On June 27 2011 10:32 lysergic wrote: •Half the time you face the Zerg AI, it does some sort of 1 base rush/cheese - there's not enough macro builds. I think you should remove the Drone Rush and 7 Pool Commanders. (Also, the frequency of BCs and Carriers are annoying too.)
Regardless of zerg... I would say that ALL the races have too many commanders. Commanders like protoss cannon rush is TERRIBLE because AIs do not know how to cannon rush. The list of random commanders which are chosen automatically should be made much more lean.
If people want to choose a specific commander that it doesn't normally choose from, that's fine, but many of these commanders should NEVER execute automatically. At the minimum they should execute much less than other commanders. More is not always better. Quantity ≠ quality.
One of the biggest issues with commanders though is with regards to not accounting for the opponent's race & build-order. commanders should be selected based off the opponents race (Don't cannon rush a terran, although like i said before, they shouldn't cannon rush at all). For an "AI" to actually be an AI, it needs to have intelligence. Intelligence consists of the knowledge of countering different things, and the ability to see such things (feedback).
I just played another game against it and it moves out with 2 phoenix, a sentry (or two) and 2 zealot.
As if that wasn't bad enough, it also didn't even used graviton beam except for once later on, which was on my LARVA.
As if that wasn't bad enough, it seems like most builds the AI runs it gets a robotics facility for some stupid reason... (apparently it transitions to immortal later which I guess is fine, but robo bay was a bit early to get) it doesn't even use the robo bay I find. Both this phoenix attack, as well as "2 gate expand" get a robo bay and don't even use it. It's especially strange with the "2 gate expand" build which goes 2 gate robo, then might attack or just sit on it's ass, THEN expand at like the 8 minute mark (really late for a 2 gate expand).
The 2 gate expand build needs to be modified, and the weird-ass phoenix build with robo should probably be removed (assuming there isn't a build that gets voidray+phoenix instead already, otherwise switch it to that).
Overall this is the problem with most of the commanders... It's just quantity over quality. There may be a bunch of commanders but they run trashy builds and/or are not executed well (I saw one game where toss got 3 bases and intentionally decided not to mine gas, preferring mostly mass zealot (which obviously got stomped).
yea terran is OK but zerg ai is terrible. Half the time it does a 1 base all-in with no transition. sometimes itll make 2 spawning pools and an inbase hatch...
@Xapti: The programmer take a good jobs in making AI that far better than the original Blizzard AI. If you find that you are a real strategist, why don't you make and AI by yourself if you thinking of making AI is easy. I understand it take him much time to develop this excellent program, but he need time to do other thing. Instead of complaining , you should support him. At least by action, there is a 9 combination in SC2 Z/P/T vs Z/P/T, so for each combination, let say 6 strategies, then 54 strategies is required. Do you have time to making this list by yourself and send to him ? No, I assume.
@Ptanhkhoa: Thanks for the great AI, but could you set the choosing time a little longer, for me 10s is too short. 15s is good enough.
On June 28 2011 04:41 lysergic wrote: Yeah I was thinking the Select Commander might not of been working for me because I wasn't doing it early enough. Having to do it within the first 10 game seconds (7 real seconds) isn't enough time to do all this:
-Make a Drone -Split workers -Overlord scout -Control Group Hatchery/Overlord -Set rally points -Read the list of Commanders "-lp" -Select Commander "-se x y"
That's why I suggested the OP to implement a map restart when you select a Commander. This would make the map more like YABOT (you can choose a new build for the AI and restart the map). I think the game clock should reset too.
It'll give you enough time to choose a build. It'll let you restart the map whenever (if you're about to win or lose, if you mess up, or if you don't like the Commander picked, etc).
I can do this. but it will mess up with the GT opening code. So I can increase the time to 15 seconds. I think that a reasonable time before the first drone is build, otherwise it will messed up with some build order. Also, when you time "-lp" at the begining, the gamespeed will reset to "Slower" give you more free time.
I update to GT AI 0.88 version which give more detail descriptions, also adding "-nc" (no cheese) command that cross out the cheese strategy from Commander list. Thanks for uploading map on the US Server ^_^.
@Xapti: Sorry, my time to develop on this project is very limited ^^. Your idea is very good, but thinking that instead of developing the strategy by my self, i will make the AI that can actually learn your BO xD, so the more you play, the AI will be better.
@ thisisSSK : Zerg AI has balance strategy between Rush and Expand, so do some more match, or in the GT AI 0.88 version, you can type "-nc" to cross out the "cheese" strategy (6pool, 7pool,... ) , so i will use more standard "2 base" opening.
@Ptanhkhoa: Try new GT AI 0.88, just increase the time to 15 seconds. Also you should using with "-lz", "-lp", "-lt" command to set the game speed to "Slower" and give you more time. After typing "-select x y" the game will be "Faster" again.
On June 28 2011 14:00 alonestar_ak wrote: @Xapti: The programmer take a good jobs in making AI that far better than the original Blizzard AI. If you find that you are a real strategist, why don't you make and AI by yourself if you thinking of making AI is easy.
First of all, if I knew 1 year ago that in 1 years time, starcraft 2 would have only one (and a half) melee AI (as far as I know at least), made by mostly just 1 person with quite limited time, I would have actually dedicated a lot of time to learn about it, if not work on it myself. However, I did not expect this future, and actions are best done according to predictions and plans. The major lack of SC2 AI developers has utterly surprised me. I thought many Starcraft 1 people would move to it, as well as a whole bunch of other people who want to innovate, help the SC community, learn things, program, get famous, etc.
Secondly, Criticism is one of the the best forms of support, as that is how it can be improved. I specifically posted in a previous post that I care about this AI and want it to get better (among other things).
Lastly, Blizzard's Very Hard AI is better than some of these commanders (Blizzard AI is also programmed based off no strong metagame, and before a bunch of nerfs and buffs, such as reaper and roach). One of the biggest issues is that the AI has too many bad build orders, or just completely bad builds/strategies (cannon rush).
I wonder if you can implement MMM kiting? Most of the time the Terran army just splits up individually as the zealots chase marauders down, reducing dps time and thus gets annihilated. If the terran bioball would just clump up and stim+run the zealots will never be able to catch up.
On June 29 2011 22:17 LookNaph wrote: I wonder if you can implement MMM kiting? Most of the time the Terran army just splits up individually as the zealots chase marauders down, reducing dps time and thus gets annihilated. If the terran bioball would just clump up and stim+run the zealots will never be able to catch up.
While I have thought of the same thing, I think it might be rather difficult to implement.
The easier thing to deal with that you somewhat bring up though, is changing shooting time before moving for stutter-stepping. Marines have an extremely fast move delay after attacking, which makes them very effective at stutter-stepping, but the AI doesn't seem to take most advantage of it (although I can't specifically remember for certain).
Oh, and a serous problem/bug I noticed, is at least sometimes the AI (at least terran) will not rebuild their expansion if it got killed. In fact after taking their 4th position expo around or after the 3rd died, they never took any additional base, not the 3rd, not the gold, or any other, even when in a situation to do so, and in a situation with low money. Another time I saw protoss build 2 nexus at 3rd expansion for some reason... I'm thinking it might have been intended for the gold, but it was on the wrong side of the cliffs .
I find the AI is very bad with managing their money that way... they just try to expand a lot while pumping units, never being more defensive or careful if they are about to run out of money, granted I don't think this sort of thing is much of a priority, mostly just rebuilding expos is important.
Another small issue is that AI attacks neural parasited units a bit. While that might not be an issue in itself, when it wears off, they will continue attacking it, even chasing it. Not sure if it eventually stops, or if it kills the unit entirely.
On June 28 2011 14:00 alonestar_ak wrote: @Xapti: The programmer take a good jobs in making AI that far better than the original Blizzard AI. If you find that you are a real strategist, why don't you make and AI by yourself if you thinking of making AI is easy. I understand it take him much time to develop this excellent program, but he need time to do other thing. Instead of complaining , you should support him. At least by action, there is a 9 combination in SC2 Z/P/T vs Z/P/T, so for each combination, let say 6 strategies, then 54 strategies is required. Do you have time to making this list by yourself and send to him ? No, I assume.
His post is filled with more constructive critisicm than yours. If the OP wants feedback, this post would be extremely useful.
To the OP, Good work but I am not that impressed so far. I will write up a more descriptive post when I get home.
Awesome work on this, had some fun messing around with a few new builds though I'll confess I'm far too slow to manage a decent commander selection at the beginning. I find some AI builds provide for far better games than others for whatever reasons, like with some openings the AI follows up properly and others it doesn't or something ^^
It's got some sick burrowed Roach micro lol, watching a replay and it's doing 1.3k apm during a big battle heh.
There are many things need to improve, but first, i modify the AI Micro a little bit: make all unit move to the same direction, using vector system instead of spiting unit in all direction. Also adding Order Build Recording, so you can make AI do better Opening. After implement the build order, i think of making AI doing general AI scripts that can counter every race, but it will take more time and tester for developing
On July 01 2011 08:24 ptanhkhoa wrote: There are many things need to improve, but first, i modify the AI Micro a little bit: make all unit move to the same direction, using vector system instead of spiting unit in all direction. Also adding Order Build Recording, so you can make AI do better Opening. After implement the build order, i think of making AI doing general AI scripts that can counter every race, but it will take more time and tester for developing
Excellent! This AI is getting more and more advanced and better too. Couple flaws I'd still like to point out: -The Build Recording system is awesome, but it also means that you have to play other races and then perfectly execute the build you want the cpu to learn. This sounds like an incredible hassle to me, so I suggest some kind of way to edit/import BO's like YABOT. Honestly I like this GTAI much better than YABOT because its so much more advanced ^_^
-This is just a minor critique: the micro of the computer might be TOO good. I don't know if theres a fix for this but if so, it might be better to actually lower the micro level because it becomes too inhuman and unrealistic.
-I haven't tried it on the new version but from what I remember the -qr and especially the -re were broken on some maps
Thats it! Again, I commend you on your hard work and dedication Please keep at it!
2) like lysergic already said there are to many strange commanders in the pool. i think its better to pick a random commander out of a smaller pool and let the other in for direct selection.
Tbh, this isn't very useful for practice purposes. The fact that the AI has an income cheat ends up encouraging you to over produce units. I've picked up this bad habit back when I used to practice against the regular insane AI.
@AndAgain: Ofcourse, that is your choice for more challenge you play at harder difficulty. But from Medium below, AI have no minerals cheating. @ thisisSSK: Well, i will think of adding a manually adding BO, beside playing and Recording. @ skeldark: yeah, when I think I get up the data recording from pro tester, I will replace the old BO with the new professional BO.