|
On December 20 2012 00:35 Moka wrote:
Why bronze and silver players have suddenly a SQ spike @ 25 minute mark? o.o It also applies to other leagues, but the spike is less noticeable. And I also wonder what is so special at the 25 minute mark.
Good catch! I realize now that must be an artifact of how I smoothed the data. Between 10 and 25 minutes inclusive, I took the average of the window (x-5,x+5). So 10 minutes and 25 minutes are the join points between smoothed data and raw data.
I smoothed the data because the raw data is noisy/jumpy, which produces weird and undesirable effects when you use the data to compute your Spending Skill.
|
On December 20 2012 04:20 GenesisX wrote:
Wow what happens at 25 minutes that makes all leagues spike in their spending?
Hive tech  . Mining out, too.
|
On December 20 2012 07:35 dsjoerg wrote:Show nested quote +On December 20 2012 00:35 Moka wrote:
Why bronze and silver players have suddenly a SQ spike @ 25 minute mark? o.o It also applies to other leagues, but the spike is less noticeable. And I also wonder what is so special at the 25 minute mark. Good catch! I realize now that must be an artifact of how I smoothed the data. Between 10 and 25 minutes inclusive, I took the average of the window (x-5,x+5). So 10 minutes and 25 minutes are the join points between smoothed data and raw data. I smoothed the data because the raw data is noisy/jumpy, which produces weird and undesirable effects when you use the data to compute your Spending Skill.
Maybe lower the window size and spread it across the whole graph (x-2,x+2) to get a nice rolling average?
|
This measure could be expanded and normalised per race if you were to calculate a seperate set of quotients.
1. Production Quotient: Percentage of time all buildings spent unused below 200 supply
2. Energy Quotient: Harder to do for all races but could very easily be weighted relative to their importance.
3. Efficiency Quotient: Damage Inflicted-Damage Taken/Damage Taken. Damage Taken will ALWAYS be non-zero.
|
What happens at 25 min?
Edit: answered
|
On December 20 2012 08:09 Evangelist wrote:
This measure could be expanded and normalised per race if you were to calculate a seperate set of quotients.
1. Production Quotient: Percentage of time all buildings spent unused below 200 supply
2. Energy Quotient: Harder to do for all races but could very easily be weighted relative to their importance.
3. Efficiency Quotient: Damage Inflicted-Damage Taken/Damage Taken. Damage Taken will ALWAYS be non-zero.
Unfortunately these things are very hard to do as the replay doesn't actually provide the information
For 1, to get a really accurate number you have to simulate the whole game. You don't get things like if a unit was built, just that there was an attempt to build the unit, nor do you get that it was delayed due to supply block, etc.
For 2, you have to simulate how much energy things have and subtract for casting.
For 3, also requires simulating the game, though I think you can get unit deaths which may be enough to get some decent numbers.
|
On December 20 2012 03:24 iEchoic wrote:
These are really cool statistics, would love to see more of this. It would be really cool to see timings by league as well (for example, upgrade timings, etc.). I have a feeling that more refined build orders are a major factor separating GM from Master's and Diamond.
Great idea iEchoic, we'll do it!
|
On December 20 2012 08:06 Wraithan wrote:Show nested quote +On December 20 2012 07:35 dsjoerg wrote:On December 20 2012 00:35 Moka wrote:
Why bronze and silver players have suddenly a SQ spike @ 25 minute mark? o.o It also applies to other leagues, but the spike is less noticeable. And I also wonder what is so special at the 25 minute mark. Good catch! I realize now that must be an artifact of how I smoothed the data. Between 10 and 25 minutes inclusive, I took the average of the window (x-5,x+5). So 10 minutes and 25 minutes are the join points between smoothed data and raw data. I smoothed the data because the raw data is noisy/jumpy, which produces weird and undesirable effects when you use the data to compute your Spending Skill. Maybe lower the window size and spread it across the whole graph (x-2,x+2) to get a nice rolling average?
I tried something like that, however for the 5 minute mark, you don't really want to average SQ from 3 to 7 -- because games don't end at 3 minutes anyway... nor do you want to use the average from 5 to 7 as the stat for 5. Which led me to the somewhat braindead approach I'm using now, simply using the average of games that lasted from 5:00 to 5:59.
Now that I'm thinking about it, it might be better for 5 to be simply 5, but 6 can be the average of 5,6,7, and 7 can be the average of 5,6,7,8,9, etc.
|
On December 20 2012 08:09 Evangelist wrote:
This measure could be expanded and normalised per race if you were to calculate a seperate set of quotients.
1. Production Quotient: Percentage of time all buildings spent unused below 200 supply
2. Energy Quotient: Harder to do for all races but could very easily be weighted relative to their importance.
3. Efficiency Quotient: Damage Inflicted-Damage Taken/Damage Taken. Damage Taken will ALWAYS be non-zero.
Agreeing with Wraithan that 1 is really tough to do. Hard to know how many workers a player has at any time, thus hard to know when they are at/below 200 supply.
#2 I think is doable, we've got something like that for Zerg up on the site now.
#3 is doable in a way, not in terms of damage, but in terms of the resource value of units killed. And what we have is not quite killed but "removed from opponent's active army" which usually means killed. We estimate the army composition at every point in the game, so based on that we can estimate death...
|
@dsjoerg :
Can you provide more statistics? Value of the current data is unknown as you do not provide your methods or deeper statistics, such as user counts (you likely have much more data from the uploaders than from their opponents), distribution to different levels, time period of the matches taken into account, sample sizes for different levels & different game lengths, etc.
How do you collect the data as SQ cannot be calculated based on the info stored in the replay? Is the data entirely based on s2gs file imports (the announced amounts, 40k for US and 80k for EU, sound very high considering the tool never seemed to become popular by looking at TL threads on it) or have users manually reported core values? With a quick view some match reports seem to be based on only replay file, some on s2gs file and some have combined data from both two sources (your description suggests that you would fetch the s2gs data based on the replay? Is the s2gs id stated in the replay file?)?
Edit: sample sizes seem to have been added
|
On December 20 2012 14:21 korona wrote:
@dsjoerg :
Can you provide more statistics? Value of the current data is unknown as you do not provide your methods or deeper statistics, such as user counts (you likely have much more data from the uploaders than from their opponents), distribution to different levels, time period of the matches taken into account, sample sizes for different levels & different game lengths, etc.
How do you collect the data as SQ cannot be calculated based on the info stored in the replay? Is the data entirely based on s2gs file imports (the announced amounts, 40k for US and 80k for EU, sound very high considering the tool never seemed to become popular by looking at TL threads on it) or have users manually reported core values? With a quick view some match reports seem to be based on only replay file, some on s2gs file and some have combined data from both two sources (your description suggests that you would fetch the s2gs data based on the replay? Is the s2gs id stated in the replay file?)?
Edit: sample sizes seem to have been added
Quick answers:
- as you've noticed, sample sizes are now available at http://ggtracker.com/spending_skill_stats. You can hover your mouse over any data point to see the # of matches for that data point.
- indeed yes, it is based entirely on s2gs file imports.
- matches were retrieved directly from SC2. To build this dataset, we retrieved s2gs files more or less at random, whether or not we had the replay for that match, keeping track of what league the players were in at the time.
- the matches are all 1v1 Ladder only
- the tool has actually been pretty popular! >200k replays uploaded
 although we have room for more
I'll add more details tomorrow -- in particular the dates from which these games are drawn, and hopefully I can find a good way to let you drill down into the individual games on which the stats are drawn so that you can confirm suspicious / interesting data points for yourself.
|
Has anyone linked Build Orders to SQ yet? Or is this just a data mining exercise to deem what is efficient assuming you are either spending all your resources, or not.
|
On December 28 2012 03:14 Allenansgar wrote:
Has anyone linked Build Orders to SQ yet? Or is this just a data mining exercise to deem what is efficient assuming you are either spending all your resources, or not.
Could you explain in more detail what you have in mind? Identifying the "build order" that was used in any particular game is quite difficult, depending on exactly what you mean. That is, in any particular game there is the list of what each player built and when, which is literally the Build Order for that game. So there are millions and millions of different possible Build Orders. However I suspect that linking these raw Build Orders to SQ is not quite what you have in mind.
|
I'm slightly surprised, because zergs spend money on spines and spores while maxed, and I rarely see Terrans and Protosses do the same with turrets/PF and cannons
|
I find this statistic deceiving...
So right now I am low-mid Master Protoss player on NA who used to be high Master but hasn't been playing much. One of my strengths is that I spend my money really well. Thus, I got an average 102.5 as my spending quotient with a game length of 24 minutes. This is well above GM.
But one my weakness is that I often don't take in quite as much as my opponent, because sometimes I cut my Probe production too soon or don't take a 3rd quick enough. So my macro in terms of raw income is poor, but my ability to spend wisely and not pool resources is good which leads to a good spending quotient.
I believe this quotient needs to have some consideration of how much resources are being taken in and when. If I should be collecting X amount at Y time, but I have less, then the fact I spending those resources more efficiently is meaningless.
|
On December 30 2012 10:07 BronzeKnee wrote:
I find this statistic deceiving...
... So my macro in terms of raw income is poor, but my ability to spend wisely and not pool resources is good which leads to a good spending quotient.
BronzeKnee, I agree that Spending Skill is limited to measuring of only one aspect of macro -- your Spending. The plan for GGTracker is to measure and track several other skills. You mentioned income, do people think it'd be a good idea to have an Income Skill? It would compare your income in a game to other players of that race in games of that length.
I'm open to any and all suggestions, especially ones that find a pretty broad consensus in the community.
|
This is nice to see. I have some thoughts on the results:
1) Looking at durations from about 8 to 20 minutes, where the great majority of games end, it is interesting to see how things have changed since September 2011.
The improvement in average SQ for Grandmasters appears to be relatively small, if non-existent. Previously the average SQ was 82. For NA Terran players, it now appears to now be around 86, but for Protoss only about 78, and for Zerg about 82 (averaging across the downward sloping line). Since there are similar numbers of players of each race, this comes out to an average of about 82, i.e., no net improvement.
Doing similar by-eye averages for the other NA leagues, I find the following improvements since the initial analysis.
As I expected, the greatest gains have been achieved by the players in the lower leagues. As a result, the SQ spread has become narrower. Previously, the difference in SQ between the average Bronze and the average GM player was about 40 points. Now it looks to be about 26 points.
For this reason, I think it may be necessary to develop a higher fidelity measure. When SQ was first proposed, it was an excellent metric for comparison between leagues and for tracking self-improvement. Now that the range of skill is narrowing, it may be worth looking for a way to increase its accuracy (see below).
2) In my initial analysis with 2084 games, I was not able to detect any significant effect of game duration or race on SQ. My later analysis suggested that there may be some disadvantage for Protoss, especially at the highest levels of play.
With many more games, it is now interesting to see what appears to be a trend towards lower SQ for longer games in these results, especially for Zerg, and especially for lower leagues. Terran also definitely appears to have some advantage over Protoss, and over Zerg for longer game durations.
Using these data, it should be possible to formulate an even more accurate version of SQ, possibly taking race and game duration into account. Some of the spread in the initial results by league may be mostly due to not accounting for these variables.
|
Not necessarily. There will always be a distribution which will result from a few things.
1. Interest in the game. I'm willing to suggest that those people who are more into Starcraft 2 (watch Day9, tournaments and so on) are likely to be more focused on "mechanics". A lot of the overlap you will see between leagues is due to this effect - people actively seeking promotion through macro. This is especially true of Terrans - FilterSC is more or less the standard terran macro build in lower leagues and it is extraordinarily SQ efficient. A terran following FilterSC standard 3 rax 1 fact 1 starport 50 SCVs with +1 can easily hit 75-85 SQ AND STILL LOSE because they either are too timid, attack late or various other reasons.
2. Percentage of games which last a certain time. In other analysis you note that the SQ of games has a tendency to drop off dramatically as the game lengthens. My question for you is, in your analysis, what percentage of those games by league got to that point? Remember that people don't really hit timings in Bronze League. They might be able to macro - they might even be able to put a huge number of units together but the moment they attack, their SQ will drop really fast.
I don't think you're ever going to get an ideal fit. That's why I suggested some iterative measures above relating to Production Quotients and so on. Harder to do, but maybe we'll get some more information out of Blizzard at some point
|
I'd also be careful about increasing Protoss spending seeing as only the Warpgate does not function in the same way as Terran buildings. You may need to add a Warpgate adjustment factor to your calculations based on the composition.
|
Terrans have the highest spending quotient, really now?
Lets look at random diamond replay: queued units on the production buildings.
Lets look at random master replay: queued units on the production buildings.
Lets look at random grandmaster replay: queued units on the production buildings.
Lets take a look at IMMvps replays, queued units on the production buildings, at all stages of the game.
I wonder why terrans have the highest spending quotient, really dont have a clue at all.
Oh right, macroing as terran works differently than macroing as protoss or zerg, of which both have to macro at the exact moment when they actually have their larvae/warpins available, while terran can lazily queue all of that stuff up.
You constantly produce as terran, thats just how it is.
|
|
|
|
|
|
EPT![[image loading]](http://i.imgur.com/Ryjn1.png)
