|
On December 23 2010 08:20 Jeffbelittle wrote:
Skrag: Thank you soo much for doing this. From what you've posted I'm doing some of my own mathematics to access different build orders. What can you conclude will be your "Go-To" opener as of right now.
Unless somebody can poke holes in the replays (if anybody can even download and watch them), I really think I'm going to go back to 11p18h on just about every map and every matchup, and 14h15p or 14h14p on shakuras and jungle basin.
Also: can you try 16 hatch 17 pool? I know it sounds soo silly but I do it against Zerg and it feels right..
No. Cause I'm 90% sure it will be worse than 16h15p because of the very late pool, and as I mentioned, it's really not easy for me to add new orders. Especially if all you tell me is "16 hatch 17 pool".
|
Invalid or Deleted File.
The key you provided for file download was invalid. This is usually caused because the file is no longer stored on Mediafire. This occurs when the file is removed by the originating user or Mediafire.
If you believe you have reached this page in error, please contact support.
|
On December 23 2010 08:43 ChickenLips wrote:
Invalid or Deleted File.
The key you provided for file download was invalid. This is usually caused because the file is no longer stored on Mediafire. This occurs when the file is removed by the originating user or Mediafire.
If you believe you have reached this page in error, please contact support.
You tried to download in like the 15 seconds it took me to reupload the files and update the links. lol
|
Same exact error still, with all 3 of them.
|
Ok, reps are not working for some reason.
I have no clue how to get it published so that people can watch reps generated from it.
The files should be there now though, but I did update the links, if you haven't refreshed. And I can click on the links and download the file. (and watch the replay, but the map is in my cache from testing it with the map editor)
Guess I could share the map itself, and you could run it from the map editor yourself if you really wanted to see what it was doing.
|
Great thread, I think the work you are doing here is both really helpful, and really awesome. A couple brief questions...
-- am I to understand that yoru most recent 11p18h tests are better than the ones shown in OP?
-- I know that the people yelling about ling/gas timings can get annoying, but I can see how the necessity to get gas might alter your results significantly if it ended up delaying a crucial queen for example, leading to a missed inject, and thus spiraling the build out of control. This is a problem that might be unique to a certain build (IE if we need gas by 18 food, we might already have our queen from 11p but have to delay our 14p queen even farther if we get gas, or cut drones) -- is this something you have considered abstractly or through testing?
|
The most recent tests have followed the builds I posted earlier, all of which include gas, ling speed, a pair of lings, a creep tumor, and a spine crawler.
And it turns out that 11p actually isn't nearly as poorly affected by adding those things as the other two builds, because it passes 13p15h (which was significantly ahead in the drone races), and narrows the gap to 14h15p significantly.
|
I've followed a few of these Zerg econ threads, and I've not seen any Pool-before-Overlord builds that perform. I'm curious if this is a matter of optimization, or if they really aren't competitive for anything but Zergling rushes. Regrettably, I don't have access to SC2 to check.
My calculations suggest these timings line up, and that there's no larva wasted after the 9 pool 8 drone. Would you please check if this is overambitious?
9 Pool
8 Drone
9 Drone
10 Overlord
10 Extractor Trick/Drone
11 Queen
|
11p only being 60ish minerals behind at 9 minutes? im definitely interested in that one 
|
Here's a link to the map:
http://www.mediafire.com/?r03jc75uw2bx2fg
To use it, load it in the map editor, make sure Use Fixed Random Seed is checked in preferences (if weird stuff happens, and drones don't look like they're doing what they should, then you've probably forgotten to check that box), and then Test Document.
The default build is 11p18h, but you can change that by either changing a line in the script, or by typing "13p15h" or "14h15p" (exactly like that) in the chat within the first 8 seconds after the map starts.
Replays get stored in Documents/Starcraft II/Replays/Unsaved, and the data logs get stored in Documents/Starcraft II/UserLogs/SkragAIBuildTester
The data is stored as two separate logs, as files named NameStats.txt and NameTimings.txt, where Name is the name of the build (13p15h, etc). The stats file is the data recorded every second, and the timings file contains start and completion times of all buildings and upgrades, and completion times of all units. (couldn't figure out an easy way to record the start time of units that come from larvae)
Once you've run it from the map editor, you can watch the replays by either just double-clicking them in an explorer window, or copying them to Documents/Starcraft II/Accounts/RandomNumbers/MoreRandomNumbers/Replays/Multiplayer
After an improvement to the settle-in times for 14h15p, 11p is closer to 100 minerals behind. Which is still really damn good IMO, and still better than 13p15h.
I've started generating graphs from the generated data comparing the three builds I've completed. I'll finish them later tonight and post them.
|
Awesome work Skrag. Is it possible to test the economics of 1 base builds? Let's say you'd want ling speed started and a handful of lings out before you make your expansion.
|
On December 23 2010 10:04 VoirDire wrote:
Awesome work Skrag. Is it possible to test the economics of 1 base builds? Let's say you'd want ling speed started and a handful of lings out before you make your expansion.
The map's posted, with a bunch of support logic for scripting build orders.
Knock yourself out.
|
Alright I've got the BO for 14gas/14pool, i.e. the probably most played Z build ever, getting real stuff (tm).
14 gas (at 100 minerals)
14 pool
15 OL
16 Ling (since you've decided to only make a pair of lings, I've allowed myself to sneak a drone in that is usually saved for lings)
17 Queen + speed (those and the lings happen at almost the same time) + pull drones off gas and put them back on at 4:20 which is the speedtiming for other builds (and surprisingly enough fits almost perfectly when you would put them back in a 'real' game)
19 Hatch
19 Queen (1 drone snuck in)
25 OL
25 Spine crawler
30 OL
36 OL
The timings are:
scout 3:00, ling 3:11, tumor 4:45, crawler 4:45, drones back on gas 4:20
your drone micro and AI testing will probably be able to make it even more efficient but this is an approximate BO. I'm really interested in how much the earlier speed costs compared to hatch first and 11p, if it turns out to be dramatically worse it would be a really strong argument to forsake this build (since there is almost no point to getting speed THAT early in 90% of games, yet people blindly do it since it's what has been done for months)
|
Ok, I've just watched the 11pool build and that was really amazing to see it reach 2 fully saturated bases with real stuff at 6:00, I'm pretty wowed :D
Also that extractor trick was probably the coolest one I've ever seen. Really nice work, Skrag.
|
On December 23 2010 10:48 ChickenLips wrote:
Ok, I've just watched the 11pool build and that was really amazing to see it reach 2 fully saturated bases with real stuff at 6:00, I'm pretty wowed :D
Also that extractor trick was probably the coolest one I've ever seen. Really nice work, Skrag.
I had to add the overlord/zergling stuff just to draw it out past 6 minutes. 9 minutes in and all the builds are pretty much hitting 100 supply on zerglings.
And yeah, I wish I could execute the extractor trick (or anything else for that matter) anywhere even remotely close to as cleanly. lol. You have to do like 4 things at pretty much exactly the same time, which is easy for a script that thinks in increments of 1/16 of a second, and that can do a lot of shit in that time increment, but I'm pretty sure it's not actually humanly possible.
Were you able to watch the replays or did you download the map?
|
CONCLUSIONS (added first as a tl;dr version of what's going to be a very long post):
Three builds (14h15p, 13p15h, and 11p18h) were put through AI-run tests that did more than just build drones. Each build was required to get gas at a reasonable time, one set of lings, ling speed, a creep tumor to connect the bases, and a spine crawler, with timings based on a replay that was posted in the thread where Idra executes a 14h14p build.
Obviously 14h15p is still the economic leader overall, but not by nearly as large a margin as expected.
11p18h seems to be a pretty big winner here compared to previous testing. Surprisingly, it seems to have actually absorbed the "reality injection" *better* than both 14h15p and 13p15h did, which didn't allow it to catch up completely, but closed the gap by quite a bit, and put it ahead of 13p15h, which was shown by previous testing to have the best economy of all the pool-first builds that were tested.
Specific notes about 13p15h:
13p15h seems to have a very awkward period from around 4:25 to 6:20 where it seems to be more resource-starved than the other 2 builds. During this time, it is forced to let significant numbers of larvae go unused, which probably accounts for a lot of the economic fallout, and might explain why a lot of people had never even heard of it as a reasonable opening until the whole "Economic 11pool" discussion started a few weeks ago, and 13p15h was offered as an economically superior pool-first build.
The larva advantage 13p15h has is interesting, but it's fairly small (probably somewhere between 1 and 2 larvae), and doesn't kick in until about 4:20, and frankly, the build doesn't gather enough resources to actually spend the extra larvae once you start doing things other than just building drones, so thes advantage doesn't seem to really be worth much.
I had started using 13p15h based on previous testing, but after this last set of tests, I'll definitely be switching back to 11p18h whenever I'm not going hatch-first, which IMO should still be preferred wherever you can pull it off.
Some graphs:
Resources collected:
![[image loading]](http://img696.imageshack.us/img696/9297/resourcegraph.png)
Total Larvae:
![[image loading]](http://img209.imageshack.us/img209/7852/larvagraph.png)
Resource difference from 14h15p
![[image loading]](http://img560.imageshack.us/img560/4183/resourcediffgraph.png)
Larva difference from 14h15p
![[image loading]](http://img593.imageshack.us/img593/5219/larvadiffgraph.png)
The spreadsheet I'm using (in .ods format, which is the default format for OpenOffice Calc), containing all the collected data, higher res graphs, and a few more graphs:
http://www.mediafire.com/file/bjr8anr5nvsehqg/AITests.ods
13h15p ends up being about 235 minerals behind, while 11p18h is only about 120 behind. Gas counts are practically identical for all 3 builds, given that I tried to construct the extractor at the same time in each one. There's never more than a difference of 4 gas, which is one drone trip worth.
The larva graphs are a lot more difficult to read usefully, but the total larva graph shows windows where one build can be ahead of the others due to inject timings. For example, 14h has a couple larva advantage between about 3:40 to 4:00, at which point 11p takes a small advantage until 4:20, when 13p takes the lead until about 4:50. Unfortunately, because these builds were executed by the AI at maximum possible speed, with perfect micro and accuracy, you can't just take times from the graph and say "at this time, I'll have a larva advantage". You'd have to get real replays of your own game, and see where the injects come in, because I guarantee the AI played more perfect than you.
14h and 13p seem to stay mostly in sync, with 13p having a small advantage on average, but the timing on injects for 11p give some pretty big swings back and forth, although appearing to be approximately equivalent overall to 14h, just with different timed surges from injects.
The larva difference graph is damn near impossible to draw useful conclusions from, because it spikes all over the place with larva injects. I tried to smooth it out by doing partial larvae from injects (counting an inject that was 20/40 done as 4 larva, for example), but there's a point where the inject is complete, but the larvae haven't appeared yet, that caused the same sort of spikes, and I didn't really come up with a good way of accounting for that, so I scratched that and just went back to calculations based on available larvae.
I'm going to try to add 16h15p next, and given how long it takes to tune each build, that will probably be the last one I add, even though I'm pretty curious now to see if 14p16h outperforms 13p in a "real" test. I'm pretty convinced it will do worse than 14h15p, but the results from 11p18h were extremely surprising and not what I expected to see, so I still think it's worth doing.
Some notes about how the tests were run:
I used the build orders that were posted in the original post of this thread, and modified them to add gas, a pair of lings, ling speed, a creep tumor, and a spine crawler, trying to time all of those things based on a posted replay where Idra performed a 14h14p.
Only 39 mineral mining drones were build, 19 for the main, 20 for the expansion. The main could only possibly support 20, but in one case where I accidentally mis-rallied one drone to the main giving it 20, the game ran past the 10 minute mark and there was *still* a bouncing drone at the main (over 4 minutes after the drone was actually added to the patch), so mining was only happening at 19 workers worth anyway. The expansion can technically support 22 workers, but there are 4 patches that are far enough away to support 3 workers, but close enough that it's actually pretty hard for 3 workers to settle into a routine. Putting 20 drones at the expansion seemed to be the "right" amount, where they would settle in relatively quickly most of the time (but I still saw cases where over 2 minutes went by before they settled in).
Drones from the main base were rallied to the main minerals until there were 16 drones there, and drones made at the expansion were always rallied to the expansion. After the main had 16 drones, any drones created there were sent to the expansion, except for the last 3, which were sent to the main giving it the full 19 workers, but delaying the settle-in penalties as long as possible.
Drones were microed as perfectly as possible, being sent to whichever mineral patch would cause the least amount of bouncing delays before they started actively working. Most of the time, there were no delays at all, but in some cases, especially as the number got closer to 16, I had to bounce a drone off a patch that had 2 workers on it already, so that it could bounce to a patch with only one worker and get into a routine on that patch.
So the final target was 39 workers mining minerals, 3 workers mining gas, 2 queens, scouting at about 3:00, ling speed, a tumor, and a crawler. Once that target was reached, zerglings were built to 52 supply, then 6 overlords were built, followed by 48 supply worth of zerglings. This was done just to give some time for all the workers to settle in, and see where things fell out, while also using all available larvae so that accurate larva counts could be generated.
13p and 14h both sent a scout drone at around the 3:00 mark, but 11p was allowed to build its lings earlier for scouting rather than pulling a drone, with the reasoning that a slight delay in the scouting was compensated for by having 2 ling scouts instead, which are faster and more generally useful.
11p was also allowed to delay the creep tumor by about 20 seconds in favor of a third inject, with the reasoning being that it has a lot more flexibility due to the earlier pool, and is actually less likely to be attacked early because the pool goes down so early.
|
Amazing job, Skrag! I downloaded the map and made it do 11p18h. Pretty neat to watch. The extractor trick, and the perfectly timed injects... a thing of beauty. Some things I noticed, though:
1) At 3:40 there are 154 minerals and 20/26 supply, so the second Queen could easily be started. However, the script waits 10 more seconds to start an Overlord at 3:50 (224 minerals), and then a Queen at 3:52. Surely that's not intentional? (Although I doubt the difference will be big.)
2) There are only 2 drones mining gas, is that intentional?
3) After the main has 17 drones mining minerals, all new drones are rallied to the expansion. Shouldn't that be done at 16 instead? (Maybe the 17th mineral-drone should have been the 3rd gas-drone?)
|
Hi there i love the ending of this many threads, finally the "testers" went thinking right for the 11pool, or pools first which favors the safe base build up.
i think they are just needed by the timings of the hatch-first defends..
i begin with sc2 late and the pressure of harrassment is simply annoying, but in fact the eco is not worth when the opponent gets more gas invested, and the zerg is everytime one or more steps behind.
something between hatch and pool first is difficult but only having no practise with it at all is not a true testing result..
when i practised with the 11pool, the way of saturating the main was totally different in some games, so i tested inbase hatching in some games, rushing for 2 worker on the near minerals and flood drones to thirty, adding a defendforce on the scouting information at 24 supply and going for a window to attack at 8-11 minutes, while expanding u perform an army trade, contain or a fast expansion punish, lairtech or what is really needed by scouting informations..
Building 6-12?drones with the 2 mainhatch larvae injects shortly before the expansion finishes and the drones off the distanced minerals will give you a quick gg or a dream macrogame..
a Clue was a hatchery trick on the lower choke to generate a forward creep Tumor, and instantly building a spinecrawler with the "trick" drone, scouting will make sure you can unroot him to position it before an attack come close
Please please read this statement not of the view, that this are noob or unexperienced rts mindgames, honestly
lets get back our lovely 11-pool Timings.
Mfg
|
On December 23 2010 19:19 Nolari wrote:Amazing job, Skrag! I downloaded the map and made it do 11p18h. Pretty neat to watch. The extractor trick, and the perfectly timed injects... a thing of beauty. Some things I noticed, though: 1) At 3:40 there are 154 minerals and 20/26 supply, so the second Queen could easily be started. However, the script waits 10 more seconds to start an Overlord at 3:50 (224 minerals), and then a Queen at 3:52. Surely that's not intentional? (Although I doubt the difference will be big.) 2) There are only 2 drones mining gas, is that intentional? 3) After the main has 17 drones mining minerals, all new drones are rallied to the expansion. Shouldn't that be done at 16 instead? (Maybe the 17th mineral-drone should have been the 3rd gas-drone?)
You forgot to check the fixed random seed box, didn't you.
|
On December 24 2010 04:46 Skrag wrote:
You forgot to check the fixed random seed box, didn't you.
Aww crap, you're right! Checking that fixed the gas mining.
However, the issue with making the second Queen remains. Just look at your own 11p18h spreadsheet:
3:40 - 154 minerals - 20/26 supply
...
3:49 - 229 minerals - 20/26 supply
(Overlord is started)
3:50 - 139 minerals - 20/26 supply
3:51 - 149 minerals - 20/26 supply
(2nd Queen is started)
3:52 - 4 minerals - 22/26 supply
Surely the Queen should be started at 3:40?
|
|
|
|
|
|
EPT
