Author of Stardust here, willing to answer any questions anyone has! I am unfortunately not at all surprised by the result. All of the bots you see in this series are programmed by hobbyists, so we don’t have the kind of resources needed to do machine learning at the scale required for StarCraft. This means every high-level strategy decision has to be accounted for explicitly in the code, so it’s really easy to make bad strategy decisions and end up looking embarrasingly bad :) I am looking forward to the upcoming games though, as Stardust’s Protoss brethren you will see in the other matchups tend to be much better at that aspect of the game.
Is it possible to replace the base games AI with yours. Would be pretty funny to watch one of your friends get toasted thinking he was just playing against a normal computer. Especially if u bet him 100 bucks he couldnt beat a computer in Starcraft. Id love to get my hands on this AI, so cool. Nice work :D
I remember when I saw SC2 AI for terran. It was using bunkers to increase gas mining. It was sendng SCV from CC to bunker and instant unload (it was teleporing by bunker width), then SCV was entering refinery. When with gas, it was put again to bunker, teleported and to CC. it was doing that trick while microing every marine separatly to keep it just out of range of enemies. that was about 20k APM.. as I remember correctly
Let's not mistake a trick implemented to AI by bot creators and tricks learned by the bot itself. The trick with the bunker was implemented by the bot creators. AI has limitless APM and the ability to be "everywhere", so such tricks are no challenge to AI.
That was tried in BW too, but it doesn't really work. You can gain some time in one direction by loading in one side and unloading on the other, but you gain no time in the other direction, because you always unload on the same side of a bunker. Actually, you lose time on the way back because you have to go around it. So you still need 3 SCVs to mine optimally. Also, the trick is not really that useful. It allows you to mine gas optimally with 2 SCVs instead of 3, but to do that you have to build a bunker. And of course, a bunker is more expensive than an SCV, and an SCV is occupied while building it. It also slows down your gas mining while it is being built. So it's a significant up-front cost to ultimately get one extra SCV on minerals. I'm surprised it's worth it at all tbh.
@@EebstertheGreatBunker can be loaded from any side. So if it's built off-side from the gas path, it doesn't prevent SCVs from going to the Refinery. It reduces the distance won by doing this trick by like 1 block of movement, but cancels out all the possible drawbacks.
I helped start this scene 15 years ago, and ran the first StarCraft AI competition as part of an academic conference. The winner was the Skynet team from UC Berkeley, and there's an article about the bot and competition on ArsTechnica. Happy to answer any questions. My thoughts on the bots: - The bots give individual unit commands rather than using control groups, which inflates the APM a bit. - The API isn't supported by Blizzard, and has some quirks to it, such as workers sometimes not properly labeled as idle - The Terran bot likely bugged out, since there's the same situation as human players, where sometimes buildings don't make. And there's not a notification when this happens. This is because the API is reverse engineered, and does a best effort at providing functionality. It probably tried to start a Starport and failed without knowing it.
But it finished a refinery and then didn't mine it with anything. There is a notification when you finish a building. And what happened to the SCV building the refinery in the first place? Shouldn't it have immediately started to mine? I agree it definitely bugged out after that, and the starport not building 100% can explain it. But the fact that it didn't mine right away is weird. Maybe it pulled the SCV from the refinery off immediately to build the depot, cause it was the closest one, so it was just bad timing?
The API does not provide callbacks for building completions or failures, or even successful starts. It's a bit tedious, but you can manually check if buildings have started.
The bots don't seem to be properly abusing their superhuman abilities. Like the lack of mineral boosting noted with the Terran one, or the Zerg one not having perfect muta micro and that sort of thing. The ling concave is pretty much the best showcase there. Protos might have some insane reaver shuttle play at some point too.
I don't think reaver would be a good showcase for micro because it has a forced attack delay after being dropped. Terran units however do not have this restriction put on them, so the AI could so some magnificent tank-drop micro
@@uxie6177@strike_jk Top bots can do a few things with Reavers which are superhuman: - Pick up and re-drop after every shot, which lets Reavers shoot almost twice as fast (because dropping Reavers doesn't delay their attack; it sets the cooldown to a fixed value!) - Micro any number of Reavers/Shuttles simultaneously - Target every Reaver shot, to attempt maximizing splash damage or just avoid shooting "AI Pylons" (Decoy Pylons humans place in the natural to draw incidental Reaver shots) On the other hand, their decisionmaking with Reavers tends to suffer due to poorly handling the impact of terrain/collisions on Reaver shots, not accurately simulating Reavers when predicting how fights will turn out, and not dealing well with the acceleration characteristics of Shuttles.
@@perfectman3077They can do it! Funnily enough, storm is actually quite weak against bots because they can instantly dodge it. Even Marines can largely survive storm with instant reactions and splitting.
It all boils down to how well versed in Starcraft the programmer is. If they implemented good strategies and micro methods into the program. A machine would execute perfectly what you tell it to do. The question is, what do you tell it to do? Either that or it's an actual AI program that can learn and get better as it plays more games. Which would explain why it does poorly in the first few - or perhaps hundreds of - games.
They did a last minute-change on the lineup for the terran bot because krasi0 said his bot isn't good at TvZ. But I think he would have done way better than what Terminus did here with the super-odd fast gas into not spending it on anything. I coded a Zerg bot some years ago and I never managed to find a way to get it to beat krasi0. And that's years ago where krasi0 still improved his bot since. Against protoss the 8-9-off-gates were incredibly hard to deal with as Zerg with any kind of standard-build-order. A build that would survive this was obviously pretty bad against any of the more standard ones. ZvP still was my bots best matchup. I also wonder about the Z-bots muta-micro. Mine had different micro depending on whether it was fighting moving or stationary targets. Against moving-targets it would use the move-shooting and against stationary ones the kiting-style that zerg here used. My bot might overall be a buggy mess that loses half of it's games by crashing or something else vital stopping to work. But it sure had some neat micro-methods that I'd invite other bot-authors to use for their otherwise better bots.
@@oneoranota There was a lot of "building on other's foundation" but due to it also being framed as a competition honor kinda prevented too much collaboration.
I've had experience playing with some of these bots a few years ago. My takeaway was that it wasn't terribly difficult to get a strategic win in the early to mid game but if you let them power up their macro engine, they're almost impossible to beat. Would be curious to see a series where some of these guys give the ai more of a fighting chance earlier in the game and then take on the 5000 apm macro engine.
Heck yeah! Thx for working on this and being open to answer questions. Is PurpleWave trained using machine learning techniques? If yes, is it augmented with any human-suggested strategies/tricks?
@@IcthiIndeed. So PurpleWave's strategies are my best human understanding of what comprise good decisions across matchups, as best as I can capture them in code. Sometimes there are concessions like avoiding situations requiring the bot to perform techniques it's not strong at.
I would be really interested in seeing a BW AI tournament being casted, or AI vs human. Something like OpenAI/Dota level- it would be great to see what strategies AI could come up with in Broodwar that defy common practice
I really like this. In my teens I wished for better AI players for Starcraft, as I dindt have internet to play against humans. Now they are here. What a time to be alive!
lol. I hope BW AI never becomes truly next level, because the meta being pushed forward by human learning is so much more interesting. Human strategies and play will always be more engaging, and the pros using machine learning and rote memorization to optimize play will be the death of BW for the fans.
@@mediocrates1937 Even if such AIs become common, I doubt anybody would turn up to watch them. Watching living beings play this clunky game at such a level is what makes it interest.
I honestly hope true AI like AlphaStar do tackle BW. They could show us things about the game we didnt know, and clue us into strategies, techniques or metas we might never even find on our own. Also, you have to consider -- not that many new people really get into BW. All the ASL players are ones who have been at it for years. There's not really a young generation right now getting into BW the same way there is for Melee. The competitive scene for BW will slowly fade away IMO (sadly) so we need to explore the game fully while we can - and we're not getting new developments in the meta or finding new micro tricks much anymore. @@mediocrates1937
I wish we could get like, some different or interesting AI opponents in the Remaster. Sometimes I want to play skirmish but don't want to get completely rekt on the ladder. But it's still the same old AI from the original game. It won't even GG when it knows it's lost. Good video anyway.
@@Ryan41707I love how much of a challenge that Campaign is but the UED units are such a chore to use. They just auto-lose defensively due to their nerfed range/non siegeable tanks
Very interesting to see these! I was actually already curious about where AI is at in brood war. Not sure if I’d like to see AI matches on a regular basis, but as an informative one-off this is really cool!
No there was a huge mistake by the AI over microing the mutas in that ZvZ. It's not just tactical issues. In the final game the AI was trying to pull marines back from a sunken...which was only reducing the damage output of the marines.
ahh yeah I remember watching the old SCAIT league on twitch way back in the day. It's really cool that people are making these things for fun. Important to know that these are AIs in the old sense of the word - like "just a set of if this, do this" instructions more or less. They aren't nearly as elaborate, powerful or complicated as the modern "AI" term (things like ChatGPT for example.) I would LOVE to see AlphaStar for Brood War, I think there's some real potential for it to show us things about the game we didn't know, and show some strategies or techniques that we never thought possible.
Wow, looking forward to that! I like watching the SC2 AI scene (there are some strong bots there, but generally not quite of the level of pro players apart from cheese strats). This will be an interesting comparison.
I love this! There is so much content for StarCraft that we haven't even seen yet. I'd love to see more of these and maybe even a series on the South American scene that you mentioned during the first game @artosis
I am sooo interested in AI and game theory in general and I can't wait for the day that AI shows us how BW works on an even higher level than humanly possible. It is so interesting to me.
Jimin's APM has to be bugged or something. I've never seen an S rank zerg with that low of an APM. If it's not bugged, then I'm actually incredibly impressed, he must be super smart and have great macro...or great cheese
With SCHNAIL the games are played in 1.16 client, and on Normal speed but sped up by the API to feel as if it is Fastest so a human wouldn't notice while playing. This impacts a few things though: Watching the replay back, the replay game timer doesn't matchup with what you normally would expect, as that timer is impacted by the game speed that was played ("Normal" instead of Fastest). I believe this impacts APM as well, and I suspect this pushes player APM lower than normal because 1.) game is treating all the information as if it is Normal speed, but 2.) since the game is running as if it's Fastest thanks to the API, the human player is putting less inputs over time than they would actually do on Normal speed. So the APM shows lower because "more time is passing" by the replay timer bar, but the human is doing less actions over that longer time.
I'd be really interested in a machine learning bot, since these SCHNAIL AIs were preprogrammed. Probably they are optimized to make a single build or maybe rotate a couple, but reacting to a hydra bust... So something like AlphaStar because it'd innovate and actual players could learn from it.
Meta and Samsung were both researching AIs using reinforcement learning for Brood War. Unfortunately AlphaStar beat them to the punch and both companies pulled the plug not too long afterwards. All of the people still actively working on Brood War bots are hobbyists without the resources needed to do any kind of proper machine learning at the scale needed for StarCraft.
@@b-qj9oz Are you sure AlphaStar was trained on BW too? I've only seen it in SC2, I'd love to take a look at BW ML gameplay. And I do understand that SCHNAIL AIs were made by enthusiastic hobbyists. I did not want to throw shade, just acknowledge the limitations of the approach. (And also, I know that most of the resources needed are not for the code but for the training and amateurs just do not have the astronomical amount of money for that kind of hardware.)
@@csucskos I formulated my comment poorly, you are of course correct that AlphaStar was only trained for SC2. However in the eyes of the public there really wasn't much difference - as soon as one company was seen to have "solved" a StarCraft game, anyone else coming later would have a difficult time differentiating their achievement from AlphaStar's. So there was no reason for Meta and Samsung to invest any more resources into it, as it had limited research value (AlphaStar had demonstrated what was widely possible with reinforcement learning applied to StarCraft) and limited publicity value. I also did not take your comment at all as throwing shade, so no worries about that :)
Bots are mostly optimized for bot-on-bot competition, and the Zerg bots have been weaker than Terran/Protoss for a few years. So bots aren't as well prepared against Zerg as they are against other races. Hoping for stronger showings against Terran/Protoss!
Great series 👏 Demand more thorough testing. Best of 3s on same map same faction and the player goes same economical late game build to limit test the AIs 'cook'. Can it rush? How well? Does it drop? How well? Does it do timings? How well? Ectect. It sure does well at getting slapped n stomped into an early checkmate.
I'd love to see more casts like this, but you should know that there are already commentated human vs AI games on UA-cam. I know of one on Falcon Paladin's channel and a number on Making Computer Do Things's channel, but there are plenty more if you look for them.
@@levonschaftin3676And while it's easy for a human to dismiss a lot of these states as junk, an AI has to actually learn that, so it's a much more difficult task for a computer.
I like that you did some ZeLot matches and now he's in the ASL. Focusing on some of the lesser known asl players this season might be nice, change them from NPCs into known entities. Like Sea in the first group.
I am thinking about doing a challenge where I try to beat the campaign using only ai. Haven't done c++ in over a year but I was surprised by how intuitive the bwapi is and setting it up wasn't as bad as I thought it would be.
To be honest, the only interesting thing here was the zergling concave. Other than that the bots got totally smashed, though your commentary on WHY they got smashed was very insightful. I'd be up to seeing some more, but maybe just make it monthly as bot skill progress seems to move at a glacial pace. One thing I'd like to know is whether a bot with maphack would be able to put up a better fight, because it seems they're not getting very much out of actual scouting.
I got into playing baduk/go bc of all the hype of seeing how ai would change sc2 and when it ended up anticlimactic I felt so let down I started playing go online and making go friends on discord. I'm rated 6k now and I have arty to thank because of his initial enthusiasm. I still watch your great content even though I don't play sc or sc2 anymore. You are amazing I love the content. ❤
there was a twitch channel that only showed bot matches but i think they haven't been streaming for a while. medrybw with the classic bw competitions also closed some time ago.
these videos from this tournament arent on a playlist, ive had trouble trying to find all these videos from this tournament. Naming the playlist the full names as well as the acronyms would help as well. Even some of the individual videos lack tournament or player descriptions, theres plenty of room in the titles, we dont mind.
the terran bot was mining minerals better than you think, @artosis .. believe it or not, there's a more efficient way of mining than just mineral boosting.
Bots use different optimization techniques than the humans. They tend to mine a little slower in the first minute or two of the game, but much faster later on. The two biggest techniques bots use are: - Mineral locking: Refusing to let workers waiting for a patch drift to a different patch - Order timing: When a worker arrives at a mineral, there's normally a 0-8 frame delay before they actually start mining. Bots can usually get that down to 0 frames every time by issuing a Gather command at the precisely correct frame And unlike humans they can do these for every worker throughout the whole game. So over the course of a game, a bot will probably mine faster than a pro. Artosis impressed me a lot when he played against bots and correctly deduced that one was mining faster than him just by how they were macroing in the midgame.
I always enjoy seeing what kind of AI people can program for the starcraft series. Unfortunately, this was a disappointing showing. Perhaps the matches would be more interesting if they fought A or B rank players instead. Or if the games just went past the 10 minute mark.
It's always weird to see the AI do something technical perfectly early on and think "oh this thing is gonna crush" and then it just borks up the midgame and that's that 🫤
I find Brood War AI to be really interesting, though kind of wish that they were a bit better in terms of offering a challenge to pro-level players. Feel like it could open up lots of opportunity for learning and strategy development.
I feel like with a couple medics that push in the sunks could have been deadly. you can use infinite apm to target the medic heal carefully. It's a shame we didnt get to late game. because infinite apm storm drops, reaver drops, lockdowns would be crazy. Or muta micro targetting could be absurd. Oh that scv took four bounce glaive, so i need two fewer muta tonfu ish it and they can start damaging the next scv. It takes 6 muta shot to kill a marine, so i can perfectly micro 3 muta groups. That sounds insane. Looking forward to it!
Hey RT! Was wondering, saw 8 rax hit 9/9/9 in a game between rush and zealot, zealot ended up being crushed and losing his nat hatch before he could get lings out and kill the bunker, he just didn't seem to have enough drones when the bunker started to use the usual defense, but lings weren't close to being ready yet. You mentioned, 9/9/9 being stronger against 8 rax, do you think the game i saw was an anomaly? What might cause this Thanks for the casts! love seeing AI in all games i play 😊
I'm honestly surprised how terrible the bots were. I feel like the theoretical bot capability (with today's technologies and techniques) is much higher than this but I could be wrong.
This is still considerably better than the AI built into SC2, which had a large professional team developing it. These bots are just made by hobbyists. And the only way to test the AI is to load up a game and see what it does.
I feel like these bots were not programmed by 2400+ players, but rather by RT's twitch chat. Edit: i see from some of the "AI" authors that these in fact were not created with machine learning but rather were programmed with the individual author's skill level and coding prowess. So calling them AI is a bit misleading, these are more like virtual intelligences (so called in Mass Effect) with a "best" approximation of intellect instead of something with the capability of adaptation or improvement through trial and error.
Looks funny. I used to watch a lot of "Mind Games" in SC2 which is dedicated to all the stuff about AIs in the game. But it seems like AIs in SC1 are still too weak compared to their relatives in SC2
I just played several matches against the Blizzard's AI. A couple against the BW bots, several against Original Terran, Protoss and Zerg. My conclusion is that the Original AI will create more actual, fighting units. If Protoss, it will build more Zealots and Dragoons instead of spamming templars like the BW AI does. If Terran, he would make more Siege Tanks and Marines. BW Terran would just spam Goliaths after a certain point. So, it has better macro. Especially during and after mid game. I'm guessing that the addition of anti-air units such as Corsairs, Devourers and Valks messed up the AI's resource usage. It will also make too many Lurkers, and just keep them in main until you come and destroy it. They really rushed it, and it shows.
I know this is refering to BWAPI AIs, but I'd like to make a digression on the Blizzard's AI, from my experience. Brood War's AI is pretty bugged. Terran AI doesn't build turrets at expansions at all. When playing against other Terrans, it doesn't even research siege mode or upgrade infantry. Vs Protoss, it would randomly mass firebats. On island maps, however, it is deadly. Mass goliaths with Charon boosters are OP. Zerg BW AI is sometimes good, but most often it is bad with upgrades and teching to tier 3. It can stay on hydra-lurker combo for the entire game, if it's tech is delayed. It also doesn't expand on island maps (since patch 1.10, I heard), at all. A bit less broken than Terrans, but still broken to not pose great challenge. But, if they go Hive, be prepared for some wicked attacks. Protoss BW AI are the best and most stable AI of the expansion. Unfortunately, they tech too fast to templars and build useless units like the Dark Archon and corsairs too early. Especially after the suicide 3 gate zealot all-in, the AI should have first rebuilt his army, instead. Later, they go Skytoss with Scouts, Corsairs and Carrier combo. Builds one Arbiter much later in the game. Expanions are poorly timed after the 12 Zealot push, plus it builds too many cannons, wasting minerals. In contrast, Original Starcraft AIs are much more robust. They would macro fine and wouldn't build any useless units. They have less bugs, and more consistent build orders. In Original SC, the Terran AI is the deadliest. His 14 Marine + 5 tanks push caught me off guard too many times. It upgrades slowly though, and doesn't use goliaths and vultures until much later. It uses Ghosts and Battlecruisers in the late game. Good expansions, defended with turrets and bunkers. Protoss AI would actually first rush with 3 Zealots, before going for 12 Zealot push. His later attacks focus on Speedlots and Dragoons with Singularity Charge. HTs are used mainly for defense. Late game, it goes Skytoss with Scouts and Carriers. It has good upgrades and expands well. OG Zerg AI is the weakest at the start. It rarely attacks with 12 hydras, then goes for mostly air. It upgrades and techs to Hive slowly. Thanks for reading, hopefully it helps.
Author of Stardust here, willing to answer any questions anyone has!
I am unfortunately not at all surprised by the result. All of the bots you see in this series are programmed by hobbyists, so we don’t have the kind of resources needed to do machine learning at the scale required for StarCraft. This means every high-level strategy decision has to be accounted for explicitly in the code, so it’s really easy to make bad strategy decisions and end up looking embarrasingly bad :)
I am looking forward to the upcoming games though, as Stardust’s Protoss brethren you will see in the other matchups tend to be much better at that aspect of the game.
How does one get into creating BW AI bots? Are there resources available online?
As a team member of thr original Berkeley overmind it's cool to see bwapi crowd alive and well haha
@@Strohsephmy links are being deleted but yes, look up BWAPI tutorials
your ai got wrecked lol
Is it possible to replace the base games AI with yours. Would be pretty funny to watch one of your friends get toasted thinking he was just playing against a normal computer. Especially if u bet him 100 bucks he couldnt beat a computer in Starcraft. Id love to get my hands on this AI, so cool. Nice work :D
I remember when I saw SC2 AI for terran.
It was using bunkers to increase gas mining. It was sendng SCV from CC to bunker and instant unload (it was teleporing by bunker width), then SCV was entering refinery. When with gas, it was put again to bunker, teleported and to CC.
it was doing that trick while microing every marine separatly to keep it just out of range of enemies.
that was about 20k APM.. as I remember correctly
Let's not mistake a trick implemented to AI by bot creators and tricks learned by the bot itself. The trick with the bunker was implemented by the bot creators. AI has limitless APM and the ability to be "everywhere", so such tricks are no challenge to AI.
Anyone have a link to that? sounds hilarious
That was tried in BW too, but it doesn't really work. You can gain some time in one direction by loading in one side and unloading on the other, but you gain no time in the other direction, because you always unload on the same side of a bunker. Actually, you lose time on the way back because you have to go around it. So you still need 3 SCVs to mine optimally.
Also, the trick is not really that useful. It allows you to mine gas optimally with 2 SCVs instead of 3, but to do that you have to build a bunker. And of course, a bunker is more expensive than an SCV, and an SCV is occupied while building it. It also slows down your gas mining while it is being built. So it's a significant up-front cost to ultimately get one extra SCV on minerals. I'm surprised it's worth it at all tbh.
@@EebstertheGreatBunker can be loaded from any side. So if it's built off-side from the gas path, it doesn't prevent SCVs from going to the Refinery. It reduces the distance won by doing this trick by like 1 block of movement, but cancels out all the possible drawbacks.
I helped start this scene 15 years ago, and ran the first StarCraft AI competition as part of an academic conference. The winner was the Skynet team from UC Berkeley, and there's an article about the bot and competition on ArsTechnica. Happy to answer any questions.
My thoughts on the bots:
- The bots give individual unit commands rather than using control groups, which inflates the APM a bit.
- The API isn't supported by Blizzard, and has some quirks to it, such as workers sometimes not properly labeled as idle
- The Terran bot likely bugged out, since there's the same situation as human players, where sometimes buildings don't make. And there's not a notification when this happens. This is because the API is reverse engineered, and does a best effort at providing functionality. It probably tried to start a Starport and failed without knowing it.
But it finished a refinery and then didn't mine it with anything. There is a notification when you finish a building. And what happened to the SCV building the refinery in the first place? Shouldn't it have immediately started to mine?
I agree it definitely bugged out after that, and the starport not building 100% can explain it. But the fact that it didn't mine right away is weird. Maybe it pulled the SCV from the refinery off immediately to build the depot, cause it was the closest one, so it was just bad timing?
The API does not provide callbacks for building completions or failures, or even successful starts. It's a bit tedious, but you can manually check if buildings have started.
Crown? I fuckin wish I knew abt this when I was a Slug in 2010
You should get back to organizing AIIDE. You did a great job back in 2010!
@@UCSCbweber That's annoying. The game has a notification when a building finishes, so I wonder why that wasn't relatively easy to do in BWAPI.
The zerglings concave was lit, I was expecting more intense micro, but on the strategic side those AIs were super weak.
Yeah a bit strange too that the zergling micro was superhuman but the mutalisk micro was a letdown.
@@nankamthere're many bots in SC AI League that execute Muta micro flawlessly. Apparently, not this one.
The bots don't seem to be properly abusing their superhuman abilities. Like the lack of mineral boosting noted with the Terran one, or the Zerg one not having perfect muta micro and that sort of thing. The ling concave is pretty much the best showcase there. Protos might have some insane reaver shuttle play at some point too.
I don't think reaver would be a good showcase for micro because it has a forced attack delay after being dropped. Terran units however do not have this restriction put on them, so the AI could so some magnificent tank-drop micro
@@StriKe_jk Yeah but I'm thinking of something like 4 shuttle reavers being used simultaneously
@@uxie6177@strike_jk Top bots can do a few things with Reavers which are superhuman:
- Pick up and re-drop after every shot, which lets Reavers shoot almost twice as fast (because dropping Reavers doesn't delay their attack; it sets the cooldown to a fixed value!)
- Micro any number of Reavers/Shuttles simultaneously
- Target every Reaver shot, to attempt maximizing splash damage or just avoid shooting "AI Pylons" (Decoy Pylons humans place in the natural to draw incidental Reaver shots)
On the other hand, their decisionmaking with Reavers tends to suffer due to poorly handling the impact of terrain/collisions on Reaver shots, not accurately simulating Reavers when predicting how fights will turn out, and not dealing well with the acceleration characteristics of Shuttles.
Reaver micro while also dropping storms would be incredible.
@@perfectman3077They can do it! Funnily enough, storm is actually quite weak against bots because they can instantly dodge it. Even Marines can largely survive storm with instant reactions and splitting.
My expectations of AI: AlphaStar
Reality: “why can’t we kill that overlord!!😵💫”
The build order can probably use some polish. Factory into marine only push on sunkens into academy.
It all boils down to how well versed in Starcraft the programmer is. If they implemented good strategies and micro methods into the program. A machine would execute perfectly what you tell it to do. The question is, what do you tell it to do?
Either that or it's an actual AI program that can learn and get better as it plays more games. Which would explain why it does poorly in the first few - or perhaps hundreds of - games.
Alphastar was made with machine learning. These bots use manually programmed rules.
The computer in SC1 was a real challenge as a kid, cool to see how far the bots have come!
They did a last minute-change on the lineup for the terran bot because krasi0 said his bot isn't good at TvZ. But I think he would have done way better than what Terminus did here with the super-odd fast gas into not spending it on anything.
I coded a Zerg bot some years ago and I never managed to find a way to get it to beat krasi0. And that's years ago where krasi0 still improved his bot since.
Against protoss the 8-9-off-gates were incredibly hard to deal with as Zerg with any kind of standard-build-order. A build that would survive this was obviously pretty bad against any of the more standard ones. ZvP still was my bots best matchup.
I also wonder about the Z-bots muta-micro. Mine had different micro depending on whether it was fighting moving or stationary targets. Against moving-targets it would use the move-shooting and against stationary ones the kiting-style that zerg here used.
My bot might overall be a buggy mess that loses half of it's games by crashing or something else vital stopping to work. But it sure had some neat micro-methods that I'd invite other bot-authors to use for their otherwise better bots.
It would be great to see the botting community come together to create the ultimate AI !
@@oneoranota There was a lot of "building on other's foundation" but due to it also being framed as a competition honor kinda prevented too much collaboration.
I've had experience playing with some of these bots a few years ago. My takeaway was that it wasn't terribly difficult to get a strategic win in the early to mid game but if you let them power up their macro engine, they're almost impossible to beat. Would be curious to see a series where some of these guys give the ai more of a fighting chance earlier in the game and then take on the 5000 apm macro engine.
Love the "AI show match" format, would love to see more of them!
I wrote the bot PurpleWave which should appear in one of the subsequent casts. Happy to answer questions about Brood War bots here.
Heck yeah! Thx for working on this and being open to answer questions. Is PurpleWave trained using machine learning techniques? If yes, is it augmented with any human-suggested strategies/tricks?
Oh now i see a different comment indicating no machine learning was used for these bots
@@IcthiIndeed. So PurpleWave's strategies are my best human understanding of what comprise good decisions across matchups, as best as I can capture them in code. Sometimes there are concessions like avoiding situations requiring the bot to perform techniques it's not strong at.
So these bots are not AI at all
@@valentinomanontroppo4675 AI is a lot broader than only machine learning. There are plenty of AI techniques that go into these bots.
I would be really interested in seeing a BW AI tournament being casted, or AI vs human. Something like OpenAI/Dota level- it would be great to see what strategies AI could come up with in Broodwar that defy common practice
these are soo cool, it reminds me of the time when i was watching SSCAIT (student starcraft ai tournament) on twitch for non stop action
holy crap the way it controls those zerglings to make a perfect concave protecting the ramp is something I've never seen before!
How long until the AI engines can develop literal gamer chills? BW coffee break is the best coffee break - thanks Arty
even if the bots lose it's still super impressive to me, props to all the people behind them
I really like this. In my teens I wished for better AI players for Starcraft, as I dindt have internet to play against humans. Now they are here. What a time to be alive!
Can’t wait for future Terran AI Bonjwa Flash-Drive.
😂
lol. I hope BW AI never becomes truly next level, because the meta being pushed forward by human learning is so much more interesting. Human strategies and play will always be more engaging, and the pros using machine learning and rote memorization to optimize play will be the death of BW for the fans.
@@mediocrates1937 Even if such AIs become common, I doubt anybody would turn up to watch them. Watching living beings play this clunky game at such a level is what makes it interest.
I honestly hope true AI like AlphaStar do tackle BW. They could show us things about the game we didnt know, and clue us into strategies, techniques or metas we might never even find on our own. Also, you have to consider -- not that many new people really get into BW. All the ASL players are ones who have been at it for years. There's not really a young generation right now getting into BW the same way there is for Melee. The competitive scene for BW will slowly fade away IMO (sadly) so we need to explore the game fully while we can - and we're not getting new developments in the meta or finding new micro tricks much anymore. @@mediocrates1937
I wish we could get like, some different or interesting AI opponents in the Remaster. Sometimes I want to play skirmish but don't want to get completely rekt on the ladder. But it's still the same old AI from the original game. It won't even GG when it knows it's lost. Good video anyway.
play UEDAIP, they have their own custom AI's that are really fun to fight.
@@Ryan41707I love how much of a challenge that Campaign is but the UED units are such a chore to use. They just auto-lose defensively due to their nerfed range/non siegeable tanks
Very interesting to see these! I was actually already curious about where AI is at in brood war.
Not sure if I’d like to see AI matches on a regular basis, but as an informative one-off this is really cool!
I really enjoyed this cast @artosiscast. It was interesting. I don't think I would watch lots of videos like this but seeing something new was fun
Love the vs AI series, I feel it's more of a game that falls apart because of tactical mistakes rather than micro-mistakes.
No there was a huge mistake by the AI over microing the mutas in that ZvZ. It's not just tactical issues. In the final game the AI was trying to pull marines back from a sunken...which was only reducing the damage output of the marines.
ahh yeah I remember watching the old SCAIT league on twitch way back in the day. It's really cool that people are making these things for fun. Important to know that these are AIs in the old sense of the word - like "just a set of if this, do this" instructions more or less. They aren't nearly as elaborate, powerful or complicated as the modern "AI" term (things like ChatGPT for example.)
I would LOVE to see AlphaStar for Brood War, I think there's some real potential for it to show us things about the game we didn't know, and show some strategies or techniques that we never thought possible.
Wow, looking forward to that! I like watching the SC2 AI scene (there are some strong bots there, but generally not quite of the level of pro players apart from cheese strats). This will be an interesting comparison.
Would love to see his reaction to UEDAIP and how they changed every unit.
Love this going off brand Arty bro , nice mix up adding some flavor/variation :))
I want to see an AI league where AI fight each other. Imagine that, paired with machine learning, we'd see some insane never before seen strats
That is sicks! I tried playing against some AI by using the SCHNAIL client and it was a fantastic experience.
I love this! There is so much content for StarCraft that we haven't even seen yet. I'd love to see more of these and maybe even a series on the South American scene that you mentioned during the first game @artosis
It'll be interesting when hardware for the hobbyist catches up and you can start to do insane things like learning while the match is going on
I am sooo interested in AI and game theory in general and I can't wait for the day that AI shows us how BW works on an even higher level than humanly possible. It is so interesting to me.
Looking forward to more AI matches!
I facepalmed IRL when I saw those
marines and that overlord
I am a simple man. I see schnail and I like it
i was hoping you'd post the tourney match between you and g5. youre such a great role model rt.
have you seen him raging before lol
I would like to see periodical videos on AI to see how they improve. Great video.
This is awesome to see.
Jimin's APM has to be bugged or something. I've never seen an S rank zerg with that low of an APM. If it's not bugged, then I'm actually incredibly impressed, he must be super smart and have great macro...or great cheese
He is just playing instead of spam clicking like pros do.
its definitely bugged
@@calvin8857he streams on twitch, he has like 250+ apm
Savior was the best player in the world for quite some time and he only had like 200 apm.
With SCHNAIL the games are played in 1.16 client, and on Normal speed but sped up by the API to feel as if it is Fastest so a human wouldn't notice while playing. This impacts a few things though: Watching the replay back, the replay game timer doesn't matchup with what you normally would expect, as that timer is impacted by the game speed that was played ("Normal" instead of Fastest). I believe this impacts APM as well, and I suspect this pushes player APM lower than normal because 1.) game is treating all the information as if it is Normal speed, but 2.) since the game is running as if it's Fastest thanks to the API, the human player is putting less inputs over time than they would actually do on Normal speed. So the APM shows lower because "more time is passing" by the replay timer bar, but the human is doing less actions over that longer time.
I'd be really interested in a machine learning bot, since these SCHNAIL AIs were preprogrammed. Probably they are optimized to make a single build or maybe rotate a couple, but reacting to a hydra bust... So something like AlphaStar because it'd innovate and actual players could learn from it.
Meta and Samsung were both researching AIs using reinforcement learning for Brood War. Unfortunately AlphaStar beat them to the punch and both companies pulled the plug not too long afterwards. All of the people still actively working on Brood War bots are hobbyists without the resources needed to do any kind of proper machine learning at the scale needed for StarCraft.
@@b-qj9oz Are you sure AlphaStar was trained on BW too? I've only seen it in SC2, I'd love to take a look at BW ML gameplay. And I do understand that SCHNAIL AIs were made by enthusiastic hobbyists. I did not want to throw shade, just acknowledge the limitations of the approach. (And also, I know that most of the resources needed are not for the code but for the training and amateurs just do not have the astronomical amount of money for that kind of hardware.)
@@b-qj9oz But Samsung did actually release a bot, right? In 2018? They just stopped improving it after AlphaStar was released for SC2.
@@csucskos I formulated my comment poorly, you are of course correct that AlphaStar was only trained for SC2. However in the eyes of the public there really wasn't much difference - as soon as one company was seen to have "solved" a StarCraft game, anyone else coming later would have a difficult time differentiating their achievement from AlphaStar's. So there was no reason for Meta and Samsung to invest any more resources into it, as it had limited research value (AlphaStar had demonstrated what was widely possible with reinforcement learning applied to StarCraft) and limited publicity value.
I also did not take your comment at all as throwing shade, so no worries about that :)
the series we've all been waiting for. artosis X Starcraft AIs
Very interesting to see the current progress of AI vs. humans!
There is nothing more satisfying than that ling ramp circling happening just without fuss, holy crap.
Pretty interesting. Though i thought that StarCraft A.I.s would be more developed by now.
Bots are mostly optimized for bot-on-bot competition, and the Zerg bots have been weaker than Terran/Protoss for a few years. So bots aren't as well prepared against Zerg as they are against other races. Hoping for stronger showings against Terran/Protoss!
This is awesome, RT. You know, if you just used 2 AIs and commentate using your voice in Tangia, you literally don't have to do anything.
cool series, hoping the devs will add a feature where their bots gg out when they know they've lost
I like the ai video, would love to see more from time to time
Great series 👏
Demand more thorough testing.
Best of 3s on same map same faction and the player goes same economical late game build to limit test the AIs 'cook'.
Can it rush? How well? Does it drop? How well? Does it do timings? How well? Ectect.
It sure does well at getting slapped n stomped into an early checkmate.
I'd love to see more casts like this, but you should know that there are already commentated human vs AI games on UA-cam. I know of one on Falcon Paladin's channel and a number on Making Computer Do Things's channel, but there are plenty more if you look for them.
5:54 now this is important apm for him doing his little dance
The first round was pretty interesting, they basically did the exact same moves but the AI was like a second behind on a few things.
Cool! I love that so many legendary games such as chess and go have been solved by AI but SC hasn’t. Testament to the game imho.
starcraft cannot be compared to those games. the number of states of the game far, far exceeds the other games. orders of magnitude more variability.
@@levonschaftin3676And while it's easy for a human to dismiss a lot of these states as junk, an AI has to actually learn that, so it's a much more difficult task for a computer.
This is a fun little sidebar of the normal content.
I like that you did some ZeLot matches and now he's in the ASL. Focusing on some of the lesser known asl players this season might be nice, change them from NPCs into known entities. Like Sea in the first group.
Sea is an old school legend, just below taekbangleessang back in the ProLeague days
sweet, i'm gonna enjoy playing these AI!!
I am thinking about doing a challenge where I try to beat the campaign using only ai. Haven't done c++ in over a year but I was surprised by how intuitive the bwapi is and setting it up wasn't as bad as I thought it would be.
To be honest, the only interesting thing here was the zergling concave. Other than that the bots got totally smashed, though your commentary on WHY they got smashed was very insightful. I'd be up to seeing some more, but maybe just make it monthly as bot skill progress seems to move at a glacial pace. One thing I'd like to know is whether a bot with maphack would be able to put up a better fight, because it seems they're not getting very much out of actual scouting.
Neat to see.
Wild that it had over 1400 gas with nothing to show
That ling arc on the ramp is very satisfying lol
Underwhelming performance from the AI team TBH, but I like the cast idea. Hope the other matchups will be more interesting.
I got into playing baduk/go bc of all the hype of seeing how ai would change sc2 and when it ended up anticlimactic I felt so let down I started playing go online and making go friends on discord. I'm rated 6k now and I have arty to thank because of his initial enthusiasm. I still watch your great content even though I don't play sc or sc2 anymore. You are amazing I love the content. ❤
It was a fun diversion. But the bot really disappointed in the second and third matches. Too one-sided
This is a fun character of pace.
there was a twitch channel that only showed bot matches but i think they haven't been streaming for a while. medrybw with the classic bw competitions also closed some time ago.
Blizzard should train a neural net on all the ladder games played on battle net.
It was interesting to watch!
Thanks Artosis!
these videos from this tournament arent on a playlist, ive had trouble trying to find all these videos from this tournament. Naming the playlist the full names as well as the acronyms would help as well. Even some of the individual videos lack tournament or player descriptions, theres plenty of room in the titles, we dont mind.
I ABSOLUTELY LOVE IT!
the terran bot was mining minerals better than you think, @artosis .. believe it or not, there's a more efficient way of mining than just mineral boosting.
Bots use different optimization techniques than the humans. They tend to mine a little slower in the first minute or two of the game, but much faster later on.
The two biggest techniques bots use are:
- Mineral locking: Refusing to let workers waiting for a patch drift to a different patch
- Order timing: When a worker arrives at a mineral, there's normally a 0-8 frame delay before they actually start mining. Bots can usually get that down to 0 frames every time by issuing a Gather command at the precisely correct frame
And unlike humans they can do these for every worker throughout the whole game. So over the course of a game, a bot will probably mine faster than a pro. Artosis impressed me a lot when he played against bots and correctly deduced that one was mining faster than him just by how they were macroing in the midgame.
I personally enjoy watching AI for inspirational / learning purposes.
I always enjoy seeing what kind of AI people can program for the starcraft series. Unfortunately, this was a disappointing showing. Perhaps the matches would be more interesting if they fought A or B rank players instead. Or if the games just went past the 10 minute mark.
It's always weird to see the AI do something technical perfectly early on and think "oh this thing is gonna crush" and then it just borks up the midgame and that's that 🫤
I'd love to see more videos like this. However I would prefer the bot be a little smarter.
Definitely interesting to see!
I wanna see that Zerg bot against Artosis. He is perfect at "just sitting in an arc" what should give him a big advantage
You'd be surprised by the normal SC computer players now. They do all sorts of wild stuff that human players came up with first
I find Brood War AI to be really interesting, though kind of wish that they were a bit better in terms of offering a challenge to pro-level players. Feel like it could open up lots of opportunity for learning and strategy development.
Well those AI zerglings on the ramp showcased what is possible; I guess it's like chess AI, only a matter of time really untill it becomes unbeatable.
I wonder what kinds of progress we might start to see if someone develops an effective general intelligence algorithm.
I wonder if it would have made a difference if the Marine didn't bug on the Overlord?
I feel like with a couple medics that push in the sunks could have been deadly. you can use infinite apm to target the medic heal carefully.
It's a shame we didnt get to late game. because infinite apm storm drops, reaver drops, lockdowns would be crazy.
Or muta micro targetting could be absurd. Oh that scv took four bounce glaive, so i need two fewer muta tonfu ish it and they can start damaging the next scv. It takes 6 muta shot to kill a marine, so i can perfectly micro 3 muta groups. That sounds insane. Looking forward to it!
Hey RT! Was wondering, saw 8 rax hit 9/9/9 in a game between rush and zealot, zealot ended up being crushed and losing his nat hatch before he could get lings out and kill the bunker, he just didn't seem to have enough drones when the bunker started to use the usual defense, but lings weren't close to being ready yet. You mentioned, 9/9/9 being stronger against 8 rax, do you think the game i saw was an anomaly? What might cause this
Thanks for the casts! love seeing AI in all games i play 😊
VERY cool video, I want more.
After the first ZvZ I'd expect an AlphaStar like AI to be nearly unbeatable in ZvZ.
Back in the day we used to play big game hunters map vs 7 PCs. It was fun for solid 60 minutes. 🙂
Wanna see AI puts down "A Pylon"? A stands for that, you know.
I'm honestly surprised how terrible the bots were. I feel like the theoretical bot capability (with today's technologies and techniques) is much higher than this but I could be wrong.
This is still considerably better than the AI built into SC2, which had a large professional team developing it. These bots are just made by hobbyists. And the only way to test the AI is to load up a game and see what it does.
hows the AI vs AI scene thooooo
MUCH FUN For SINGLE PLAYER WITH an AI
I feel like these bots were not programmed by 2400+ players, but rather by RT's twitch chat.
Edit: i see from some of the "AI" authors that these in fact were not created with machine learning but rather were programmed with the individual author's skill level and coding prowess. So calling them AI is a bit misleading, these are more like virtual intelligences (so called in Mass Effect) with a "best" approximation of intellect instead of something with the capability of adaptation or improvement through trial and error.
This is promising at least, it is much more versatile, that the built in AI.
That ling positioning is so satisfying
Looks funny. I used to watch a lot of "Mind Games" in SC2 which is dedicated to all the stuff about AIs in the game. But it seems like AIs in SC1 are still too weak compared to their relatives in SC2
I wonder if they would get really scary later on in a macro style game
I just played several matches against the Blizzard's AI. A couple against the BW bots, several against Original Terran, Protoss and Zerg.
My conclusion is that the Original AI will create more actual, fighting units. If Protoss, it will build more Zealots and Dragoons instead of spamming templars like the BW AI does. If Terran, he would make more Siege Tanks and Marines. BW Terran would just spam Goliaths after a certain point. So, it has better macro. Especially during and after mid game.
I'm guessing that the addition of anti-air units such as Corsairs, Devourers and Valks messed up the AI's resource usage. It will also make too many Lurkers, and just keep them in main until you come and destroy it.
They really rushed it, and it shows.
I'm interested in this.
I know this is refering to BWAPI AIs, but I'd like to make a digression on the Blizzard's AI, from my experience.
Brood War's AI is pretty bugged. Terran AI doesn't build turrets at expansions at all. When playing against other Terrans, it doesn't even research siege mode or upgrade infantry. Vs Protoss, it would randomly mass firebats. On island maps, however, it is deadly. Mass goliaths with Charon boosters are OP.
Zerg BW AI is sometimes good, but most often it is bad with upgrades and teching to tier 3. It can stay on hydra-lurker combo for the entire game, if it's tech is delayed. It also doesn't expand on island maps (since patch 1.10, I heard), at all. A bit less broken than Terrans, but still broken to not pose great challenge. But, if they go Hive, be prepared for some wicked attacks.
Protoss BW AI are the best and most stable AI of the expansion. Unfortunately, they tech too fast to templars and build useless units like the Dark Archon and corsairs too early. Especially after the suicide 3 gate zealot all-in, the AI should have first rebuilt his army, instead. Later, they go Skytoss with Scouts, Corsairs and Carrier combo. Builds one Arbiter much later in the game. Expanions are poorly timed after the 12 Zealot push, plus it builds too many cannons, wasting minerals.
In contrast, Original Starcraft AIs are much more robust. They would macro fine and wouldn't build any useless units. They have less bugs, and more consistent build orders. In Original SC, the Terran AI is the deadliest. His 14 Marine + 5 tanks push caught me off guard too many times. It upgrades slowly though, and doesn't use goliaths and vultures until much later. It uses Ghosts and Battlecruisers in the late game. Good expansions, defended with turrets and bunkers.
Protoss AI would actually first rush with 3 Zealots, before going for 12 Zealot push. His later attacks focus on Speedlots and Dragoons with Singularity Charge. HTs are used mainly for defense. Late game, it goes Skytoss with Scouts and Carriers. It has good upgrades and expands well.
OG Zerg AI is the weakest at the start. It rarely attacks with 12 hydras, then goes for mostly air. It upgrades and techs to Hive slowly.
Thanks for reading, hopefully it helps.
Havnt seen you do any ai stuff in yeeeeeaarrrsss i really liked it then i an really like it now always kkeeepp it commijnnnggg
ai's vs ai's are the best.
Just goes to show that even with 7000+ apm... you have to actually do things that are useful to win, not just spam apm
So at 12:45 protoss had 3300 apm and not a single real command was sent. that's how apm describes how good a player is in skillwise.
Fascinating vid