So, i created a Protoss bot, which vs terran focuses on attacking early with +1 speedzealots with 1-2 dark templar and securing its expansion at the same time as its attack. The follow up is, i guess a little too simplistic (i'm still working on it..), is to get MASS zealots with MASS storm, and the end game is zealot templar arbiter. Only 8 goons are made in the early game, they are supposed to be acting as mobile base defenses. Neways, The opponent is also a Bot. I do not know quite how it works, since he is also in the competition, but vs protoss it always goes mech, with a vulture drop after it takes its third. I've played this bot several times, and thats about all i can figure of its play. To wrap up, i've uploaded the replay, it needs to be run with a repfix, cause bots hit 14k apm at some points. My biggest problem with this bot is micro management. This is the part i would like to discuss with everyone, still anything is open for discussion. I've implemented a micro management for storming, i'm not sure if storm placement is ideal, but i have it storming one tile behind the target, where behind is away from the caster. and i have also stasis working in a similiar manner. My main problem is, ways of microing zealots. In this match, they were all just a-moving, previous attempts at micro was BAD! Is my army composition weak? I just felt as the game progressed, i was wininng, but maybe i was jsut lucky. Well lastly, these Bots do not have perfect information, they only are able to see what a human can see. Well neways... its an intresting match between AIs, there are some other ones, TL has a bot, forgot teh name, and there are a couple others. EISbot, and SkyNet are others. Mine is still nameless.. Dunno how to end this, i guess: Discuss
Well I have NO knowledge about making AI for SC but good luck with yours! It sounds REALLY cool. Hey if you polish your AI enough, it might play like Stork! lol
Theoretically a perfectly made SC AI should be unbeatable since it can run like 27 simultaneous harass attempts and fight a center battle while still making perfect macro. Take advantage of the strong point of the AI: the 14k apm, as you say, and emphasise storm raids and DT drops.
As for how to micro the zealots; one experiment you can try is have the zealots directly attack tanks if tanks are in seige mode so they run past the vultures and mines. You can also have ht in shuttles to storm raid seiged tanks. Make them drop a single zealot first to absorb tank fire.
I also think you should mix in a few dragoons. If your opponent knows your ai doesn't make goons he's going to be able to abuse you with goliaths and vultures.
On July 25 2010 15:26 Hautamaki wrote: Theoretically a perfectly made SC AI should be unbeatable since it can run like 27 simultaneous harass attempts and fight a center battle while still making perfect macro. Take advantage of the strong point of the AI: the 14k apm, as you say, and emphasise storm raids and DT drops.
As for how to micro the zealots; one experiment you can try is have the zealots directly attack tanks if tanks are in seige mode so they run past the vultures and mines. You can also have ht in shuttles to storm raid seiged tanks. Make them drop a single zealot first to absorb tank fire.
I also think you should mix in a few dragoons. If your opponent knows your ai doesn't make goons he's going to be able to abuse you with goliaths and vultures.
The amount of coding required to replicate the strategic and tactical decision making that comes naturally to the human brain is absurd. While I agree a human cannot rival a machine in macro that doesn't change the fact that a human can choose when to macro and when to power far more effectively. Bots won't be getting anywhere near the top level of players without maphack anytime soon.
Hi. Ex-Programmer here with some background in SC.
Many of us don't know the extent to which decisions can be made in SC AI. What parameters are there? Simply saying that it sees what we see is not enough. ie...
Can the bots keep track of how much health each unit has? Count how many peons per gas geyser/min patch? Bases taken? Production/tech building/depot/overlord/pylon count/existence? Opponent army composition? Lights on building (core-goon range/factory/rax production being run, etc) Which upgrades have been taken? For example PvT, if a unit gets hit by mine or sees one lain, can you trip a flag/register/boolean var to say that mines has been researched? Or can it check for +1/0 in the unit portraits etc?
Basically, what intel can the bot gather? (Tech/Unit/Terrain) And to what extent of flexibility do you have to construct strategies/tactics for the bot to follow? How about movement? Can you mimic SC2 style unit movements rather than sc1 bumper cars style?
I'm very interested in contributing to such a project via pseudo-code for algorithms. I specialize in constructing elaborate control structures to dictate optimal strategies.
Let me know how we can help.
(PM me and I'll give you easier ways to contact if you think I can help.)
On July 25 2010 15:26 Hautamaki wrote: Theoretically a perfectly made SC AI should be unbeatable since it can run like 27 simultaneous harass attempts and fight a center battle while still making perfect macro. Take advantage of the strong point of the AI: the 14k apm, as you say, and emphasise storm raids and DT drops.
As for how to micro the zealots; one experiment you can try is have the zealots directly attack tanks if tanks are in seige mode so they run past the vultures and mines. You can also have ht in shuttles to storm raid seiged tanks. Make them drop a single zealot first to absorb tank fire.
I also think you should mix in a few dragoons. If your opponent knows your ai doesn't make goons he's going to be able to abuse you with goliaths and vultures.
The amount of coding required to replicate the strategic and tactical decision making that comes naturally to the human brain is absurd. While I agree a human cannot rival a machine in macro that doesn't change the fact that a human can choose when to macro and when to power far more effectively. Bots won't be getting anywhere near the top level of players without maphack anytime soon.
Hence why I said theoretically =p
It would take a team of top-level AI programmers working together with progamers tens of thousands of man-hours to create an AI that could beat any human player, and obviously that's never going to happen. But still even a single programmer with a high level of knowledge about the game should be able to create a pretty decent AI. I've played around with some AI tools in SC before and have made an AI that came in second place in an AI vs AI tournament hosted by SCLegacy way back in the day. Though of course it didn't come close to being able to beat a good human player for the reasons you mentioned.
Well these AIs that are being created through the use of BWAPI: http://code.google.com/p/bwapi/ BWSAL and BWTA have the potential of being far greater. Ill link some youtubes to show you the potential: EISBOT, (Probably the most advanced full game bot): + Show Spoiler +
On July 25 2010 16:43 Emperor_Earth wrote: Hi. Ex-Programmer here with some background in SC.
Many of us don't know the extent to which decisions can be made in SC AI. What parameters are there? Simply saying that it sees what we see is not enough. ie...
Can the bots keep track of how much health each unit has? Count how many peons per gas geyser/min patch? Bases taken? Production/tech building/depot/overlord/pylon count/existence? Opponent army composition? Lights on building (core-goon range/factory/rax production being run, etc) Which upgrades have been taken? For example PvT, if a unit gets hit by mine or sees one lain, can you trip a flag/register/boolean var to say that mines has been researched? Or can it check for +1/0 in the unit portraits etc?
Basically, what intel can the bot gather? (Tech/Unit/Terrain) And to what extent of flexibility do you have to construct strategies/tactics for the bot to follow? How about movement? Can you mimic SC2 style unit movements rather than sc1 bumper cars style?
I'm very interested in contributing to such a project via pseudo-code for algorithms. I specialize in constructing elaborate control structures to dictate optimal strategies.
Let me know how we can help.
(PM me and I'll give you easier ways to contact if you think I can help.)
Let me know if you
Huh?
And lol I was gonna be like "wtf you're not a progamer" until I realized its "programmer". heh, a little too much time watching replays for me lol
On July 25 2010 12:27 Xiphos wrote: Well I have NO knowledge about making AI for SC but good luck with yours! It sounds REALLY cool. Hey if you polish your AI enough, it might play like StorkBeSt! lol
The Doh says No.
On topic: I haven't learned anything about any sort of programming, so all I can say is: good luck to you, sir. :D
On July 25 2010 15:26 Hautamaki wrote: Theoretically a perfectly made SC AI should be unbeatable since it can run like 27 simultaneous harass attempts and fight a center battle while still making perfect macro. Take advantage of the strong point of the AI: the 14k apm, as you say, and emphasise storm raids and DT drops.
As for how to micro the zealots; one experiment you can try is have the zealots directly attack tanks if tanks are in seige mode so they run past the vultures and mines. You can also have ht in shuttles to storm raid seiged tanks. Make them drop a single zealot first to absorb tank fire.
I also think you should mix in a few dragoons. If your opponent knows your ai doesn't make goons he's going to be able to abuse you with goliaths and vultures.
Yah, except the problem is that the opponent also has 14k apm.
Ya those vids changed colors on me, but not all of the colors, for some reason the green stayed green Hopefully ill have some more matches up on youtube, im working on getting Storm to be more effective, and maybe include zealot bombs, but thats a nightmare to program.... god why do i suck at programming! How do i describe to my AI when it is a good time to do a zealot bomb? If there are 10 tanks within a certain distance? any suggestions?