|
Germany511 Posts
I can't really go through seven pages right now cause the posts are all so lengthy (and I think that speaks to the high level of discussion) but here are some points:
Data is skewed towards equality because lower league players don't think about trade offs in terms of queen usage. Higher level players might inject or get a creep tumor, depending on whether they can use the larvae in time (if they stay idle the gain is near zero as the base-larvae don't regenerate). They might also get earlier expansions before pools which explains the lower inject percentage before 5min. Higher level players are also under a lot more aggressive pressure and use queens in their defense, often times missing injects because they block chokes and fight units. Lower level players will at the same time just go by conventional wisdom and "inject, inject, inject". We can also assume that it's easy to inject when people constantly look at their base which happens more in lower leagues as higher level players are more active around the map, scouting and fighting early skirmishes. It's an attention and APM trade off that could just as well show that higher level players are so good at injecting that they keep their hatcheries injected even under the greater requirements of higher level games. Queens might also be pulled off hatcheries in late game for offensive transfusing. Finally it could be the case that because every Zerg is always taught that injecting is key that all Zerg tend to learn this skill first, putting the learning curve about injects at a very different point than the learning curve about the game in general. It makes sense as it is a very binary thing with only one positive outcome. To get better results we would need to take equal players with random inject rates which is obviously impossible.
Overall I think it's WAY too early to draw conclusions from data like this. It should still be important for things like the Staircase although adjusted for game time. It is however very cool to see it and I thank you for your work. Maybe after a peer review it will get published for a broader audience 
|
You guys need to calm down, the article can be explained a better way:
Good macro leads to better injects, Better injects don't lead to good macro.
(also, you'll notice that master's DO outperform silver, but its not by much).
|
On May 22 2013 08:06 LaLuSh wrote:Show nested quote +On May 22 2013 07:40 Shousan wrote:
I am in no way condoning the research and work that you put in this article, the data alone is pretty interesting, but the analysis and 'conclusions' are just way off and can be very misleading.
As others have mentioned, inject percentage comparison is meaningless as master players need to put up with far more things throughout the game, you need to give numbers a lot more perspective, cross-relating harrass, multitasking and actions with injects will give you a far better scope of what to look for in these statistics.
If we looked at professional car racing and had a percentage of accidents per competitor and then compare it to the percentage of accidents from an amateur or casual driver we could see that they'll maybe have a similar percentage, but this in no way means that the amateur is near the level of a pro driver, it may be because a pro will have tougher competition, trickier tracks, and many more things to factor in. This is why a correct approach to the numbers is far more important than just having a bunch of data and immediately assume that your comparison is correct.
Because Starcraft is a game that has an enormous amount of variables to account for, you'll have a ton of variation in replays and scenarios when the winning factor is more than just a 'key mechanic', how much of this silver level percentage would drop if he was facing a masters opponent? Probably a lot. Therefore, even if we're talking about the same 'metric' (inject percentage) it's just bad practice to even suggest that those percentages can be compared and have a valid conclusion.
The problem with these articles is that many people don't even stop to analyse and just take it for granted because there's a lot of data and a nice spreadsheet and then try to account the data as 'proof' that a race is easier or harder than the others, which just invites more ignorant discussion.
TLDR nice idea, great effort, cool statistics, pretty blind analysis and conclusions. Their theory and analysis is solid. Rather it's posts like yours, filled to the brim with pseudo-statements, that provide the blind analysis. Let's see what their analysis is: 1. larva to resource ratio decreases as the game progresses and you tend towards using higher tech units (ultras, broodlords etc cost both more resources and supply). 2. Mined out locations keep supplying larva. Less risk of "larva blockage" late game. I think they are solid explanations. Only thing I'd like to see added is that you spend a considerable time maxed out during the late game (partly covered by larva to resource ratio and larva blockage explanation). As they state in their analysis: When you resupply and/or remax lategame you tend to resupply on high tech, high resource, high supply demanding units and less so on cheep larva demanding ones. I can't spot what's so blind about their analysis? Apart from the fact that you and most others similar posters in the thread probably didn't even bother reading it. The truly blind ones are those who don't bother addressing specific points but rather use vague language to cover up their contribution to an "ignorant discussion".
"That means (assuming normally distributed scores) that about 30% of masters zerg games have an inject % that's worse than the average silver inject %. They did worse than the average silver game, yet somehow they are in masters with sub-silver injects."
That's a blind statement, saying that "somehow they are in masters with sub-silver injects" is just wrong, as a silver player in the same situation (vs a master player) would have a considerably lower percentage, add the fact that silver players won't choose correctly whether to use a tumor or inject and just plainly inject every time.
If I can manage to make 60 drones in 8 minutes in bronze league it doesn't mean that I have the same macro as a GM (at least for those 8 min), no matter how much the numbers look alike, since a GM will probably have more things happening in those 8 minutes and still manage to get those numbers.
There's a lot to factor in to make this proof that inject efficiency IS an important skill, their conclusions said it was NOT, since they decided to not incorporate inject consistency as a part of their training program. The points you mentioned are theories behind their reasoning while the conclusion is that inject efficiency is not an important skill (not important enough to teach thoroughly, anyway), they back this up with a standard deviation which in this case doesn't really prove a lot and is still consistent with other races and their respective macro mechanic. This is a poor analysis, because they're rushing to a conclusion without hard evidence behind it and just a couple of statements that are pretty clear without looking at the data even.
You can get far more interesting conclusions from the information gathered than a plain "Since higher tier units consume less larvae, then perfect injects are less important the longer the game goes, also, you have more bases etc..." well, no kidding, but that's not the sole reason for it, the longer the game goes, you need to focus on a lot more things, hence the mechanic being more forgiving, and that's true not just for zerg, but for the other races as well, the difference between leagues in MULE percentage or Chrono-Boost percentage with the given standard deviations are still not far from one another, and both follow the same tendency to drop the longer the game goes, is it because dropping MULES is less important in the late game? Don't think so, it's probably because there's simply more going on and the mechanic is built around that fact, so it'll be more forgiving if you are not as precise with it given that you have a stable economy and good numbers of infrastructure to produce.
The fact that there are people thinking that this is not a good analysis doesn't mean that we didn't bother reading it, more like some might recognize that in a study it's pretty difficult to do a great analysis even with good information if you don't approach what you want to prove correctly and just decide something with poor basis. It's not about making pseudo statements or using vague language, is about looking objectively at something that has the potential to be relevant and it's not because of how the information is treated.
|
Super important early game and gets less important as the game progresses.
|
Countless times zergs at the pro level have lost in the late game, because of not enough larva to defend quickly enough.
|
The primary conclusion of this article is that there is no useful benchmark for inject %.
Those who disagree, please propose a benchmark for inject %. I would love to have one!
|
Obviously injects become less important in a lategame scenario. You're spending more money on fewer units--larva becomes an abundant resource. However in the early-midgame injects are VITAL. Say you get hit by a timing from a terran or protoss. If you haven't been keeping up with your injects (or don't during the fight) you won't have enough units to keep up. It's the same as late game you will have many more hatches to make larva on their on.
And to people saying this is obvious proof that zerg macro is easy, that's bullshit and you know it. Macro for all 3 races is DIFFERENT. It's like comparing Starcraft and Chess. You just can't because they're just so different. Starcraft is a game of incomplete knowledge, whereas chess is a game of complete knowledge. That's just ONE difference out of a vast array. Calm the fuck down people, this isn't a warrant to whine.
This isn't as gamechanging as people think.
|
I think a more correct approach would be to base it on queen energy effiency, rather than just injecting, since creep tumors also play a huge role in the outcome of the game and cost of opportunity forces you to choose when is better to use one or the other at least in the early stages of the game, this is just basing on the fact that the goal of that study was to determine whether inject efficiency is relevant enough to incorporate to your methodology, and it sure has more variables to it (queens just to spread creep, to attack, etc.) but it'd be more indicative of skill throughout leagues
|
I'd love to see where Life is on this chart, or other Code S level Zergs.
|
Injecting is hard bc you can't spam them all at once like MULES and to a lesser extent, Chronos (on multiple upgrades/production facilities, etc)
|
On May 22 2013 08:06 LaLuSh wrote:Show nested quote +On May 22 2013 07:40 Shousan wrote:
I am in no way condoning the research and work that you put in this article, the data alone is pretty interesting, but the analysis and 'conclusions' are just way off and can be very misleading.
As others have mentioned, inject percentage comparison is meaningless as master players need to put up with far more things throughout the game, you need to give numbers a lot more perspective, cross-relating harrass, multitasking and actions with injects will give you a far better scope of what to look for in these statistics.
If we looked at professional car racing and had a percentage of accidents per competitor and then compare it to the percentage of accidents from an amateur or casual driver we could see that they'll maybe have a similar percentage, but this in no way means that the amateur is near the level of a pro driver, it may be because a pro will have tougher competition, trickier tracks, and many more things to factor in. This is why a correct approach to the numbers is far more important than just having a bunch of data and immediately assume that your comparison is correct.
Because Starcraft is a game that has an enormous amount of variables to account for, you'll have a ton of variation in replays and scenarios when the winning factor is more than just a 'key mechanic', how much of this silver level percentage would drop if he was facing a masters opponent? Probably a lot. Therefore, even if we're talking about the same 'metric' (inject percentage) it's just bad practice to even suggest that those percentages can be compared and have a valid conclusion.
The problem with these articles is that many people don't even stop to analyse and just take it for granted because there's a lot of data and a nice spreadsheet and then try to account the data as 'proof' that a race is easier or harder than the others, which just invites more ignorant discussion.
TLDR nice idea, great effort, cool statistics, pretty blind analysis and conclusions. Their theory and analysis is solid. Rather it's posts like yours, filled to the brim with pseudo-statements, that provide the blind analysis. Let's see what their analysis is: 1. larva to resource ratio decreases as the game progresses and you tend towards using higher tech units (ultras, broodlords etc cost both more resources and supply). 2. Mined out locations keep supplying larva. Less risk of "larva blockage" late game. I think they are solid explanations. Only thing I'd like to see added is that you spend a considerable time maxed out during the late game (partly covered by larva to resource ratio and larva blockage explanation). As they state in their analysis: When you resupply and/or remax lategame you tend to resupply on high tech, high resource, high supply demanding units and less so on cheep larva demanding ones. I can't spot what's so blind about their analysis? Apart from the fact that you and most others similar posters in the thread probably didn't even bother reading it. The truly blind ones are those who don't bother addressing specific points but rather use vague language to cover up their contribution to an "ignorant discussion".
He actually did have some solid points about the easiest to mess up aspects of statistics, it's not technically pseudo-statements. His example made the point of interpolation vs extrapolation (statistics doesn't allow for extrapolation), and about hidden variables possibly skewing the data (correlation does not equal causation). Not that he's necessarily right in this case (I don't see any extrapolation in the article, though possible hidden variables that could arise from the inject % calculation method weren't really discussed much), but he's not wrong for being skeptical of the analysis with his more major point of hidden variables at least, what with statistics being inherently fuzzy math (though the justification of the analysis in the study adequate enough for this to be a good jumping off point for more study and has good "Directions for Future Exploration") and starcraft being incredibly complex with the interactions of its many macro variables.
|
I remember I has such good injects in gold, but then got worse at them when I started focusing on other aspects of the game. In diamond I actually don't think my injects were as good as they were in gold when that was all I cared about. Now that im highish masters I think im finally back to gold level injects xD
|
On May 22 2013 09:47 dsjoerg wrote:
The primary conclusion of this article is that there is no useful benchmark for inject %.
Those who disagree, please propose a benchmark for inject %. I would love to have one!
.
Could you calculate an opportunity value of missed injects? it could be a simple formula of %missed * # of hatcheries * time and then regress that against win rate.
|
I hate to be that guy, but apart from theoretizations on these very unexpected results, there also seems to be something odd about the data you're using.
This image suggests that IdrA injects his third hatchery only haphazardly:
![[image loading]](http://a1.ggtracker.com/assets/inject0-4703359242c5e3c0620066d190e44dbf.png)
It does not seem to account for the fourth (macro) hatch right next to it that gets injected by the same queen:
![[image loading]](http://i.imgur.com/VPwudFBl.jpg)
However, to add insult to injury, if you account for the fourth hatchery (never mind the fifth, it doesn't get any injects), IdrA's inject efficiency goes down to 55% (according to SC2gears, which computes Main building control (Zerg) for four hatcheries). In other words, because he injects faster than one hatch can produce larvae, his injects are to be considered less efficient. That can't be right?
Honestly, I don't see "Inject % = (total # of minutes all hatches spent with injected larva) / (total # of minutes all hatches were active)" actually measuring what you're aiming for... like, at all. 
|
On May 22 2013 06:17 Eventine wrote:
I swear, every time someone provides some data point, instead of sparking interesting debate, half the responses are like your data is wrong, my intuition is perfect and therefore i reject your findings.
This. I find it amusing how every Zerg (well Zerg mains because I do play Zerg) player suddenly attempted to become a statistics major. Quite convenient, eh? To quote someone else:
A while back, someone collected similar data for workers-produced, time-supply-blocked, and surplus-resources-banked. Their results showed vast differences between the different divisions, implying that the above metrics are a major differentiating factor between the different divisions & skill levels. The empirical approach does work.
The above metrics are also subjected to the same "real active game world" scrutiny in the same exact way, and yet they scale non-marginally with skill level.
|
On May 22 2013 10:54 Poffel wrote:
I hate to be that guy, but apart from theoretizations on these very unexpected results, there also seems to be something odd about the data you're using.
Thank you for checking the data! I love "that guy"  usually he's me.
I ran my code in "debug mode":
+ Show Spoiler +
Active:09.40 Injects:06.40 Last:14.24
Hatchery [3B80001] Injects: 04.44 05.27 06.15 07.01 07.44 08.25 09.12 10.00 11.13 13.02
Active:06.27 Injects:02.40 Last:14.24
Hatchery [4680001] Injects: 07.57 08.42 10.37 11.45
Hatchery [5D80001] Injects: 09.49
Active:09.39 Injects:06.40 Last:14.24
Lair [32C0001] Injects: 04.45 05.30 06.12 07.00 07.45 08.32 09.26 10.35 11.47 13.01
The Hatchery [5D80001] you're writing about gets only one inject, at 09.49. I dug into my code and there's an extra rule that I didn't describe in the article: a base must receive more than one inject to be considered at all for the inject % measurement.
So, no bug here.
|
On May 22 2013 03:54 Embir wrote:
Finally solid confirmation that Zergs macro is the easiest - we already knew they had it easy with only one production building and easiest tech switches in the game, now we know that they macro mechanic is also forgiving - and note that supposed unforgiveness of zerg mechanics was main argument for zerg's macro difficulty.
Ye cause Terran/Toss is not unforgiving at all. If you don't hit those crucial mule's before capped 200/200 you're gonne be in trouble! Same with Chronoboost, imagine waiting 50 seconds longer than u could've, instant GG. Terran/Toss has easier macro mechanics as it doesn't penalize you by waiting with chrono/mule (more than delaying timings).
|
Yeah you really can't get much information from data like this. So many other things to take into consideration such as the number of bases, macro hatches, queens being used for other purposes. A masters player is far more likely to use a transfuse during a rush or a creep tumor early on, etc. But of course people are just going to read the summary and instantly cry about how Zerg takes no skill or something.
|
On May 22 2013 03:54 Embir wrote:
Finally solid confirmation that Zergs macro is the easiest - we already knew they had it easy with only one production building and easiest tech switches in the game, now we know that they macro mechanic is also forgiving - and note that supposed unforgiveness of zerg mechanics was main argument for zerg's macro difficulty.
What the... how did you manage to get the idea Z has easiest macro from data that says inject might not be the most deciding macro? If anything, Z is the only race that cant constantly build troops without going back to their base screen.
|
On May 22 2013 11:09 dsjoerg wrote:Show nested quote +On May 22 2013 10:54 Poffel wrote:
I hate to be that guy, but apart from theoretizations on these very unexpected results, there also seems to be something odd about the data you're using.
Thank you for checking the data! I love "that guy" usually he's me. I ran my code in "debug mode": + Show Spoiler +
Active:09.40 Injects:06.40 Last:14.24
Hatchery [3B80001] Injects: 04.44 05.27 06.15 07.01 07.44 08.25 09.12 10.00 11.13 13.02
Active:06.27 Injects:02.40 Last:14.24
Hatchery [4680001] Injects: 07.57 08.42 10.37 11.45
Hatchery [5D80001] Injects: 09.49
Active:09.39 Injects:06.40 Last:14.24
Lair [32C0001] Injects: 04.45 05.30 06.12 07.00 07.45 08.32 09.26 10.35 11.47 13.01
The Hatchery [5D80001] you're writing about gets only one inject, at 09.49. I dug into my code and there's an extra rule that I didn't describe in the article: a base must receive more than one inject to be considered at all for the inject % measurement. So, no bug here.
Ok, there's no bug but a conscious exclusion... still, that's one inject lost in the void, and IdrA receives a lower efficiancy rating than if he had applied that inject to the other hatchery, whereas it stands to reason that it does not matter in the slightest which hatchery he injects. Don't get me wrong, your results are indeed quite surprising, and I - like many others - would have expected a huge difference between the leagues even under your conditions of observation. Still, a macro hatch is far from an obscure incident. In your model, if a player distributes a queen's injects between several hatcheries, his efficiency drops by a large margin. Frankly, I'm unsure whether that's the kind of noise we have to tacitly accept for useful modelization or if it has to be considered a severe flaw of the model you're using.
|
|
|
|
|
|
EPT
