|
Hi TL, this is my first post to the forums. I'm an old-time noncompetitive SC1 player who has been actively following SC2. I often see people discussing the ideal saturation points of mineral lines. It's common knowledge that you want at least two workers per mineral patch, and no more than three workers than mineral patch unless you're planning on transferring soon or using workers for other things in a crisis (bring extra SCVs on your thor push :D ).
Anyway, I've done some experimental testing of worker saturation for (assuming 8 mineral patches) 1 through 24 probes mining, and compiled the results in a little report that took about two or three hours to get together (keeping myself sharp for report-writing and uncertainty analysis). I used qxc's build tester map (which is awesome, thanks qxc!) and protoss for all tests. There's a TL;DR at the end, and pretty graphs. The curve fits are just that: curve fits to experimental data. Take them with a grain of salt. I probably kept one too many significant decimal points which are not really significant due to time and mineral resolutions.
In general the takeaways are: Avoid having three workers per mineral patch until you have two workers at each mineral patch. If you have a patch with three miners and others with one miner (which is possible!), that mineral patch will quickly deplete relative to the other patches, causing an oversaturation of the other mineral patches leading to wasted worker efficiency. After two workers per mineral patch, each additional worker suffers significant diminishing returns in the relative increase in the next worker. Saturation really pays off when you're planning on expanding, since your expansion and main will then drop off into the linear region of worker income efficiency and you can just build workers from there.
No groundbreaking revolutions; just an additional source of information for people to look at if they're curious of one set of experimental results to try to quantify worker efficiency. It's evident though at higher numbers of saturation, the AI can have trouble finding minerals and oversaturation may even be a bit detrimental at exceedingly high numbers.
Anyways: The paper (it's not that long; you can skip to graphs and tl;dr at end!) I prepared is here: http://docs.google.com/fileview?id=0B3X9N0NR7fRtYzAzN2RlZTAtZDlkNS00M2Y2LTlkMGQtZGZkM2M5ZTA4Mjc0&hl=en
I don't know if google-docs has a bandwith limit or if anyone is even interested, but hopefully this can help someone curious about the results.
Also, here's a graph that you can find better explained in the report:
![[image loading]](http://i.imgur.com/QY2mr.jpg)
PS. I know there's at least one spelling or grammar error, the focus is the content and data analysis. Also, the (95%) thing next to the calibration means 'to 95% confidence on a student's t distribution'. I'm not a statistician, just an engineering u-grad with a small uncertainty bkg.
<3
|
excellent contribution, most peoples first post isn't of this quality.
|
impressive efforts in the write-ups...
but i find your conclusions to be lacking or the amount of datas u've gathered, until we reach 2 per mineral patch saturations the mineral gathering rate pretty much goes up linearly from the data's u've presented, but i think what's actually important to know about is how much increase in mineral gathering rate on the 17th probe onwards, in fact would've been alot more helpful if u've calculated the gradient of the graph u've presented, one for the linear line for the first 16th probe and then for the line of sub-sequent probe until maximum saturation is reached...
|
Alright, a LaTeX paper on mining efficiency, been waiting for this. Great work!
From your raw data, would it be possible to calculate how much the "pay-off" time would increase for saturation higher than 2/minerals? Because that would be useful to know when trying to do a timing push.
So, when mining 3 drones/patch, some fumble and some stay cool? Any idea if this is correlated to patch distance? If it is, optimal could be ~ 20. 2 on each patch and 3 on the non-fumbling patches.
Now I build 16 + a few (3-4) spare ones which can be used to build defense or tech without giving up linear saturation.
|
Shizuru, look at the graph? Eyeballing it makes me think that the last 8 bring in half the increase that each of the first 16 did.
|
Nice graphs. It's always good to have one's intuition confirmed by actual tests. I presume you made the graphs with gnuplot?
Now for some aesthetic comments: A) Is that \documentclass[a4wide]{article} with \linespread{1.3} [1]? I thought lshort.pdf discouraged making the margins smaller because it reduces readability. B) In Equation 1 I'd use \cdot instead of *.
[1]: It saddens me a bit that I can actually make an educated guess about things like this.
|
Impressive, thanks for the information! I'll have to seriously change my drone production now that I've seen this.
|
I read the document, and I understand that using an averaging approach it won't answer the following question directly.
But since you pointed out not to do 3 workers on a patch if there is a patch with 1 worker, is it ever better to put 2 workers on a (perhaps closer) patch than to put 1 worker on a farther patch?
|
|
|
|
|
Good Job, I did a similar work earlier (sorry did not publish it because i sort of forgot about it) and got a similar result, how ever i recalculated the result to mining/min*#probes*Mineral patch (that is according to if you have 1, 2, 3 or more probes /mineral patch) and got the following result:
![[image loading]](http://dl.dropbox.com/u/3351176/WorkerMiningGraf.jpg)
I'm sorry i didn't have the time to correct the graphs so they had the same time spawn, but i hope you will get the big picture anyway.
The tests was also done with Protoss but on blistering sand against computer, and the time for each set of workers was at least 100sec and the times was started after the workers had been split.
the first graph shows the mean income / sec when you have 1,2,3,4,5 and 6 workers/mineral patch, and as it was tested with the standard 8 patches the unit count is 8,16,24,32,40,48.
at 1 probe /patch it will harvest 42 mineral/min and probe (335 minerals total/min) 100% of M.cap.* at 2 probe /patch it will harvest 40 mineral/min and probe (639 minerals total/min) 95% of M.cap.* at 3 probe /patch it will harvest 33 mineral/min and probe (791 minerals total/min) 79% of M.cap.* at 4 probe /patch it will harvest 25 mineral/min and probe (811 minerals total/min) 60% of M.cap.* at 5 probe /patch it will harvest 21 mineral/min and probe (835 minerals total/min) 50% of M.cap.* at 6 probe /patch it will harvest 18 mineral/min and probe (839 minerals total/min) 43% of M.cap.*
*M.cap. = % of maximum capacity
|
On June 03 2010 17:07 artanis2 wrote: Shizuru, look at the graph? Eyeballing it makes me think that the last 8 bring in half the increase that each of the first 16 did.
beats the whole purpose of a technical paper if we have to approximate visually by the graph instead of getting a concrete value out of it don't you think?
|
What exactly is the point of the curve fitting? I mean, we already know from the other 20 threads on the topic that the last 8 workers are far less efficient than the first 16. What's special about the last 8 being fit by a cubic?
|
Thank you very much for your article. It's not tremendously enlightening (and it shouldn't be), but there is a real gem in having most people's suspicions corroborated by hard data.
I hope you'll produce more in the future!
|
Could you elaborate on your point regarding avoiding 3 harvesters per node where others have only 1? I imagine that the AI would motivate harvesters to end up at empty nodes rather than waiting for the assigned node to open up.
|
On June 03 2010 23:01 PJA wrote: What exactly is the point of the curve fitting? I mean, we already know from the other 20 threads on the topic that the last 8 workers are far less efficient than the first 16. What's special about the last 8 being fit by a cubic?
The point of the curve fitting for the first set of workers (linear region) was to give a concrete experimental calibration to what we already know (and even a little bit of uncertainty to show what we don't know!). The cubic fit is just if you wanted to look at the relative difference between the linear region of worker efficiency compared to the 2-3/patch region. Looking at the difference between the two fits is an interesting thing also if anyone was interested, since that shows what you're trading off by saturating instead of having an expansion and main in the linear region (but I didn't do it). Nothing special at all whatsoever about a cubic. It just fits the data nicely and can model the results. I literally just noticed an inflection point and said: "Well it's not linear, it's not parabolic. Aha, look at that. Mmph. Cubic fit look good. Ungh. Cubic!"
Infenwe; I'm newer to LaTeX, and the fact that I put it in this form is basically just because I'm trying to get myself used to it. I'll be sure to take your advice when making papers for school and read over lshort (I just skimmed it originally) :D.
igotmyown: That's an interesting question. Reasonably it seems like 2 on a closer patch would be better when you're starting off in workers. I think that's probably the case. Again, I completely neglected the tiny differences between near and far mineral patches for the first 16 workers by averaging and just saying "aha, it looks linear as expected". In reality there's probably little ways like this to sneak an extra 20 or 25 minerals out while you're in the early stages of the game by choosing carefully and doing different subtle things with workers that ordinarily might be standing around for a few secs between buildings or what have you. Hell, that might even be the difference between a stalker and a zealot earlier on. I haven't collected any data on it though.
LordOFswe: Similar results are good, and lends to the idea that we're both on the right track, as well as the community of starcraft players in general; which is good. Your graphs show how additional workers beyond 3 per patch DO increase mineral yield over a given time (if only slightly), which I completely didn't consider looking at. All the more reason to never stop producing workers (especially when looking at transferring!).
Makh: Yeah. Try it yourself, or look at a replay from a recent game and you might be able to see it. I think the best way to do it is to get all patches having two workers/patch except for one patch with one worker on it. Rally your nexus/hatch/cc to a mineral patch with two workers and watch closely. If it doesn't work the first time, try it one or two more times and you'll probably see it (I did at least); three workers will contently mine from one mineral patch WITHOUT SEARCHING for an additional patch to mine from. It's like they'll each delay mining and just sit there in queue to mine from it. More often though yes, the AI will likely try to find an empty node. Maybe it's just the conditions under which I watched the probes, but I'm relatively certain I've seen it during normal ladder games (as I go watch a replay with my curiosity piqued to make sure).
also; thanks everyone for positive comments <3
|
WOW very nice job, happy that you did this.
|
On June 04 2010 02:19 Cheticus wrote: Your graphs show how additional workers beyond 3 per patch DO increase mineral yield over a given time (if only slightly), which I completely didn't consider looking at. All the more reason to never stop producing workers (especially when looking at transferring!).
Hmmmm, well i would say that it proves that you DON'T want more then 3 workers/patch, as if you look on the extra mineral you got, that is an increase of 50 minerals/min if you double the amount of workers from 3/patch to 6/patch, on 8 patches it is 24 workers and as they 24 workers will earn you an additional 50 minerals/min it will take them 24 minutes to pay for them self. And if you go from 3/patch to 4/patch the extra probes will earn you one single extra mineral per minute and patch (3x33=99 and 4x25=100). So you should only produce workers over 3/patch if you are planing for an expansion.
|
On June 04 2010 02:48 LordOFswe wrote:Show nested quote +On June 04 2010 02:19 Cheticus wrote: Your graphs show how additional workers beyond 3 per patch DO increase mineral yield over a given time (if only slightly), which I completely didn't consider looking at. All the more reason to never stop producing workers (especially when looking at transferring!). Hmmmm, well i would say that it proves that you DON'T want more then 3 workers/patch, as if you look on the extra mineral you got, that is an increase of 50 minerals/min if you double the amount of workers from 3/patch to 6/patch, on 8 patches it is 24 workers and as they 24 workers will earn you an additional 50 minerals/min it will take them 24 minutes to pay for them self. And if you go from 3/patch to 4/patch the extra probes will earn you one single extra mineral per minute and patch (3x33=99 and 4x25=100). So you should only produce workers over 3/patch if you are planing for an expansion.
I would even think that you probably shouldn't build any extra workers until you begin the expansion. Then build like 4 workers while the expansion is going up take another 8 probes to the expansion to start with about 12 probes at expansion then start macroing up again. It seems like it would be better to have spent those minerals on extra units
|
What matters to a real game is the incremental income per worker. If the first probe harvests 40 minerals/min, how many do you get from probe #17, 18, etc?
As far as the paper goes, your curve fittings make no sense. The linear portion of the curve has a non-zero intercept. At 0 workers you lose minerals, and the 1st worker is less effective than the 2nd worker. The cubic (why cubic?) portion of the curve eventually turns downward, meaning that if you extrapolate to 100 workers, you again lose minerals. You can use a spreadsheet to fit any curve to any data set. It's only a valid technique if your curve has some basis in reality.
The most mathematically sound approach is to use intercepts and asymptotes that make logical sense. That is, you need a curve that starts at 0 income for 0 workers and eventually levels off at a fixed income with infinite workers.
One logical approach would be a two-part linear-hyperbolic curve: W<16: I = a*W W>=16: I = (b * W) / (c + W)
This function starts at 0, and trends toward b as W goes to infinity. You could use a more sophisticated curve such as a logistic function for W>=16, but it would be a similar concept.
Ideally, you would also want to constrain the second curve to keep a smooth slope.
|
|
|
|
|
|