To test the muta's ability to get the fuck out of a shitstorm, I spawned scourges on all sides of the mutas so they would be surrounded. The anti scourage micro still needs a bit of work though.
12 Mutas vs 5 archons
Mutas still will get hit because they're only 32 range shorter than an archon, and so when archon and a muta walk towards each other at max speed, that 32 range will get exceeded when muta turn around to run away because the muta will lose some speed when turning.
Teamliquid AI Team owns.
EDIT: NVM it is actually 7 archons.
This one seems a bit more boring I guess because you can't see that much action in one screen. But I wanted to see how the wraiths fared on flat ground where hydras can chase them around.
Impressive. Now create an AI what would use muta to form a circle around a sair so it can attack only one muta at a time and you have +40% winrate against FE
Not much macro with 35k apm tbh. Not very impressive against auto attacking archons, should try against a human... or make a archon micro and see who wins :/
o.O Very impressive. I remember long long ago playing special AI and they had terrible micro. Can these things do that in an actual game, or are they only for specifically made situations like that?
Yeah I don't think the team will bust my balls for this because I took off all the debugging markers before filming, so this just reveals very very cool micro rather than anything important.
On November 27 2009 14:36 Count9 wrote: Not much macro with 35k apm tbh. Not very impressive against auto attacking archons, should try against a human... or make a archon micro and see who wins :/
Do you know what it takes to construct AI like that?
Wow...that's pretty impressive. I wonder if there's any way to make them not lose speed while turning and keep them moving at top speed...then they would really be pretty unstoppable.
On November 27 2009 14:56 lazz wrote: more proof that with perfect mechanics zerg would be the best race
Actually, wraiths would be better because their range is longer. With wraiths I could get 0 damage taken at all. If anyone wants I can upload the video with wraiths.
On November 27 2009 14:36 Count9 wrote: Not much macro with 35k apm tbh. Not very impressive against auto attacking archons, should try against a human... or make a archon micro and see who wins :/
Do you know what it takes to construct AI like that?
Not impressive to watch, not not impressive to make.
On November 27 2009 14:56 lazz wrote: more proof that with perfect mechanics zerg would be the best race
Actually, wraiths would be better because their range is longer. With wraiths I could get 0 damage taken at all. If anyone wants I can upload the video with wraiths.
uhhhh we sort of want you to upload every single video you can
On November 27 2009 14:56 lazz wrote: more proof that with perfect mechanics zerg would be the best race
Actually, wraiths would be better because their range is longer. With wraiths I could get 0 damage taken at all. If anyone wants I can upload the video with wraiths.
Would it be too much to ask how complicated the algorithm was to implement? High level it seems straightforward (switches between getting to minimum distance to attacking closest target), but I'm curious since alot of things turn out more complicated than first thought.
Also I'm curious about the capabilities; is this a "kill everything" alogrithm or can you specify targets? I'm sure this takes less damage than clumping, but is it faster? I think it'd be interesting if you had clumping too had had them battle.
Wouldn't you be able to get the Muta micro perfect if you stopped them from moving towards Archons, when the Archons are moving at Max speed? With the chaos of the micro you have it should be able to still be effective.
On November 27 2009 14:56 lazz wrote: more proof that with perfect mechanics zerg would be the best race
Actually, wraiths would be better because their range is longer. With wraiths I could get 0 damage taken at all. If anyone wants I can upload the video with wraiths.
On November 27 2009 14:36 Count9 wrote: Not much macro with 35k apm tbh. Not very impressive against auto attacking archons, should try against a human... or make a archon micro and see who wins :/
Do you know what it takes to construct AI like that?
Not impressive to watch, not not impressive to make.
If you don't, then you can't appreciate it. Like saying Cubism, or surrealism is bad art because it's just a bunch of random colors, and thus not impressive to make.
On November 27 2009 14:36 Count9 wrote: Not much macro with 35k apm tbh. Not very impressive against auto attacking archons, should try against a human... or make a archon micro and see who wins :/
Macro? What does Macro have to do with what this guy is trying to do. By this guy i mean the creator, he made an AI to micro not to macro...
On November 27 2009 14:36 Count9 wrote: Not much macro with 35k apm tbh. Not very impressive against auto attacking archons, should try against a human... or make a archon micro and see who wins :/
Not enough stupid with 157 characters even do you really tried hard.
On November 27 2009 15:13 AssuredVacancy wrote: Sit tight I'm recording the video of 12 wraiths vs 4 control group of hydras.
holy shit is that even possible? the hydras range upgraded ? i had no idea wriaths outranged hydras
Not range upgraded, without range upgrade hydras have 1 less range than wraiths. Doesn't look as fun because they're getting chased all over the map. With cliffs to hide it's way to easy to kill the hydras.
Not knocking you work - this is incredibly impressive, but I think I noticed a way to improve it.
It looks like the units retreat until they are ready to attack again, rather than timing their movement towards an enemy unit to coincide when they are ready to attack. If you did this instead (not sure how), you may be able to cut down the time required to kill this shit by up to half.....
Cool and awesome work! I like the videos. Btw, for big ones like the wraith vs hydras, you could probably use resolution expander hack to view the whole action.
On November 27 2009 15:48 lMPERVlOUS wrote: Not knocking you work - this is incredibly impressive, but I think I noticed a way to improve it.
It looks like the units retreat until they are ready to attack again, rather than timing their movement towards an enemy unit to coincide when they are ready to attack. If you did this instead (not sure how), you may be able to cut down the time required to kill this shit by up to half.....
BTW - FUCKING INCREDIBLE!
You mean like bouncing back and forth between units as opposed to hitting one, retreating and coming back?
That sort of thing would definitely do damage faster, but also make the units more susceptible to taking damage, because they are increasing the amount of time they spend charging at enemy units. Probably you'd still be able to kill all them, but end up losing a wraith/muta or 2, as opposed to not losing a single one.
On November 27 2009 15:48 lMPERVlOUS wrote: Not knocking you work - this is incredibly impressive, but I think I noticed a way to improve it.
It looks like the units retreat until they are ready to attack again, rather than timing their movement towards an enemy unit to coincide when they are ready to attack. If you did this instead (not sure how), you may be able to cut down the time required to kill this shit by up to half.....
BTW - FUCKING INCREDIBLE!
That seems like a good idea at first, but the thing is that you don't know what direction THEY will move in. If they come straight at you when you go back to attack they will meet you when your CD is not over, thus fucking you over because you gotta retreat and waste more time.
Of course I could make it the worst everytime and ASSUME they will come straight at me, that will improve it by a very small margin though.
On November 27 2009 15:48 lMPERVlOUS wrote: Not knocking you work - this is incredibly impressive, but I think I noticed a way to improve it.
It looks like the units retreat until they are ready to attack again, rather than timing their movement towards an enemy unit to coincide when they are ready to attack. If you did this instead (not sure how), you may be able to cut down the time required to kill this shit by up to half.....
BTW - FUCKING INCREDIBLE!
You mean like bouncing back and forth between units as opposed to hitting one, retreating and coming back?
That sort of thing would definitely do damage faster, but also make the units more susceptible to taking damage, because they are increasing the amount of time they spend charging at enemy units. Probably you'd still be able to kill all them, but end up losing a wraith/muta or 2, as opposed to not losing a single one.
That makes no sense... it might increase the time the unit is headed towards an enemy unit but it won't widen the window that the enemy unit has to shoot the attacking unit. So it shouldn't affect damage taken.
Haha oh god that second vid. 2 port wraith would be every zerg's nightmare, if it isn't already. Still, I find it incredible that you guys can do something like this, mucho props.
On November 27 2009 15:48 lMPERVlOUS wrote: Not knocking you work - this is incredibly impressive, but I think I noticed a way to improve it.
It looks like the units retreat until they are ready to attack again, rather than timing their movement towards an enemy unit to coincide when they are ready to attack. If you did this instead (not sure how), you may be able to cut down the time required to kill this shit by up to half.....
BTW - FUCKING INCREDIBLE!
You mean like bouncing back and forth between units as opposed to hitting one, retreating and coming back?
That sort of thing would definitely do damage faster, but also make the units more susceptible to taking damage, because they are increasing the amount of time they spend charging at enemy units. Probably you'd still be able to kill all them, but end up losing a wraith/muta or 2, as opposed to not losing a single one.
That makes no sense... it might increase the time the unit is headed towards an enemy unit but it won't widen the window that the enemy unit has to shoot the attacking unit. So it shouldn't affect damage taken.
Notice that the wraiths did end up taking damage from time to time? If you're heading towards the enemy more often, obviously you're going to end up taking damage more often. But that wasn't the point I intended, I meant more like getting sandwiched between units because you're constantly going between them, and the distance is going to decrease as they move in to attack you as you run.
edit : I'll try to clarify what I meant.
A x B
Say you're x, and alternating between attacking A and B. When you attack A and run, it moves in and chases you when you go to attack B. Then when you attack and run from B, A ends up being closer, and eventually the distance between them decreases to where it becomes impossible to micro and you just get sandwiched between them. I hope that makes a little more sense, the wording in my initial post didn't quite get across the idea I had in my head.
i think it would be interesting to see a news post introducing the members of the tl.net ai team and their efforts to us a whole. who are they, what have they done, what are the working on? maybe some way we can follow their progress in a less sporadic way? get some publicity! =)
looks like it's going to be a bunch of people with high lvl knowledge of Starcraft (us) against a bunch of people with advanced programming knowledge. Except... we apperantly have a bunch of people with programming knowledge as well. GL TL!
looks like it's going to be a bunch of people with high lvl knowledge of Starcraft (us) against a bunch of people with advanced programming knowledge. Except... we apperantly have a bunch of people with programming knowledge as well. GL TL!
Wow, that's some heavy company there. Best of luck to TL.
On November 27 2009 15:48 lMPERVlOUS wrote: Not knocking you work - this is incredibly impressive, but I think I noticed a way to improve it.
It looks like the units retreat until they are ready to attack again, rather than timing their movement towards an enemy unit to coincide when they are ready to attack. If you did this instead (not sure how), you may be able to cut down the time required to kill this shit by up to half.....
BTW - FUCKING INCREDIBLE!
That seems like a good idea at first, but the thing is that you don't know what direction THEY will move in. If they come straight at you when you go back to attack they will meet you when your CD is not over, thus fucking you over because you gotta retreat and waste more time.
Of course I could make it the worst everytime and ASSUME they will come straight at me, that will improve it by a very small margin though.
This is the part that I am unsure about. I mean, if the opposing unit comes straight at the muta (for simplicity, I'll use mutas for the example, but it could go for any unit), you could back up a little bit, then go at them again, timing it to happen just after the cooldown finishes. If they are moving straight backwards, you follow them, catching up when the cooldown finishes, or as soon as possible.
What I'm suggesting is to basically stay just outside of range until the cooldown finishes, at which time you move into range, attack, then fall back again. But instead of falling a large distance away, you fall back only a small amount.
I am going to have nightmares with that wraith video.
EDIT: This actually reminds me of interceptor movement, now that I'm watching it again. I remember when I used to mod SC I'd switch up the interceptor sprites to wraiths or scouts and watch them zoom around dodging hits. Wonder if that's something useful to look into.
That is, of course, reeeeeeeeeeeeally impressive. Nice job you guys.
It's funny to me that people are suggesting usual build orders to take advantage of that unit control. The whole point is that having unit control like this completely redefines the dynamics of the game. Air units become a lot more powerful, or speedy units in general. The subsequent AIs after the first generation ones at the tournament will incorporate the new AI metagame dynamics and games will look nothing like pro matches. To reiterate a previous post, who is the AI team? I don't mind dying to find out. : )
On November 27 2009 14:36 Count9 wrote: Not much macro with 35k apm tbh. Not very impressive against auto attacking archons, should try against a human... or make a archon micro and see who wins :/
Do you know what it takes to construct AI like that?
Archon? Easy algorithm for single-unit ai... If Muta in range, attack -else identify closest muta, walk toward closest muta. (this part is tricky, because if they're abusing -cliff, you may not be able to find a way to click past a muta for ideal pathing.)
An optimized "groupthink" Ai, however would be much more difficult, as there would then have to be a fast weighting calculation between nearest muta and nearest archon... A preliminary algorithm using a simple "recentering" weighting algorithm using nearest neighbor would be something like: If muta in range, attack --else, ----identify closest friendly archon, weight of -Weight* vector difference. Ideal muta to pursue is the ----muta with the closest "weighted" distance after subtracting the vector difference of the friendly archon. ----I.e. Archon is at (5,3) from current location. A muta is at (2,2) from current location would have a ----new weight value of (2-W*5,2-W*3). In essence the "distance" is from a point W* 100% between the current archon and its nearest neighbor. ----Pursue said muta.
EDIT: The "groupthink" algorithm is too heavy for 35000 APM performance. NVM.
There should almost be a limitation of "1 command per frame"..... right now, every single unit can be controlled flawlessly, each frame (basically). That's insane.
This is a little off-topic but could someone tell me what the piano piece name of the second video is called? I know I've heard it alot and im sure its quite famous, but i just can't put my finger on it. If someone could tell me that would be really appreciated =P ty
On November 27 2009 15:48 lMPERVlOUS wrote: Not knocking you work - this is incredibly impressive, but I think I noticed a way to improve it.
It looks like the units retreat until they are ready to attack again, rather than timing their movement towards an enemy unit to coincide when they are ready to attack. If you did this instead (not sure how), you may be able to cut down the time required to kill this shit by up to half.....
BTW - FUCKING INCREDIBLE!
That seems like a good idea at first, but the thing is that you don't know what direction THEY will move in. If they come straight at you when you go back to attack they will meet you when your CD is not over, thus fucking you over because you gotta retreat and waste more time.
Of course I could make it the worst everytime and ASSUME they will come straight at me, that will improve it by a very small margin though.
So on that note, it appears that the units prioritize their own safety above dealing damage. Does that mean they can be cornered? If so, do they bother counterattacking or just continue fleeing? Do they find a default waypoint to escape to and just sustain minor damage?
Great example videos, by the way. Looks like things are shaping up nicely.
On November 27 2009 16:58 Pieguy314 wrote: This is a little off-topic but could someone tell me what the piano piece name of the second video is called? I know I've heard it alot and im sure its quite famous, but i just can't put my finger on it. If someone could tell me that would be really appreciated =P ty
On November 27 2009 16:58 Pieguy314 wrote: This is a little off-topic but could someone tell me what the piano piece name of the second video is called? I know I've heard it alot and im sure its quite famous, but i just can't put my finger on it. If someone could tell me that would be really appreciated =P ty
On November 27 2009 17:32 KwarK wrote: Oh, if it's possible, do the dt in the minefield next. I'd love to see one of those fuckers just strolling through.
On November 27 2009 15:48 lMPERVlOUS wrote: Not knocking you work - this is incredibly impressive, but I think I noticed a way to improve it.
It looks like the units retreat until they are ready to attack again, rather than timing their movement towards an enemy unit to coincide when they are ready to attack. If you did this instead (not sure how), you may be able to cut down the time required to kill this shit by up to half.....
BTW - FUCKING INCREDIBLE!
That seems like a good idea at first, but the thing is that you don't know what direction THEY will move in. If they come straight at you when you go back to attack they will meet you when your CD is not over, thus fucking you over because you gotta retreat and waste more time.
Of course I could make it the worst everytime and ASSUME they will come straight at me, that will improve it by a very small margin though.
Well... One way around this is not to assume they will come straight at you but to assume that they will continue moving the direction they are currently moving. (which, for now is pretty safe.)
On November 27 2009 17:32 KwarK wrote: Oh, if it's possible, do the dt in the minefield next. I'd love to see one of those fuckers just strolling through.
What do you mean? A dt killing mines as it goes?
apparently dts have like 1 range or whatever so they can actually destryo the mines before they explode o.o
On November 27 2009 17:32 KwarK wrote: Oh, if it's possible, do the dt in the minefield next. I'd love to see one of those fuckers just strolling through.
From a programmers standpoint that would be really easy to make right? Compared to having 12 wraiths vs. 48 hydras it should be simple for these guys.
So impressive, there are so many different possibilities of this... I am so glad you made these vids. Have you tried playing against it? I mean it looks good... but does it fall apart if you group your hydras and time your attacks with theirs or do they respond enough to move away?
I'm interested to see what happens if you for example take control of a M&M group vs the mutas, do they fail due to not being able to focus fire properly?
On November 27 2009 16:58 Pieguy314 wrote: This is a little off-topic but could someone tell me what the piano piece name of the second video is called? I know I've heard it alot and im sure its quite famous, but i just can't put my finger on it. If someone could tell me that would be really appreciated =P ty
I think it was on Cowboy Bebop?
Omg I didn't realize I had sound turned off the first time. Just watched it again with sound on... so fitting... I cannot describe.
So, is this on Singleplayer or on UDP via LAN? I can't really tell because I don't know how much prediction you used, but I would assume Singleplayer. If it isn't UDP, and UDP has that 5 frame delay I heard it has, the video isn't half as impressive as it could be.
Also, imo there is a lot of wasted dps. My suggestion for perfect muta vs archon micro would be to zig-zag away from the archons to keep the muta speed at the max, and also keep the perfect distance so that you can do backwards shots without getting hurt. Can't say for sure that this actually works because obviously noone has been able to do that yet. But if it works (the main issue should be how much a muta decelerates during a backwards shot), the dps would be much higher and you wouldn't get any hits at all because you never have a muta and an archon going towards each other at top speed.
The best Archon AI against that imo would have to think in much longer timespans, and try to trap the mutas. Obviously this isn't going to work if there is a cliff, so all you could do is chase the mutas out by attempting a surround and then retreating as far as possible. But a much better unit composition against mutas would be archon+sair anyways. The sairs deal the main damage and the archons punish any mutas that try to kill the sairs.
Even though I guess that AI implementation is comparatively simple and straightforward, it still is impressive.
On November 27 2009 17:32 KwarK wrote: Oh, if it's possible, do the dt in the minefield next. I'd love to see one of those fuckers just strolling through.
What do you mean? A dt killing mines as it goes?
If a dt is walking directly towards a mine and goes into hold position just before it reaches the dts maximum range it swipes and kills the mine. If a dt is attempting to kill a mine it walks right up to its attack range then swipes which unfortunately allows the mine to kill it. The dt has slightly more range but the mine has faster attack speed. It is relatively easy to micro a dt against a mine you know the whereabouts of because you can send the dt towards the mine and be ready to hold position at the last moment. It is much harder to do it blindly because you'd have to change direction to face the mine instantly and be ready to hold. Perfect situation for AI exploitation.
Oh, and dragoons microing vs mines would be good too. Stepping back and manually targetting the mines, two on each, would be sick. PvT you could just skip robo for sick mass goon attacks.
I think an improvement can be made to the AI, by having air units basically "orbit" the opponent just out max range as opposed to the "head toward and back" strategy. Instead it would orbit just outside max range, move in a shallow angle to shoot and move out back into orbit depending on if the opponent gives chase, this lowers the approach rate between units and allow good exploitation of cooldown.
For group AI, i think just perfect stacking would be good if no range advantage exists against ground units. Use the above orbiting mechanism as well as groups big enough to one shot should do the job.
On November 27 2009 16:58 Pieguy314 wrote: This is a little off-topic but could someone tell me what the piano piece name of the second video is called? I know I've heard it alot and im sure its quite famous, but i just can't put my finger on it. If someone could tell me that would be really appreciated =P ty
i was listening to the same chopin album containing that piece, I thought I hit the next track on my keyboard when I played the video, but I hit pause. Confused me good for a few seconds.
On November 27 2009 22:46 BookTwo wrote: i was listening to the same chopin album containing that piece, I thought I hit the next track on my keyboard when I played the video, but I hit pause. Confused me good for a few seconds.
so cool that there is more people listening to the same thing as me around here :D
4 pool beats a 4 probe rush, and has good chances against any opening, so I think that is what most teams will be working on. It's the micro of six zerglings simultaneously that is key.
On November 27 2009 16:58 Pieguy314 wrote: This is a little off-topic but could someone tell me what the piano piece name of the second video is called? I know I've heard it alot and im sure its quite famous, but i just can't put my finger on it. If someone could tell me that would be really appreciated =P ty
I think its Liszt's Liebenstraum
Yes! This is one of my favorite piano pieces.
The AI is awesome. Would love to see some stacking action for more concentrated fire!
On November 27 2009 23:41 datscilly wrote: 4 pool beats a 4 probe rush, and has good chances against any opening
not true. 6 pool is known to be 50% better than 4 pool.
Not true on 2 player maps. 5 probes absolutely own the 4 drones they find mining when they get there. That means no minerals for lings = gg. Strange but true. Rushing with probes hard counters 5 pool.
On November 27 2009 23:41 datscilly wrote: 4 pool beats a 4 probe rush, and has good chances against any opening
not true. 6 pool is known to be 50% better than 4 pool.
Not true on 2 player maps. 5 probes absolutely own the 4 drones they find mining when they get there. That means no minerals for lings = gg. Strange but true. Rushing with probes hard counters 5 pool.
would be really cool if some progamer gambled the other was goin 5 pool, and was right -_- winning with worker rush in osl finals would be memorable
On November 28 2009 01:06 arb wrote: would it be more benficial to the ai microing if you included a stacking unit? such as a larvae or something or would you have to write a new AI?
otherwise this is incredible
Well it sure looks cooler without it. Also with a stacking unit, all the muta would take damage every time it got hit, so it would actually be worse against the archons, but I think the wraiths would be much crazier if they had a stacking unit.
You actually can't control units in groups with BWAPI like in SC. You'll have to keep them grouped in your own code, and have to do all the stacking manually. It shouldn't be too hard to do, but it probably is not trivial either. Well, at least I don't think there is a way to control groups, correct me if I missed something there (but I must be blind if I missed that).
Stacking only makes sense against units you can't outrange properly. Against units with splash it doesn't make much sense in either case (though there might be a few exceptions). Against units that WILL hit you while you hit them, stacking minimizes the number of units that can hit you and maximizes the damage you can dish out in a short time. If you wouldn't stack against marines, you would recieve considerably more damage while not dealing any more damage than with stacking.
Very cool! How does the AI act against enemies with superior range, like goliaths or turrets? Is it already able to micro ground units, and how? Are there already BO and macro measures in place?
looks like it's going to be a bunch of people with high lvl knowledge of Starcraft (us) against a bunch of people with advanced programming knowledge. Except... we apperantly have a bunch of people with programming knowledge as well. GL TL!
is TL fielding a team? didn't see them on the official team list
Nice micro from the Mutas / Wraiths. Terrible micro from the Archons / Hydras. Make a new video where the micro from the Archons/Hydras isn't so terrible, and you might impress me.
On November 28 2009 02:53 Zato-1 wrote: Nice micro from the Mutas / Wraiths. Terrible micro from the Archons / Hydras. Make a new video where the micro from the Archons/Hydras isn't so terrible, and you might impress me.
On November 28 2009 01:40 spinesheath wrote: You actually can't control units in groups with BWAPI like in SC. You'll have to keep them grouped in your own code, and have to do all the stacking manually. It shouldn't be too hard to do, but it probably is not trivial either. Well, at least I don't think there is a way to control groups, correct me if I missed something there (but I must be blind if I missed that).
Yeah, that's why it's 35000apm. BWAPI executes every command by selecting an individual muta and issuing the command to that one muta.
Music: 1. Liszt - Liebestraum No. 3 2. Chopin - Piano Concerto No. 2, movement 3
Videos/code are almost entirely by AssuredVacancy afaik (might be based on Catyoul's code, I'm not sure)
On November 28 2009 02:27 GogoKodo wrote: Went to the bwapi forum to see if other teams have any videos. Nothing quite as cool as the muta/wraith micro but still neat
On November 28 2009 01:40 spinesheath wrote: You actually can't control units in groups with BWAPI like in SC. You'll have to keep them grouped in your own code, and have to do all the stacking manually. It shouldn't be too hard to do, but it probably is not trivial either. Well, at least I don't think there is a way to control groups, correct me if I missed something there (but I must be blind if I missed that).
Yeah, that's why it's 35000apm. BWAPI executes every command by selecting an individual muta and issuing the command to that one muta.
Music: 1. Liszt - Liebestraum No. 3 2. Chopin - Piano Concerto No. 2, movement 3
Videos/code are almost entirely by AssuredVacancy afaik (might be based on Catyoul's code, I'm not sure)
yeah that block of code from catyoul that calculates the direction to move back. I was too lazy to think about all the math so i just copy pasta'd it.
On November 28 2009 01:40 spinesheath wrote: You actually can't control units in groups with BWAPI like in SC. You'll have to keep them grouped in your own code, and have to do all the stacking manually. It shouldn't be too hard to do, but it probably is not trivial either. Well, at least I don't think there is a way to control groups, correct me if I missed something there (but I must be blind if I missed that).
Yeah, that's why it's 35000apm. BWAPI executes every command by selecting an individual muta and issuing the command to that one muta.
Music: 1. Liszt - Liebestraum No. 3 2. Chopin - Piano Concerto No. 2, movement 3
Videos/code are almost entirely by AssuredVacancy afaik (might be based on Catyoul's code, I'm not sure)
yeah that block of code from catyoul that calculates the direction to move back. I was too lazy to think about all the math so i just copy pasta'd it.
Whats the next steps in this? like what are the future projects right now?
On November 28 2009 02:27 GogoKodo wrote: Went to the bwapi forum to see if other teams have any videos. Nothing quite as cool as the muta/wraith micro but still neat
Everyone is overreacting so much Oo Seriously, this AI isn't even close to perfection, and without lots of work, mutas won't be as devastating as it might look. What you see in those videos is most likely all the AI can do at the moment. Flying back and forward, and sometimes shooting. It most likely has no idea what to do when faced with a surround. It especially can't anticipate one.
I would love to see someone make a good AI for vultures. Hit and run with two control groups etc. If you could make an AI that automatically surrounds an opponents army and plants mines in key locations before pulling back it would be even more awesome.
On November 28 2009 04:11 sArite_nite wrote: This is beautiful. If you could, why not put the AI into a UMS map for humans to train against perfect micro?
You can't. The AI has to be run on the user's end, it basically injects commands for you and gets to look into the games memory. The AI is a run as a program and therefore cannot be stored in a map. What could be done is if they packaged it as an executable and released it, and you had two computers, you could play with the AI on a UDP LAN game by having two computers, one running the AI, and you actually playing on the other.
On November 28 2009 04:26 DrainX wrote: I would love to see someone make a good AI for vultures. Hit and run with two control groups etc. If you could make an AI that automatically surrounds an opponents army and plants mines in key locations before pulling back it would be even more awesome.
It would be most awesome if the AI would judge the chances of success based on terrain, positioning, army composition, army size and leftover HP, as well as chance of reinforcements/unsouted units for the enemy, and in case that chance is worth the risk, it would do what you said.
Pretty good. Although the problem with the way it works is that you need a lot of open space. If the units are inside a base or something, a human player can easily trap the AI. Also the human can flank from two sides. Actually, this AI will probably work best for air units.
On November 28 2009 04:11 sArite_nite wrote: This is beautiful. If you could, why not put the AI into a UMS map for humans to train against perfect micro?
On November 28 2009 04:22 spinesheath wrote: Everyone is overreacting so much Oo Seriously, this AI isn't even close to perfection, and without lots of work, mutas won't be as devastating as it might look. What you see in those videos is most likely all the AI can do at the moment. Flying back and forward, and sometimes shooting. It most likely has no idea what to do when faced with a surround. It especially can't anticipate one.
Actually it deals with surrounds pretty well. Unless the surround is perfectly symmetrical, it will find an edge with the most space to escape.
On November 28 2009 04:22 spinesheath wrote: Everyone is overreacting so much Oo Seriously, this AI isn't even close to perfection, and without lots of work, mutas won't be as devastating as it might look. What you see in those videos is most likely all the AI can do at the moment. Flying back and forward, and sometimes shooting. It most likely has no idea what to do when faced with a surround. It especially can't anticipate one.
Actually it deals with surrounds pretty well. Unless the surround is perfectly symmetrical, it will find an edge with the most space to escape.
SHOW THE DISBELIEVERS :@ This is awesome... I can't believe there are people who don't think this is crazy cool. People have been doing AI for years and years, and the closest thing to this I ever saw was some upgraded peon harassment that only worked in conjunction with map triggers. Most AI up until now has just been cheat AI that gave the computer a lot of resources and instructions to spend them.
To test the muta's ability to get the fuck out of a shitstorm, I spawned scourges on all sides of the mutas so they would be surrounded. The anti scourage micro still needs a bit of work though.
45k apm? i see you upgraded ;-) The problem with the videos you are showing us is that the opposing AI is just a-moving. Anybody can script an AI that beats an a-moving opponent, and while it's certainly the place to start, you seem to be past that. Show us some of you AI vs AI battles. Personally what I want to see is corsairs vs scourge, or muta/scourge ZvZ battles. Take this AI (the one in previous videos) as the default, then show us what is capable of beating it. It shouldn't be to hard.
On November 28 2009 07:04 flx! wrote: 45k apm? i see you upgraded ;-) The problem with the videos you are showing us is that the opposing AI is just a-moving. Anybody can script an AI that beats an a-moving opponent, and while it's certainly the place to start, you seem to be past that. Show us some of you AI vs AI battles. Personally what I want to see is corsairs vs scourge, or muta/scourge ZvZ battles. Take this AI (the one in previous videos) as the default, then show us what is capable of beating it. It shouldn't be to hard.
Great work so far, AI scripting can be very fun.
I really want to try to play against my AI, but I only have one machine and I haven't got VM set up for it. And from what I heard from catyoul, it lags it up by having 2 ais playing against each other.
The thing with scourge vs muta micro is even if there was a better AI behind it, I doubt the results would be any different since the scourge can't get near the mutas for any micro(cloning) to take effect.
On November 28 2009 04:54 Avidkeystamper wrote: Wouldn't this be kinda bad against MnM seeing how the mutas don't attack at once?
Not really. So long as each muta is hitting marines, it would be pretty hard for medics to heal the marines, and since the mutas all act independently, they get individual attacks in more frequently than a stack would be able to, since they can use openings that wouldn't be available to a stack.
It wouldn't be optimal, but by no means could it be called bad.
@TheYango: Well isn't the reason mutas are stacked in the first place are to kill marines fast which unstacked mutas wouldn't really be able to do unless like just enough mutas to kill a marine target a single one which still won't work because without the stacking mutas are easier to hit. That is unless the mutas don't target a single marine all at once but close to the same time which still doesn't work too well since it makes it easier on the medic to heal the marine. ALTHOUGH it still isn't bad because it makes it much easier for mutas to work around marines and harass several spots.
On November 28 2009 07:32 bp1696 wrote: Wow, this is really awesome.
Will computer be able to beat the pros one day soon? Reminds me of Deep Blue and chess .
Honestly, by the end of the AIIDE contest, the AIs will beat anyone on the 4 maps in the pool. The difficulty of AI programming is rarely making something that beats humans (it was with chess and still is with Go), it's giving the user the impression of playing against a human, who makes mistakes like him. It's all about creating an enjoyable experience. That's really hard. Also if you limit the APM of the AI to something reasonable, like 1000, then all the problem is about making choices and decisions, and that is also quite hard.
On November 28 2009 07:32 bp1696 wrote: Wow, this is really awesome.
Will computer be able to beat the pros one day soon? Reminds me of Deep Blue and chess .
no thanks, having a computer play better than the best human players destroys the game imo
I don't see chess as being a "destroyed" game despite a wide range of AIs. Even a middle of the road AI that a normal consumer computer processor can run is easily capable of destroying the majority of human players. But I would still rather play chess with another person rather than with a computer. Same for tournaments.
Maybe a team of Jaedong, Effort, Calm, Luxury, Yellow, Kwanro, Hyuk, Zero, Hero and July, each controlling one mutalisk, could replicate that result. And I guess also get something like 4,000 apm.
On November 28 2009 07:32 bp1696 wrote: Wow, this is really awesome.
Will computer be able to beat the pros one day soon? Reminds me of Deep Blue and chess .
no thanks, having a computer play better than the best human players destroys the game imo
I don't see chess as being a "destroyed" game despite a wide range of AIs. Even a middle of the road AI that a normal consumer computer processor can run is easily capable of destroying the majority of human players. But I would still rather play chess with another person rather than with a computer. Same for tournaments.
In either game that person could be cheating by using an AI though. Of course with SC, it'd be wayyyy more obvious.
Stupid question but why put wraiths against unupgraded hydras? Show off the AI a bit? The Dragoon micro is crazy though, I'd like to see more like that where the AI shows off against bad or more even odds and destroys the other units.
On November 28 2009 08:54 iCCup.Raelcun wrote: Stupid question but why put wraiths against unupgraded hydras? Show off the AI a bit? The Dragoon micro is crazy though, I'd like to see more like that where the AI shows off against bad or more even odds and destroys the other units.
To have an advantage while microing against range upgraded hydras I need to stack the wraiths. And so far I have not found a way to do that yet.
On November 28 2009 07:39 zergnewb wrote: @TheYango: Well isn't the reason mutas are stacked in the first place are to kill marines fast which unstacked mutas wouldn't really be able to do unless like just enough mutas to kill a marine target a single one which still won't work because without the stacking mutas are easier to hit. That is unless the mutas don't target a single marine all at once but close to the same time which still doesn't work too well since it makes it easier on the medic to heal the marine. ALTHOUGH it still isn't bad because it makes it much easier for mutas to work around marines and harass several spots.
It's sort of a tradeoff. A muta stack hitting single marine makes the most real gain in an engagement (since any injured marines can get healed), but it also wastes damage, because often your stack is large enough to do more damage (which is why with a large stack, hold position get's used, so some of the damage gets spread around). If each muta is getting controlled individually, then damage will be put on every marine, because they're attacking from a wide angle. The medics simply can't keep up with the damage being dished out, because they have to travel back and forth to heal damaged marines (whereas with concentrated fire, they only need to heal the same 2-3 marines). Obviously, this also means that this takes longer for the marine force as a whole to die, and they will therefore put out more damage before dying, but with the AI's micro, putting out a significant amount of damage could be difficult.
On November 28 2009 07:32 bp1696 wrote: Wow, this is really awesome.
Will computer be able to beat the pros one day soon? Reminds me of Deep Blue and chess .
no thanks, having a computer play better than the best human players destroys the game imo
I don't see chess as being a "destroyed" game despite a wide range of AIs. Even a middle of the road AI that a normal consumer computer processor can run is easily capable of destroying the majority of human players. But I would still rather play chess with another person rather than with a computer. Same for tournaments.
In either game that person could be cheating by using an AI though. Of course with SC, it'd be wayyyy more obvious.
Yeah, the ICC is pretty good about detecting hackers by monitoring your currently level of play. If you suddenly start playing a lot better they're going to know you're abusing computers. The same could be done with Starcraft in theory.
On November 28 2009 09:08 XinRan wrote: If unit groupings don't exist in Starcraft programming, how does the game handle hotkeyed control groups?
It's not that it doesn't exist in starcraft, it's just that the BWAPI doesn't hook into it and provide us interface functions to use them.
On November 27 2009 14:43 AssuredVacancy wrote: Yeah I don't think the team will bust my balls for this because I took off all the debugging markers before filming, so this just reveals very very cool micro rather than anything important.
Well, you know my opinion. I prefer we don't make this kind of videos public, but it's no big deal for this kind of stuff. I know it's a great feeling to show them and to have your work recognized. It's somewhat hard emotionally to keep them private, I'm really fighting hard to keep mine under wraps honestly :D But I still will, at least for now. Maybe in a couple of weeks/months when they'll be outdated.
On November 28 2009 09:31 motbob wrote: Hey guys, look! Another team managed to make a full game Terran AI! It seriously looks like it could beat C players on iccup... just watch.
On November 28 2009 09:31 motbob wrote: Hey guys, look! Another team managed to make a full game Terran AI! It seriously looks like it could beat C players on iccup... just watch.
On November 28 2009 09:31 motbob wrote: Hey guys, look! Another team managed to make a full game Terran AI! It seriously looks like it could beat C players on iccup... just watch.
On November 28 2009 07:32 bp1696 wrote: Wow, this is really awesome.
Will computer be able to beat the pros one day soon? Reminds me of Deep Blue and chess .
Honestly, by the end of the AIIDE contest, the AIs will beat anyone on the 4 maps in the pool. The difficulty of AI programming is rarely making something that beats humans (it was with chess and still is with Go), it's giving the user the impression of playing against a human, who makes mistakes like him. It's all about creating an enjoyable experience. That's really hard. Also if you limit the APM of the AI to something reasonable, like 1000, then all the problem is about making choices and decisions, and that is also quite hard.
I dont believe that for a second... Yes, the AIs can micro air units and ground units on open ground well, and macro, but hows the strategical aspect, target prioritizing, army positioning, risk assessment, terrain, filling out missing info etc etc done?
Afaik theres no existing examples of computers doing extraordinarily well on a limited information game. So to me its ridiculous to say its a given that AI will destroy everyone in 2010, you dont have any solid backing for that atm. It might happen, some matchup on some map, but surely its not easy...
Micro AI is really nice. But what about all the rest? I remember my AI battles in WC3 and the biggest problem I had, was even if I managed to get awesome build orders (and AI actually following it like it should) I had absolutely no influenco on where/how the buildings were placed. For example, if your strategy required some towers to be built on the perimeter, AI would put them randomly in center of the base etc. Building placement is very important. I don't even mean wall-in or stuff like that, but just look how much depends on it: unit pathing and distance they need to travel to waypoints, flanking options, defence etc. etc. I guess with Zerg most of the problems go away, but still sunken placements can mean difference between winning and losing.
On November 28 2009 11:56 Manit0u wrote: Micro AI is really nice. But what about all the rest? I remember my AI battles in WC3 and the biggest problem I had, was even if I managed to get awesome build orders (and AI actually following it like it should) I had absolutely no influenco on where/how the buildings were placed. For example, if your strategy required some towers to be built on the perimeter, AI would put them randomly in center of the base etc. Building placement is very important. I don't even mean wall-in or stuff like that, but just look how much depends on it: unit pathing and distance they need to travel to waypoints, flanking options, defence etc. etc. I guess with Zerg most of the problems go away, but still sunken placements can mean difference between winning and losing.
On November 27 2009 18:37 TryThis wrote: this is jsut artful, i cant stop watching it.
isnt bisu stuck playing these things? im excited to see his face when this happens to him
It will of course be interesting to see "strategy" and "perfect micro" clash, but I always assumed this would be where the TL team would shine: They actually know how to play starcraft.
On November 28 2009 07:32 bp1696 wrote: Wow, this is really awesome.
Will computer be able to beat the pros one day soon? Reminds me of Deep Blue and chess .
Honestly, by the end of the AIIDE contest, the AIs will beat anyone on the 4 maps in the pool. The difficulty of AI programming is rarely making something that beats humans (it was with chess and still is with Go), it's giving the user the impression of playing against a human, who makes mistakes like him. It's all about creating an enjoyable experience. That's really hard. Also if you limit the APM of the AI to something reasonable, like 1000, then all the problem is about making choices and decisions, and that is also quite hard.
I dont believe that for a second... Yes, the AIs can micro air units and ground units on open ground well, and macro, but hows the strategical aspect, target prioritizing, army positioning, risk assessment, terrain, filling out missing info etc etc done?
Indeed those are the hard parts in AI programming.
Afaik theres no existing examples of computers doing extraordinarily well on a limited information game. So to me its ridiculous to say its a given that AI will destroy everyone in 2010, you dont have any solid backing for that atm. It might happen, some matchup on some map, but surely its not easy...
Well FPSs are limited information games, and non-cheating bot still rape humans on it.
I think you underestimate the importance of mechanics in games such as these. If your mechanics are just ridiculously good, you only need to avoid making really stupid mistakes in order to beat a human. Yes mechanics in FPSs are much more important relatively to tactics and strategies than in RTSs, still, I believe a half witted bot with crazy micro/macro and even just random decision making (within reasonable choices) will beat most people. No doubt as people play against those bots, flaws will appear to be exploited and maybe difficult to fix, but I doubt it.
Also, Texas Hold'Em poker is example, perhaps more interesting, where bots are currently slightly better than pros. The mechanics of poker are very low (knowing probabilities) and it's all about learning, and figuring out your opponents. check this out if you're interested: http://www.wired.com/special_multimedia/2008/ff_poker_1612 http://poker.cs.ualberta.ca/
So basically I believe that, if people care enough to create it, it's a given that Starcraft AI bots will rape any human on those maps. Now creating a bot that also figures out any maps, that may be a bit harder.
On November 29 2009 08:04 meeple wrote: I'd love to see that muta vs scourge micro perfected... but seeing video makes me wonder more about why you can't stack the muta AI?
I think with micro like that it's more effective for them not to be stacked, at least vs Archons
On November 29 2009 08:04 meeple wrote: I'd love to see that muta vs scourge micro perfected... but seeing video makes me wonder more about why you can't stack the muta AI?
I think with micro like that it's more effective for them not to be stacked, at least vs Archons
I meant for the scourge, so they could one-hit them
On November 29 2009 08:04 meeple wrote: I'd love to see that muta vs scourge micro perfected... but seeing video makes me wonder more about why you can't stack the muta AI?
I think with micro like that it's more effective for them not to be stacked, at least vs Archons
I meant for the scourge, so they could one-hit them
i guess he has a different object for each mutalisk and they all acting as individuals at the moment. Maybe later he'd create a "team" object where he groups say 3 mutas together and treat all of those as a unit
not only would the AI have 'almost' perfect muta micro, it would be simultaneously macroing every single resource it gets and expoing to assure map control. that is very impressive, however AI has always been made very methodically with set answers to certain situations and not really with a freeform ability to think independently.
another example of good AI in a RTS game was Sins of a Solar empire. if you gave them time to set up an infrastructure without cheesing them, they seemed to utterly outplay humans because of being able to execute far more commands in a short period of time than a human possibly could.
i am definitely looking forward to future AI that doesn't exclusively rely on perfect mechanics to outmuscle a human opponent but rather incorporate strategy to beat us.
On November 28 2009 22:11 flx! wrote: Also, Texas Hold'Em poker is example, perhaps more interesting, where bots are currently slightly better than pros. The mechanics of poker are very low (knowing probabilities) and it's all about learning, and figuring out your opponents. check this out if you're interested: http://www.wired.com/special_multimedia/2008/ff_poker_1612 http://poker.cs.ualberta.ca/
Definitely not even close to true... if there were bots that were better than pros then why arent there bots playing NL1K online and making their owners millions of dollars? EDIT: Didn't realize this was for limit Hold em....
On November 27 2009 14:33 AssuredVacancy wrote: The anti scourage micro still needs a bit of work though.
Did a lot to improve the AI. But being really good means that I'll have to keep the video private so the competition doesn't see it in action. I'll release vids/reps when they become outdated though. Although I do think that forum vets should be allowed to the stuff we're working on because so much of it is amazingly cool.
On November 27 2009 14:36 Count9 wrote: Not much macro with 35k apm tbh. Not very impressive against auto attacking archons, should try against a human... or make a archon micro and see who wins :/
Ummmm what difference would it make? It's clearly based on distance to archon, distance that muta can shoot, distance of archons projectile...human micro, computer micro, makes no difference at all.
Cant you just make a "unitgroup" in the code that sends the exact same orders to several units. This should allow mutas to stack. Basicly this would make the computer calculate the moves it makes as one supermuta and execute them with al?
Btw im quite interested in this AI programming contest, is there any further information ?
Im wondering if you'd be willing to share the code with me ? Im just starting on using the BWAImodule and would like to see it. Maybe i can try implement stacking for mutas.
Ofcourse i understand you might want to ceep it to yourself for the contest.
The program is quite ez to start up now im gonna see what i can do with it
EPT
!
