Replays? If so where from. Did you look at sc2 gears data? Did you look at score screens? Build listings from games?
2) How many games in your data set? How are those divided between leauges?
3) "I averaged the worker count every 30 seconds for all players within a league." Does this mean you summed all the workers at each 30 second increment, in every game you reviewed, then divided by the number of games, for each leauge?
Again "diversity", as you call it, aka randomness, is not an error. It's something that random sampling will correct for. "Error" is either your native error from statistical sources (which is why i ask abour sample size), or it's systemic, from thigns like using 1 matchup, or 1 map, or pulling all your games from games people chose to upload, etc.
Also I don't think partial derivatives are the analogy you should be going for.
Restricting your data to "macro" games is fine, but I'd be more comfortable if your criteria was strictly designated, not partially based on your gut feel about whether the game was "macro" oriented or not.
Hopefully others with some math/science background will comment here and point out that these are not extreme demands of an empirical study.
I'm pretty sure my analytical, and reading skills are at or above the grammar skills in your first sentence too.
I'm pretty sure my analytical and reading skills are at or above the level of grammar in your first sentence.
1.) I got the replays from sc2replayed.com. I browsed through countless replays after I set the search parameters as (league), 1v1, ZvP, Shakuras Plateau. I got the data by going into the game, and writing down the number of workers there are every 30 seconds.
2.) A total of 24 games, 5 of which were omitted, so 19 games across the leagues.
3.) Yes.
4.) By error I mean deviation, like standard error/average deviation.
Again, it is incredibly hard to find adequate replays that fit within my control parameters. Especially below diamond.
Odd that you compare analytical and reading skills to grammar. Great scientific method there.
How should my criteria be more "strictly designated"?
|
what level master players did the have that they averaged 35ish workers at the 8 minute mark, if we're not talking about a zealot pressure or something that forces spines and wastes larva, then i find it hard to believe that any competent masters zerg doesn't have 55-60 drones at 8 minutes
|
It's nice that you took inititive on this, but I'm not really sure what you're trying to say. There's no real conclusion here, and the data doesn't appear to be immediately useful. Your method, especially sample size (3 games or less per league), is really not sufficient. Also, I think it should be clear that when you say, "A large fraction of the players I studied in these leagues," you mean like 2-3 people per league. That assumes that you didn't use more than one game form a single player.
There also seems to be something seriously wrong with the platinum data. I'm not sure if you just got unlucky when picking games (easy to do when you chose 3 at random) or if you made some sort of honest mistake (mislabeling sounds like a reasonable explanation). I'd need to see a lot more before I believed that platinum players were making about 2/3 as many workers at 10 min as bronze players, and about 1/2 as many as diamond players. In any event, the explanation you provide doesn't sound at all satisfying.
The spending quotient thread you linked to has much better data (read larger sample size and over longer durations of game time) on worker production, and also provides useful analysis on spending. This thread looks redundant at best.
|
Interesting idea, but definitely not even close to a big enough sample size to say anything. There's just no way platinum players on average has that low, even though your theory that platinum is where zergs stop doing 2base allins might be valid to some degree.
Personally I'm platinum and always have more than 50 workers at the 8 minute mark, and I don't think I'm some exception.
|
On March 31 2012 17:17 Omnipresent wrote: It's nice that you took inititive on this, but I'm not really sure what you're trying to say. There's no real conclusion here, and the data doesn't appear to be immediately useful. Your method, especially sample size (3 games or less per league), is really not sufficient. Also, I think it should be clear that when you say, "A large fraction of the players I studied in these leagues," you mean like 2-3 people per league. That assumes that you didn't use more than one game form a single player.
There also seems to be something seriously wrong with the platinum data. I'm not sure if you just got unlucky when picking games (easy to do when you chose 3 at random) or if you made some sort of honest mistake (mislabeling sounds like a reasonable explanation). I'd need to see a lot more before I believed that platinum players were making about 2/3 as many workers at 10 min as bronze players, and about 1/2 as many as diamond players. In any event, the explanation you provide doesn't sound at all satisfying.
The spending quotient thread you linked to has much better data (read larger sample size and over longer durations of game time) on worker production, and also provides useful analysis on spending. This thread looks redundant at best.
It feels like people ignore 90% of what is in the OP. First of all, the purpose of this is to do a detailed comparison where you don't have extra factors and variables. In terms of the spending quotient thread, I wanted to investigate if the ideas in it still held true (which they did) when you eliminate a majority of the average deviation.
Again, this isn't about the spending quotient. It's about the rate of workers produced and how that correlates with a player's league. I can wait until I'm about to die, lose my whole army, then spend all my idle larva and my 2k unspent minerals on drones before I lose. Viola! I have 89 drones at the end of the game. See the problem with not being selective in terms of data?
It's also hard to find games where you have a ZvP on Shakuras Plateau, where neither player does an early game allin or pressure. One of the replays for master players I used, the protoss doesn't have a single unit go to the zerg's half of the map that isn't a probe in the first 9 minutes. PM me if you want proof of this.
This data is only relevant for ZvP on Shakuras Plateau in the event of non-early aggression. I can say with confidence that platinum zergs playing protoss on shakuras plateau who aren't planning on allining will make drones similar to what is on the graph. If you have a replay of something different, please send it to me and I'll change it.
|
Good job! With all the constraints you made 3 isn't perfect but not as bad of a sample size as others accuse you of either. For my own work/analysis you study actually is a good contribution: Yesterday, analysing my own replays, I happend to discover that around plat people quite often tend to cut probes early. Your study seems to confirm this!
[edit]Ok, talking about confirmation is a little too much, but you clearly saw what my replays very often show across maps and races at the upper end of gold/lower end of plat leagues: people cut workers early.[/edit]
So in order to get into/through plat one needs to learn to handle tons of early aggression. At least that's my new theory ...
|
I really like the premise because there are just too many factors that make most “statistical” results simply "generic". I appreciate the effort, just bear in mind 3 maps per league really isn't enough, even though it might take a lot of time to go through.
|
On March 31 2012 17:50 PeanutsNJam wrote:Show nested quote +On March 31 2012 17:17 Omnipresent wrote: It's nice that you took inititive on this, but I'm not really sure what you're trying to say. There's no real conclusion here, and the data doesn't appear to be immediately useful. Your method, especially sample size (3 games or less per league), is really not sufficient. Also, I think it should be clear that when you say, "A large fraction of the players I studied in these leagues," you mean like 2-3 people per league. That assumes that you didn't use more than one game form a single player.
There also seems to be something seriously wrong with the platinum data. I'm not sure if you just got unlucky when picking games (easy to do when you chose 3 at random) or if you made some sort of honest mistake (mislabeling sounds like a reasonable explanation). I'd need to see a lot more before I believed that platinum players were making about 2/3 as many workers at 10 min as bronze players, and about 1/2 as many as diamond players. In any event, the explanation you provide doesn't sound at all satisfying.
The spending quotient thread you linked to has much better data (read larger sample size and over longer durations of game time) on worker production, and also provides useful analysis on spending. This thread looks redundant at best. It feels like people ignore 90% of what is in the OP. First of all, the purpose of this is to do a detailed comparison where you don't have extra factors and variables. In terms of the spending quotient thread, I wanted to investigate if the ideas in it still held true (which they did) when you eliminate a majority of the average deviation. Again, this isn't about the spending quotient. It's about the rate of workers produced and how that correlates with a player's league. It's also hard to find games where you have a ZvP on Shakuras Plateau, where neither player does an early game allin or pressure. One of the replays for master players I used, the protoss doesn't have a single unit go to the zerg's half of the map that isn't a probe in the first 9 minutes. PM me if you want proof of this. I have faith in the collective intelligence of the people here, so I'll assume you're just skimming. Please read carefully. I didn't misunderstand any of that. In fact, the only part of this post that I disagree with is where you say your results line up with the results in the SQ thread. They do not. The platinum numbers are wildly inconsistent between your numbers and the numbers in that thread. As far as I can tell, this is the only conclusuion you've drawn from this information, and it's demonstably false.
I understand what you're trying to do: examine data in a specific matchup on a specific map given a specific set of circumstances. You're aming for specificity. You've arbitrarily chosen to look at games where players are essentially left alone to build drones. I just don't know why you did that. It's not useful in any way, and isn't particularly interesting either. Why should someone reading this care about the data? When I say "I'm not sure what you're trying to say," I don't mean that your english is unclear. I mean I read it all, understood it, and was left wondering why it mattered. I quite literally want to know why it matters how many workers are produced in this specific matchup, on this specific map, and under these conditions. What made you think of this? What kind of conclusions do you expect to draw? What point are you making?
That's all taking for granted the premise that your data is somehow representative, which doesn't appear to be the case. If it's not representative, why bother looking at it? I don't blame you for the data. Based on the difficulty of finding enough games that fit your criteria, it probably isn't reasonable to expect your information to be representative. Again, that's true even if you could use the information to draw useful conclusions (which seems unlikely in the first place). So in addition to asking why this matters, I'm asking why it matters so much that you're willing to use such a small set of information to "examine" it. Why is it so important to look at this that you're willing to look at such a small snapshot just to get insight into this situation?
I don't doubt that you used real replays and that those replays complied with the set of conditions you put forward. I don't feel like I need to see the actual games. I essentually trust you. You have no obvious reason to fake this information, and would have likely claimed a larger sample size if you were going to. That being said, 3 games per league isn't enough to indicate anything at all. The fact that your platinum numbers seem so wildly off base is indicitive of that fact. This is based on both personal experience and the drone production numbers, not the spending quotient numbers, in the spending quotient thread. + Show Spoiler +That thread puts platinum worker production numbers much closer to 30, which is higher than your number, not lower. Any of the complicating factors you chose to eliminate would only serve to lower that number. That is, your number should be equal to or greater than that number.
I understand exactly what you're saying. It's not complicated. It just looks like a clumsy attempt to demonstate something that is neither useful nor interesting in the first place. I'm sure it's not entirely your fault. Finding replays to fit these criteria is probably both difficult and time consuming. But why settle for less?
|
If you're struggling to find a 10 minute game with no aggression, perhaps you should simply broaden your search criteria? Even by the 7 minute mark there's already quite a clear deviation in workers produced.
|
What i dont understand is, if you look at the very first point plotted on the table, Stephano has the most drones at 3 mins but bronze has the 2nd best :S are you telling me that a bronze zerg player is going to drone and have his build orders off better than a masters player?
|
On March 31 2012 18:59 Darkcaster wrote: What i dont understand is, if you look at the very first point plotted on the table, Stephano has the most drones at 3 mins but bronze has the 2nd best :S are you telling me that a bronze zerg player is going to drone and have his build orders off better than a masters player?
There's very little that can happen to fuck up your drone count only 3 minutes in.
|
On March 31 2012 19:09 Monkeyballs25 wrote:Show nested quote +On March 31 2012 18:59 Darkcaster wrote: What i dont understand is, if you look at the very first point plotted on the table, Stephano has the most drones at 3 mins but bronze has the 2nd best :S are you telling me that a bronze zerg player is going to drone and have his build orders off better than a masters player? There's very little that can happen to fuck up your drone count only 3 minutes in.
Bronze people are still forgetting to make queens.
And yes its possible and very plausible to have made 82 workers at the 8 minute mark. They turn into buildings. Maybe you're talking about workers alive.
My macro is not better than stephanos, but it takes me all but 100 apm to macro next to perfectly in zvp earlygame. Every high master should be able to match stephano.s drone count
|
Yeah, I'm counting workers alive, so when people do double extractor trick right when it's a 30 second mark, they're down 2 drones, or when you lose a scouting drone, etc.
On March 31 2012 18:53 Omnipresent wrote:Show nested quote +On March 31 2012 17:50 PeanutsNJam wrote:On March 31 2012 17:17 Omnipresent wrote: It's nice that you took inititive on this, but I'm not really sure what you're trying to say. There's no real conclusion here, and the data doesn't appear to be immediately useful. Your method, especially sample size (3 games or less per league), is really not sufficient. Also, I think it should be clear that when you say, "A large fraction of the players I studied in these leagues," you mean like 2-3 people per league. That assumes that you didn't use more than one game form a single player.
There also seems to be something seriously wrong with the platinum data. I'm not sure if you just got unlucky when picking games (easy to do when you chose 3 at random) or if you made some sort of honest mistake (mislabeling sounds like a reasonable explanation). I'd need to see a lot more before I believed that platinum players were making about 2/3 as many workers at 10 min as bronze players, and about 1/2 as many as diamond players. In any event, the explanation you provide doesn't sound at all satisfying.
The spending quotient thread you linked to has much better data (read larger sample size and over longer durations of game time) on worker production, and also provides useful analysis on spending. This thread looks redundant at best. It feels like people ignore 90% of what is in the OP. First of all, the purpose of this is to do a detailed comparison where you don't have extra factors and variables. In terms of the spending quotient thread, I wanted to investigate if the ideas in it still held true (which they did) when you eliminate a majority of the average deviation. Again, this isn't about the spending quotient. It's about the rate of workers produced and how that correlates with a player's league. It's also hard to find games where you have a ZvP on Shakuras Plateau, where neither player does an early game allin or pressure. One of the replays for master players I used, the protoss doesn't have a single unit go to the zerg's half of the map that isn't a probe in the first 9 minutes. PM me if you want proof of this. I have faith in the collective intelligence of the people here, so I'll assume you're just skimming. Please read carefully. I didn't misunderstand any of that. In fact, the only part of this post that I disagree with is where you say your results line up with the results in the SQ thread. They do not. The platinum numbers are wildly inconsistent between your numbers and the numbers in that thread. As far as I can tell, this is the only conclusuion you've drawn from this information, and it's demonstably false. I understand what you're trying to do: examine data in a specific matchup on a specific map given a specific set of circumstances. You're aming for specificity. You've arbitrarily chosen to look at games where players are essentially left alone to build drones. I just don't know why you did that. It's not useful in any way, and isn't particularly interesting either. Why should someone reading this care about the data? When I say "I'm not sure what you're trying to say," I don't mean that your english is unclear. I mean I read it all, understood it, and was left wondering why it mattered. I quite literally want to know why it matters how many workers are produced in this specific matchup, on this specific map, and under these conditions. What made you think of this? What kind of conclusions do you expect to draw? What point are you making? That's all taking for granted the premise that your data is somehow representative, which doesn't appear to be the case. If it's not representative, why bother looking at it? I don't blame you for the data. Based on the difficulty of finding enough games that fit your criteria, it probably isn't reasonable to expect your information to be representative. Again, that's true even if you could use the information to draw useful conclusions (which seems unlikely in the first place). So in addition to asking why this matters, I'm asking why it matters so much that you're willing to use such a small set of information to "examine" it. Why is it so important to look at this that you're willing to look at such a small snapshot just to get insight into this situation? I don't doubt that you used real replays and that those replays complied with the set of conditions you put forward. I don't feel like I need to see the actual games. I essentually trust you. You have no obvious reason to fake this information, and would have likely claimed a larger sample size if you were going to. That being said, 3 games per league isn't enough to indicate anything at all. The fact that your platinum numbers seem so wildly off base is indicitive of that fact. This is based on both personal experience and the drone production numbers, not the spending quotient numbers, in the spending quotient thread. + Show Spoiler +That thread puts platinum worker production numbers much closer to 30, which is higher than your number, not lower. Any of the complicating factors you chose to eliminate would only serve to lower that number. That is, your number should be equal to or greater than that number. I understand exactly what you're saying. It's not complicated. It just looks like a clumsy attempt to demonstate something that is neither useful nor interesting in the first place. I'm sure it's not entirely your fault. Finding replays to fit these criteria is probably both difficult and time consuming. But why settle for less?
The purpose in picking only replays where players were essentially allowed to macro under similar circumstances is to get a reliable comparison.
Think of it this way, it's like giving every single person the same physical body to play basketball vs each other. That way, you can evaluate their skill despite their differences in physical height/strength/etc.
Similarly, the goal was to evaluate the ability to create drones quickly in the early game between leagues where conditions are optimal for making drones, and how that correlates with skill.
I'm not saying all platinum players suck at making workers. I'm saying that platinum zerg players hesistate to dedicate larva to drones in the early game because of fear. I'm pretty sure that at the end of all the games, the platinum zerg players have more workers because they continue to build them. However, in regards to the rate at which workers are built in the early game, platinum seems to fall behind.
If I have to have an all-encompassing statement:
The rate at which you create workers in the early game without outright losing the match due to overdroning (which the bronze/silver/gold players did) is indicative of your skill.
or
To improve, build more workers more quickly (at least as zerg).
Again, the only reason the platinum is making workers more slowly (not less workers) than the bronze/silver/gold is because when you tell bronze/silver/gold to macro better, they make drones blind.
|
The title of your graph should be workers alive over time then, not workers built over time. Hence why I was confused as to the ridiculously low numbers
|
On April 01 2012 04:04 Jombozeus wrote: The title of your graph should be workers alive over time then, not workers built over time. Hence why I was confused as to the ridiculously low numbers
I c, good point. I think all the dips in workers are when a building is built or a scouting drone dies. I think in one game a zealot walked in and killed like 2 drones (gold) or something.
|
I'm sorry OP but with such an incredibly small sample size you can't really come to any justifiable conclusion, as its far too probable that you could've got those results simply by chance with a null hypothesis (no relationship).
|
How many games did you sample? Because it's impossible that platinum players' drone production is THAT terrible. I'm a Platinum, and I just checked one of my replays, and I had 52 drones at 8 minutes. Unless you were watching replays of people who had been 6 pooling all the way up until Plat, and then decided to learn how to macro, you must have seriously messed up somewhere.
|
On April 01 2012 04:06 PeanutsNJam wrote:Show nested quote +On April 01 2012 04:04 Jombozeus wrote: The title of your graph should be workers alive over time then, not workers built over time. Hence why I was confused as to the ridiculously low numbers I c, good point. I think all the dips in workers are when a building is built or a scouting drone dies. I think in one game a zealot walked in and killed like 2 drones (gold) or something.
I think you all are onto something here. If I were to add to or revise your study (I do social sciences and humanities, so I wont be), I'd look at drone use rather than drone count. Perhaps you could measure the ratio between total drones built and total drone alive/harvesting.
In other words, your work is prompting a conversation about an interested aspect of Zerg decision making: how to efficiently balance drones for harvesting vs drones for buildings. We are accustomed to talking about Zerg decision making as a case of drones versus army (larva allocation). Your work has suggested a possible need to look into drone allocation.
Perhaps Stephano and other pros have a more efficient method of balancing drone workers vs drones for buildings.
I bring this up because, while I think you need more data to draw conclusions, I am interested in your overall hypothesis that players in plat might be unnecessarily defensive, failing to drone as hard as they could. I notice some plat players dropping spines at really inefficient times.
The difference in dropping a spine 5 minutes before you are attacked and dropping a spine one minute before you are attacked has large economic consequences. I'd hypothesize that someone like Stephano makes more drones for harvesting per drones mades for buildings than someone in masters.
So to the OP, while I don't have a background in stats, I agree with those who suggest that, as of now, your parameters are a bit arbitrary and that you need a large sample size. THAT SAID, I think you are onto something interesting and I would like to it explored in more detail.
tl;dr I am particularly interested in efficiency/lack of efficiency with which Zerg at different levels manage the ratio of drones for economy versus drones for buildings.
|
On April 01 2012 04:23 XenoX101 wrote: I'm sorry OP but with such an incredibly small sample size you can't really come to any justifiable conclusion, as its far too probable that you could've got those results simply by chance with a null hypothesis (no relationship).
This would be true were matchmaking non-existent. If I randomly took 19 players, and made a chart, and said something like "all people who wear blue while playing Starcraft have poor macro", then yeah, the sample size is too small.
However, the fact that matchmaking already puts people into separate leagues makes the sample size significant. I know that platinum/diamond players are worse than a master league player (it's not a hypothesis). I wanted to know in regards to drone production in the early game, how much?
Going back to the example of basketballs, I know that NBA shooting guards are better than Division 1 shooting guards. If I want to know by how much, I think it's entirely reasonable to take 3 shooting guards from the NBA and Division 1, and compare their averaged shooting %. It's the comparison that matters.
On April 01 2012 04:25 ipwntbarney wrote: How many games did you sample? Because it's impossible that platinum players' drone production is THAT terrible. I'm a Platinum, and I just checked one of my replays, and I had 52 drones at 8 minutes. Unless you were watching replays of people who had been 6 pooling all the way up until Plat, and then decided to learn how to macro, you must have seriously messed up somewhere.
A.) What was the rate at which your drones were produced? Did you suddenly make 10-20 at the 9 min mark? Or did you consistently make drones and kept your money low without being supply blocked, and only made army when absolutely necessary (like pros do)? B.) Did you win, or did you drone blindly and die?
On April 01 2012 04:33 skatbone wrote:Show nested quote +On April 01 2012 04:06 PeanutsNJam wrote:On April 01 2012 04:04 Jombozeus wrote: The title of your graph should be workers alive over time then, not workers built over time. Hence why I was confused as to the ridiculously low numbers I c, good point. I think all the dips in workers are when a building is built or a scouting drone dies. I think in one game a zealot walked in and killed like 2 drones (gold) or something. I think you all are onto something here. If I were to add to or revise your study (I do social sciences and humanities, so I wont be), I'd look at drone use rather than drone count. Perhaps you could measure the ratio between total drones built and total drone alive/harvesting. In other words, your work is prompting a conversation about an interested aspect of Zerg decision making: how to efficiently balance drones for harvesting vs drones for buildings. We are accustomed to talking about Zerg decision making as a case of drones versus army (larva allocation). Your work has suggested a possible need to look into drone allocation. Perhaps Stephano and other pros have a more efficient method of balancing drone workers vs drones for buildings. I bring this up because, while I think you need more data to draw conclusions, I am interested in your overall hypothesis that players in plat might be unnecessarily defensive, failing to drone as hard as they could. I notice some plat players dropping spines at really inefficient times. The difference in dropping a spine 5 minutes before you are attacked and dropping a spine one minute before you are attacked has large economic consequences. I'd hypothesize that someone like Stephano makes more drones for harvesting per drones mades for buildings than someone in masters. So to the OP, while I don't have a background in stats, I agree with those who suggest that, as of now, your parameters are a bit arbitrary and that you need a large sample size. THAT SAID, I think you are onto something interesting and I would like to it explored in more detail. tl;dr I am particularly interested in efficiency/lack of efficiency with which Zerg at different levels manage the ratio of drones for economy versus drones for buildings.
Unfortunately, I thought about stuff like that, and it's not something that can be realistically statistically analyzed without some kind of super computer. Building timings, gas timings, maynarding workers, optimal saturation, hatchery timings, spine timings, sim city, etc. Not unlike predicting the weather, which takes a ton of equipment and manpower.
|
On April 01 2012 05:03 PeanutsNJam wrote:Show nested quote +On April 01 2012 04:23 XenoX101 wrote: I'm sorry OP but with such an incredibly small sample size you can't really come to any justifiable conclusion, as its far too probable that you could've got those results simply by chance with a null hypothesis (no relationship). This would be true were matchmaking non-existent. If I randomly took 19 players, and made a chart, and said something like "all people who wear blue while playing Starcraft have poor macro", then yeah, the sample size is too small. However, the fact that matchmaking already puts people into separate leagues makes the sample size significant. I know that platinum/diamond players are worse than a master league player (it' snot a hypothesis). I wanted to know in regards to drone production in the early game, how much?
That wasn't your hypothesis though, if your hypothesis was simply whether platinum/diamond players are worse than master league players then you would have no reason to conduct this 'study' as that's already been proven by the matchmaking system. Your hypothesis is specifically that drone production is greater in certain leagues than other leagues, which cannot be inferred at all by their league since league is only reflective of win rates against similarly skilled players.
|
I really do appreciate your effort but I dont' think you have the experience with designing a study to make this kind of thing work. XenoX101 is onto something when he talks about your hypothesis. He's right.
Quit telling everyone to re-read the original post. That's not the issue. We understand what you're doing (at least the parts you've told us about).
In addition to the facdt your sample size is way too small (sample size for each leauge, because you're comparing between them, is what counts), here are some other sources of bias in your study: * Your pulled your results from replays that people chose to upload. These games may be different in some way than others. * You chose 1 map * You chose 1 matchup * It's possible you didn't choose games at random, but rather at a particular timeframe, from particular people, etc. * You used a partially-subjective criteria to filter games
The irony is your sarcastic retort of taking 19 players and seeing if wearing blue improved macro... that's actually coming closer to a real study, and 20 people in that study might actually approach a reasonable sample size. But you didn't do that. You took 19 games across 6 leagues and used that to draw this conclusion. Between group versus within group comparison would be the term.
Also, you mentioned standard deviation a few posts back... what is the standard deviation here?
|
Your sample size is ridiculously small. Absolutely nothing can be drawn from this analysis. I mean... the data is completely worthless, sorry.
There's also lots of errors you are failing to adjust here for, such as what build the opponent is going. In higher leagues, people are more apt to FFE rather than gateway expand.
|
I think what you have here is a potential jumping off point for more study. But I agree that you can't draw any conclusions from what you have currently. Here's what comes to mind for me.
On March 31 2012 14:24 PeanutsNJam wrote: - Diversity of skill within a league. Expanding upon this, it could be possible that I happened upon 3 platinum ZvP's on Shakuras Plateau that didn't involve early aggression, where all three zergs sucked about equally as much (check out how close the data points are to the linear fit). There will most likely also be the most deviation in bronze and master, with the least in silver/gold/plat.
The first thing you guys notice after inspection is probably that platinum players have the lowest drone production of all the leagues. This looks like I screwed up bad, but upon reflection and closer study of the replays, it makes sense. The platinum replays were by far the most consistent. I can upload the replays if you guys don't believe me.
Here is what I believe the progression of Bronze -> Diamond is. You do 1/2 base allins until plat, then you start learning how to play. The people who struggle to "macro better" in bronze-gold are floundering around, not knowing when to drone, when to build army, etc. In platinum, they are constantly scared of aggression and under drone.
It's also likely that bronze/silver/gold players make drones blindly, while platinum players are more cautious. Again, notice the difference here is around 6 drones before the 10 min mark. I'm sure by the end of the game, platinum players generally have more total workers built.
I think it's interesting that you have a pretty reasonable correlation between total drones constructed and league, with the exception of Platinum. I don't think you can draw the inferences you are from the data that you have. I don't think food counts are necessarily useless, because they can be used to corroborate the drone counts. You state that food counts wound up between 60-80 at 10 minutes, and hey, that's pretty much the drone count difference too. So it lines up and nothing funny is going on.
So, you have a couple of hypotheses:
1. That Bronze-Gold league Zergs in general will overdrone blindly 2. That Platinum league Zergs in general will play a much more risk averse style blindly. 3. That Diamond-Master league Zergs can identify correctly when it is safe to drone.
Now suddenly getting samples of a specific map & match up & "macro game" is no longer actually necessary, and probably detrimental. Sticking with ZvP is fine (or at least staying away from ZvZ), take a sample of maps, and set the criteria to something broader, such as "Games must have lasted longer than 10 minutes".
That way you can get a much larger sample size, have a good variety of data to work with, and if Platinum league zergs are still showing these crazy low drone counts on average, you've got a lot more data to support the assumptions you're making here.
|
On March 31 2012 17:08 Coramoor wrote: what level master players did the have that they averaged 35ish workers at the 8 minute mark, if we're not talking about a zealot pressure or something that forces spines and wastes larva, then i find it hard to believe that any competent masters zerg doesn't have 55-60 drones at 8 minutes
I agree with this. The data shown in the OP is just so far away from expectations that I have a really hard time swallowing it. This is the type of scenario where you really do need a larger sample size.
@OP: You said you analyzed only 19 replays, since the number is small, could you upload them all as a replay pack to dop.sc or a similar site?
|
Edit : Can we remove posts ?
|
|
|
|