@@sebber7992 Dear Harstem. Today I had a wild game against a protos and I just don't know how to beat protos as a zerg. I had 20x higher APM than the protoss (heck, he didn't even fastmine, just left his drones on automine like a diamond player) and I out-eco'd him due to this I just couldn't win any fights. I spread my creep across the entire map, and he didn't even make an effort to clear it up. The protoss killed like 20 creepstumors during the entire game. Despite me microoing my units back and forth, burrowing and so on my army lost against a stupid protoss basically A-Moving (barely 400 APM vs my 10.000) into my army. And the most offensive thing about this were the disruptor shots. If I can't dodge them properly with my 10.000 APM how is zerg supposed to dodge them at all? It feels like protoss is just A-Move and Disruptorshot, with no counterplay. Please review and advice - is it is imba or do I suck?
That drone micro to get minerals faster is very interesting. I once saw a game where a Terran bot would purposefully build bunkers in their own mineral line and with several thousand APM would load individual SCVS into the bunkers and immediately unload them out the other side when crossing the mineral line, effectively having each worker 'teleport' through the bunker instantaneously to drop off their mineral cargo and get back to the mineral line to continue harvesting. It actually increased the mining speed by a significant amount, quite quickly paying off the cost of the bunker and then earning a steady profit.
I'm not sure if you are talking about my bot MicroMachine or a different bot that I never saw, but MicroMachine builds bunkers between the CC and double geysers to saturate them with 4 workers instead of 6.
@@BlindMidget I may have misremembered it doing it with minerals instead of Gas. ...Could you do both? 2 workers per gas and also speed up mineral mining, or even require fewer SCVs per base to optimally mine minerals?
@@Supergoon1989 Yes that could be possible, however building a bunker in the mineral line reduces the mineral income while the bunker is not finished, so it would take more time to pay for itself. And it would be a bit hard to determine which mineral patchs are close enough to the bunker to force the workers to use the bunker instead of just walking.
I'm one of the authors of this bot, thanks to Harstem for this video! Eris can tech to later in the game with spellcasters / broods / spores etc but Harstem made her panic, hence it was trying to constantly pump out units to stay alive Hopefully it be more competitive next time :)
Is the scouting optimized? He could have killed Eris with his first Adept push and Eris seemed to not be prepared for that. The bot seems fun. I'm happy that it's a Zerg bot :)
@Karttibone In the most recent version of Eris we have improved reactions to Adept timings which may have helped. The verion Harstem played here is a couple months old Having fun is our number one priority :)
I really wish we could have kept a version of alphastar on the ladder, or have the ability to face alphastar in customs, this was really cool to watch though. Great stuff as always
The problem with that is that the current bots are better than Alf* so there would be no point. It doesn't keep improving when released on ladder, the version that ladders is usually a snapshot of the current learning progress. It just isn't useful for training the bot to have it learn from random matches since the volume of matches is to low and the variety of strategies used by players is to variable to get improvement.
@@Konnkord Yeah but then he could just play against one of the existing bots. It isn't that different an experience. I'm sure some of the bots can be APM limited as well if that is the issue.
Dear Harstem, I recently played against this progamer human playing protoss and he beat me hard. He had only an average over 400APM while I was getting 10k APM. Protoss is too OP, just immortals and disruptors are OP, none of my zerg units can beat this human. No matter how much more APM I got over him, he was just able to use 400APM and beat my super micro army. Is this IMBA or do I suck? - Eris Server EU, League Grandmaster. 🤣
Fun game, As a note while a lot of bots do not do learning in the same way as AlphaStar we have different ways of "learning" so while we do script a lot of the bots in order to massively speed up development and reduce the needed computation costs we do have some things that do learn, such as keeping track of wins and losses vs another bot. These are then factored into what build is selected. Not every bot does this but quite a few of the top bots do
@@shadowlordalpha I was referring neural networks and its automatic self-balancing based on lot of data. Useful tool, but I would say that naming it as AI was a bad idea. People got wrong impression. I was trying to say, that I consider program more intelligent if there was some intelligence used in process of its making rather than just crunch of a lot of data. Btw I don't think there is actual singular definition of AI.
@@lordook5413 Normally the definitions are simplified down to a program that can solve a problem in a "smart" manner, though this also includes most algorithms or anything that takes in and does something with external data, as opposed to always saying go left, though even that could be considered one as well depending on how its used. I do agree though its a dumb definition For Neural Networks we don't define the algorithm itself but we do define what the inputs and outputs are as well as the rules that it is trying to maximize. So while you can think of it a lot like statistics its very different in the fact that it could get stuck at a local max instead of the global max or it could come to some outcome that you never expected. For example if you give it the rule to finish a SC2 game as quickly as possible you may think it will build an army and attack as quickly as it can, something like a roach rush or drone rush. However what you are more likely to get is the AI killing its own hatch right at the start of the game because that ends the game faster then building an army and trying to kill the opponent.
@@shadowlordalpha So I as "intelligent" being would press F10 then click surrender. :) Damn I would not like your AI drive cars :D.. I know it is tempting to use Neural Networks for anything that can be trained by data. It is something like fashion nowadays. Like it is a stamp of quality. I think it makes programmers a bit lazy. And data can be flawed as well. I watched a lot of Alfastar matches. It was not much intelligent. As time goes I am more and more afraid that this style of creating programs prevails so much, that we won't be able to fix problems, debug and humanity and IT will get stuck in this loophole. It is MY fear. Maybe unfounded but maybe not. So I give a praise for those who just don't blindly jump on this train. Keep a great work.
This was a great video, love how much different content you provide all centered around SC2, like it's fanatstic, way more effort and production value than any other SC2 player. You're really really good at this and with Hamster you're golden!
@@jli8225 Yeah but now that it is smart they could try and remove the apm limit. Would be quite interesting to see how it compares to these other bots.
@@nicolaunionsspezialfiliall3962 Alphastar would not be better with its APM unlocked in its current iteration/build. It "learned" to be good with its APM capped, it wouldn't know to use the extra APM unless they completely retrained it. That's just not how current AI work.
@@rayzerot Wasn't there a previous version that would save up apm and use it in fights? I'm pretty sure it would at least micro a lot better than it currently is. Maybe the strategies wouldn't be optimised for high apm but saying that it couldn't use them at all seems od.
As an extremely unskilled player, the idea of making a bot to fight other bots sounds really fun! Cool tricks here but hammers home how important build orders and responses are. The micro can help a huge amount in an even fight, but the adept timing was all that was needed for a kill.
@@N0noy1989 they would wreck them then... check how they used machine learning on Dota2. Bots smashed pros in 1 v1 and trashed them 100% od the time. even started baiting them with low hp. But was still loseing in 5 v5 comabt. PPls just was forced to roam as 5 and this is how they won.
6 minuets in and the fact that the ai speed mines not just the main, not even also the expansion but also the 3rd base all at once. Yet still manages to pull workers from the 2nd base, micro roaches and burrow them is absolutely insane lol.
As someone who designed a fairly okay-ish StarCraft 2 bot as well, here are some couple of comments explaining certain things about its behaviour. 1. The thing that is clearly hidden to everyone is the fact that bots on StarCrat 2 AI Ladder are made to play against other bots. With that being said, I mean that they do not play on the "real time" scale, but all of these bots are configured to just plug in as many commands as they want inside a single frame, and when both bots finished putting all the commands, the frame passes (thus resulting in whole minutes of waiting just to process one frame of a game), unlike a multiplayer game, where the time passes regardless if you did a command or not. Since you were able to play against the bot on "real time", it means that you played against the much weaker version of this bot, which may indirectly skip certain commands that it would normally do. 2. The APM is really high, because in a normal game, you'd normally choose a group of units, and put them to do an action, this counting as one action being made. However in the API provided, if you do the same, it counts as "n" actions, where n=number of units selected, this fact alone increasing the APM drastically. However it is fair to say that the bot does a lot of actions as well, this being another reason why the APM is that high. 3. At some point near the end of the video you mentioned that you noticed the bot didn't split the army until you destroyed a creep tumor. In code, it is fairly hard to make a set of conditions where you explain when an enemy unit wants to attack you, so most likely it was set to a trigger which detects if you are are attacked inside a base, to go and defend that base. 4. You also mentioned about the part where you cannot really catch a zerling, unless you corner it in such a way that it blocks itself in the terrain. Even if the API does provide ways to check if there is terrain somewhere, it is hard to "evaluate" the shape of the terrain such that you have a better understanding of how to move the units. 5. The fact that the creep tumors weren't spreading might be either that they are programmed to spread only at maximum possible distance, or might be a reference to the point of 1., having too many other things to do, and basically skipping multiple frames the spreading. 6. One reason for why the Disruptor shots were that effective, might be that the units were coded to check "local" short-sighted danger, and when the danger of the shot was located, it was basically too late because it was not possible to outrun it anymore. (That's my intuition, so do not take it as a fact) Anyway, great video! It was enjoyable to see the units kite so well and to have a formidable opponent.
My first experience with SC2 was actually watching some AlphaStar games, super cool to see more bots playing, even if it isn't the machine learning kind it's still interesting to see what SC2 will look like in 20 years when people figure out how to add cyber augments to their body! I for one am terrified of a world where Cyber-Serral exists and responds to the Overmind, unlike the current Serral of course.
at about 11:05 you can actually see the bots main base double evo chamber sim city has bricked the high speed mining micro and is actually resulting in worse mineral mining, maybe you can tell the bot programmer about this?
If I made a bot I would program it to respond in kind to BM, like if you say 'OUT!' or some variation is would spit back one of a few hundred pre-made responses like 'git gud, bitch-boi!' or something along those lines :P
having WAYY to much fun watching the drone chase that probe around in the beginning. RUN HIS ASS DOWN and you cut away on the first hit when he does all that work. nobody gets to recognize that epic chase for the first bite.
It's always cool watching stuff like this and seeing bots push the limits of what's possible. There are similar reasons for why TASes are often really fun to watch.
There was similar thing to the drone mineral micro in Warcraft III, back in very old versions of game, you could place buildings between the main base building and the goldmine such that it restricted the way workers could move, which somehow sped them up. Thanks to this, you could saturate a gold mine with 4 workers instead of 5. Then devs removed it, and for a long time you could micro the workers similar to the method displayed in this video (I've seen people do it, just without the queueing - shift button, and only one way, not bidirectional). I don't know if it's still in the game as I don't play it but I might have heard that they tried removing the micro and now it only works in certain specific situations where the goldmine is diagonally to the left and up in respect to main base building. Also, there is a weird thing where the workers traveling between a goldmine and a main base building have varying movespeed depending on distance between the two.
this is excellent content - seems like its very interesting for you too. Please make more. its really interesting from the consideration of human vs ai in starcraft; similarly resourced (at least relatively when considering resources allocated to AlphaStar) - would a human with unlimited financial resources + a team of highly intelligent people coaching and otherwise assisting be able to beat AlphaStar… I was typing this before the end where you said most are diamond league, poor bots
Very interesting to watch. Two things come to mind: Those disruptor shots were brutal. Every time you fired off a Disruptor, the bot lost a group of units. It might very well have lost due to that alone. Why is is the bot so bad at using active tumors when they're ready? That seems like a relatively easy thing to program, especially compared to all the wacky splitting and stutter step micro that the bot can do. Real players struggle with tumors because they're using their APM on everything else instead, but that obviously isn't an issue for the bot.
If you ever wanted to play against SC1 (broodwar 1.16) AI check out SCHNAIL. It doesn't support Starcraft remastered on battlenet but does have very easy setup and custom keybinds. I'll put the best bots to play against below. Team games don't work for 95% of the bots. Please play 1v1 and on the maps on SCHNAIL. Terran: Krais0, adias, Hao Pan Protoss: PurpleWave, McRave, Stardust, Locutus Zerg: Monster, Steamhammer, McRaveZ. Once again use only the maps provided with SCHNAIL in a 1v1 for the best results. Also these bot's play a different meta than SC1 human play. More of a bot v bot meta because they can be more cost effective with their units via micro/macro. Also another thing is they don't cheat, they can only "access" enemy player info from units they can see (Aren't in the fog of war) once they are visible they can be accessed. Some bots will "Lag out" for the first 0-10 seconds, just give them some time. Lastly most have "learning" and will decide the best builds based on previous performance against you. So keep playing them for different results. Bot APM can be ridiculous but have can have bad decision making. Happy hunting.
its impressive what the people did with this bot but i helped me just to realize how amazing the archivement from alphabet was to destroy the best players with a limited apm bot
10:27 & 10:33 are showing (new) masterful use of changelings to box in a disruptor, I gotta say this revealed more new tactics than anything i remember AlphaStar using. What a cool bot this is - is this Scarlett's bot or someone else that she helped make the rules for? Either way really really cool seeing all these new strats (burrow used as in-fight technique to dodge shots, changeling masked inside army entire time moving with army, etc)
As a programmer (not a progamer lol), I found this super interesting. I wish AlphaStar and bots that use deep learning were more a part of the bot community though
I think I know how the computer gets more minerals. I noticed immediately how the number of workers on any hatchery is constantly fluctuating. When the number drops it must be the computer reassigning the drone to the nearest mineral patch that will be available exactly when the drone will arrive back at the mineral line. So it is not technically mining for a micro-second while it is reassigned. Every time it does this it saves that drone from walking from one patch to another. It could even choose between two patches that are next to each other based on which one will require the drone to wait an extra 1/100 of a second for the last drone to finish mining.
I didn’t know this existed, but that’s such like a cool thesis project for somebody studying ai or something. Normally I just watch for the hilarious things you say, but this is really ill
@harstem at 8:40 ''this makes no sense'' its to do with the acceleration/deceleration values that the probe scv and drone all share, while the acceleration is 1000 like alot of non massive/non air units, giving them virtually instant full speed movement, the deceleration is the time/distance the workers/other units take to go from full speed to a complete stop, the harvest ability (what you refer to as return cargo) has a flag in the return cargo half of the harvest ability, which is actually 2 abilities, known as worker attack/harvest and return/return cargo, known as ''deceleration'' which allows an ability or attack to be used while in a ''deceleration'' this is most commonly seen with air units that appear to ''slide'' when given attack commands such as banshees and tempests shortly after a move command. the ai is exploiting deceleration by using the move commands, to prevent the drones from ever truly stopping while avoiding the issue of the return cargo ability from being used while moving, since it cannot be used while moving, but since deceleration takes a certain amount of distance to finish happening, and since return cargo has a specific range, this allows faster harvesting, which is pretty cool, but would be fixed by simply removing or altering workers deceleration speed, or changing return cargos deceleration flag from negative to positive.
would it be possible to play archon with one of these bots? Imagine designing a macro bot to execute a sharp timing and then just controlling the units... or a micro bot that always kites units/scouts/stops pushes while you practice a build order...
Nope. Blizzard would have to open an archon bot server, and not happening. Plus, the bots can micro at 10,000 APM. They just don't have strategic vision. Maru is a tiny bit like a bot that way, in that a Zest or SoS can outstrat. And machine learning over lots of games (SC2 is harder computationally than chess or go) needed. From my chess bot days, the challenge is to correctly "chunk" the data to learn many thousands of patterns. There are ways.
Some bots (mine included, named MicroMachine and open source) have an archon mode implemented in. With the arcon mode activated in MicroMachine, it will play normally but will not issue commands to units that are selected. So that way a player can control whatever he likes, but as soon as he deselects some units, then MicroMachine will do whatever it likes with them and might not understand the goal that the human had in mind. Still fun though, but can only be played against other bots unfortunately.
Interesting how the Changelings really mess up your micro. Essentially, the bot is using the Changelings to either stop you from getting closer or to help trap you from moving away
The highest ranked bots are Eris (Zerg), NegativeZero (Protoss) and my own bot MicroMachine (Terran). They all have their own tricks and quirks and even the meta is different from the human ladder!
As you said "the bot didn't try defend that base" my phone registered it and tried to correct it as "book Uber to home" or whatever and I was like "come on. I wasn't even talking. Just let me watch, please :( "
So when Blizzard makes the AI, do they limit the number of actions per minute to something that's possible for a human? It's interesting to see how it executes what it knows how to do, but then kind of falls apart when someone does something creative to it. You could imagine it targeting each interceptor with the exact number of units necessary to kill it, like focus fire but en masse. Warp in or do a BC flyby and they probably fall apart. In any case, it's obviously easier to maintain vision of the map when you have four mice to control your units.
That APM hitting 30k as the bot quite literally micros every single individual unit in it's army multiple times a second, while also microing all 70 workers is just... jaw dropping. One of those points where you really know you're playing a bot, not a player.
I never understood this thought process. Of course it can do that? That's like saying seeing someone put on a shirt is jaw dropping. It's something that is the baseline expectation of our capability.
I just discovered the AI SC2 Bots few days earlier and I was really impressed of them to be honest. Maybe due to the fact that I'm a developer ^^ But I'm really surprised that you have suceed to defeat it so easily. Could you provide the amount of kill of each disrupters please ? Excited to watch an other game like that after developpers patch the disrupters management to be honnest :) GG anyway
Beating grandmasters with stupid stuff idea: Fastest Possible Twin Adept Rush I've terrorized diamond NA players this rush about 250 times over the last year and am 4000-4150 MMR with it, and 3500-3900 when playing standard or other cheeses. Build is: 12 pylon on low ground 13 gateway 15 assimilator, but prioritize mineral mining 16 gateway 17 cybernetics core 18 pylon 20 2x Adepts and 2x chrono (should have 50 gas as 2nd gate, pylon, and cyber core finish) 24 nexus 26 3rd adept Adepts arrive at opponents natural at 2m 34s. The build works in all matchups but is best pvt as it can delay a low ground CC for over a minute. My most common follow up is a 2 base chargelot all-in with a prism and 9 gateways that hit at 5m 17s.
Roaches and hydras are cool and all, but why this bot didn't get any infestors? I feel like that's a major advantage of bots to controll individual units and what better than to control spellcasters in that case?
The opposite of accelerate is accelerate. Accelerate simply means to change velocity over time. Some people use the word "decelerate" to mean slow down, but this one of those commonly used yet improper words, as even "deceleration" is still acceleration.
Acceleration means to change velocity over time when it is used strictly in the world of physics. And deceleration is (from what I understand) not a term in physics. But in general language accelerate means to speed up and decelerate means to slow down. I guess this means that the opposite of accelerate is decelerate unless it is used in the world of physics where it has no perfect opposite - in which case the opposite of accelerate is accelerate like you said.
Is it possible to play vs ai bot im custom games to train a specific match up. I quitted sc2 because I could not increase my skills by laddering only. But I would love to play vs alphastar in custom games to train a specific matchup
I was curious about why the rar file is in Dropbox instead of GitHub and why there are not a README and LICENSE files so I checked the rest of the files to give it a watch. I just went to the folder of the Ketroc bot's libraries and found it is using log4j library (last update of this files on June) and then get some Vietnam flash backs. So I suggest you not to run specifically Ketrock bot. Here is an scary and detailed explanation about that exploit: th-cam.com/video/BhEZieTl0TI/w-d-xo.html
Bots have to have special thought for (human easy) details like - can't see enemy unit - must not exist. Else the bot will charge up ramps and into tanks. But several ways to fake human memory.
That was intentional. However the problem was is it broke the interaction with the manual drone mining, making the drones not mine anything at all. One of the two need to be disabled or modified/improved (or at least when they overlap)
Did you see that? His apm is so high because he continues to do the worker trick throughout the whole game. That’s why the income is more because he’s doing it with the workers. Or not he, it.
The reason it is mining better than you with less workers is because the drones are being rallied infront of the mineral patches which removes the deceleration. His drones are just moving faster. That would be my best guess from what I observed. I used to watch stephano do the same thing in the early game to get a few extra minerals. Doing this is impractical as the game goes on but it seems to maintain it on all the bases through out the game. would be crazy if humans could do the same things. Starcraft would be so different XD
Typical 500 apm protoss winning against a 10000 apm zerg
Ok, that was funny. Props
smh protoss imba
Waiting here for the bot to drop his Imba complain form...
You need to play 20 times faster than Toss and still lose.
@@sebber7992 Dear Harstem. Today I had a wild game against a protos and I just don't know how to beat protos as a zerg. I had 20x higher APM than the protoss (heck, he didn't even fastmine, just left his drones on automine like a diamond player) and I out-eco'd him due to this I just couldn't win any fights. I spread my creep across the entire map, and he didn't even make an effort to clear it up. The protoss killed like 20 creepstumors during the entire game. Despite me microoing my units back and forth, burrowing and so on my army lost against a stupid protoss basically A-Moving (barely 400 APM vs my 10.000) into my army. And the most offensive thing about this were the disruptor shots. If I can't dodge them properly with my 10.000 APM how is zerg supposed to dodge them at all? It feels like protoss is just A-Move and Disruptorshot, with no counterplay. Please review and advice - is it is imba or do I suck?
That drone micro to get minerals faster is very interesting. I once saw a game where a Terran bot would purposefully build bunkers in their own mineral line and with several thousand APM would load individual SCVS into the bunkers and immediately unload them out the other side when crossing the mineral line, effectively having each worker 'teleport' through the bunker instantaneously to drop off their mineral cargo and get back to the mineral line to continue harvesting. It actually increased the mining speed by a significant amount, quite quickly paying off the cost of the bunker and then earning a steady profit.
That's awesome
I'm not sure if you are talking about my bot MicroMachine or a different bot that I never saw, but MicroMachine builds bunkers between the CC and double geysers to saturate them with 4 workers instead of 6.
@@BlindMidget I recognise the name! yes, I think it was Micromachine! You made it!?? That's awesome! :D
@@BlindMidget I may have misremembered it doing it with minerals instead of Gas. ...Could you do both? 2 workers per gas and also speed up mineral mining, or even require fewer SCVs per base to optimally mine minerals?
@@Supergoon1989 Yes that could be possible, however building a bunker in the mineral line reduces the mineral income while the bunker is not finished, so it would take more time to pay for itself. And it would be a bit hard to determine which mineral patchs are close enough to the bunker to force the workers to use the bunker instead of just walking.
I'm one of the authors of this bot, thanks to Harstem for this video!
Eris can tech to later in the game with spellcasters / broods / spores etc but Harstem made her panic, hence it was trying to constantly pump out units to stay alive
Hopefully it be more competitive next time :)
Is the scouting optimized? He could have killed Eris with his first Adept push and Eris seemed to not be prepared for that. The bot seems fun. I'm happy that it's a Zerg bot :)
@Karttibone In the most recent version of Eris we have improved reactions to Adept timings which may have helped. The verion Harstem played here is a couple months old
Having fun is our number one priority :)
Well done sir! How can I play against it?
@@atutor1 See Ketroc's video here th-cam.com/video/ALfoEsyb54g/w-d-xo.html
Nice job your bot is impressive ! Do you think it can be possible to have Google Alphastar to play in the ProBots competition ? Or it's not possible ?
Watching that 10000 APM micro, I feel that is how the Zerg Swarm SHOULD behave. It's so skittery and insectoid, and also combat-optimal.
My idea of the Zerg race did feel very much at home seeing that APM in the thousands.
APM exceptional
Movement optimal for survival
Will consume essence, spin new strands, improve swarm
Carpal tunnel strain evolved.
Must eliminate. Ineffective.
The machine overlords will look kindly on this in the future. Perhaps they'll put you in a matrix where it's always the void ray balance patch
Machine overlords sure, but what about machine overseers?
You mean this matrix?
That sounds awful. What's the alternative, the WoL brood lord era? haha
I really wish we could have kept a version of alphastar on the ladder, or have the ability to face alphastar in customs, this was really cool to watch though. Great stuff as always
Ever heard of MaxPax?
The problem with that is that the current bots are better than Alf* so there would be no point. It doesn't keep improving when released on ladder, the version that ladders is usually a snapshot of the current learning progress. It just isn't useful for training the bot to have it learn from random matches since the volume of matches is to low and the variety of strategies used by players is to variable to get improvement.
@@ruukinen I think he wants to play for his own enjoyment, not to teach the bot his strats.
@@Konnkord Yeah but then he could just play against one of the existing bots. It isn't that different an experience. I'm sure some of the bots can be APM limited as well if that is the issue.
@@ruukinen Yes, they can. And then they'd only use speed mining in the early game, as it doesn't have anything better to do.
Dear Harstem,
I recently played against this progamer human playing protoss and he beat me hard. He had only an average over 400APM while I was getting 10k APM. Protoss is too OP, just immortals and disruptors are OP, none of my zerg units can beat this human. No matter how much more APM I got over him, he was just able to use 400APM and beat my super micro army. Is this IMBA or do I suck? - Eris Server EU, League Grandmaster.
🤣
XD
xD
Comment approved
Fun game,
As a note while a lot of bots do not do learning in the same way as AlphaStar we have different ways of "learning" so while we do script a lot of the bots in order to massively speed up development and reduce the needed computation costs we do have some things that do learn, such as keeping track of wins and losses vs another bot. These are then factored into what build is selected. Not every bot does this but quite a few of the top bots do
All AI "learning" is just weighted statistics. Change my mind. I much prefer intelligently designed algorithms.
@@lordook5413 are we talking the perceived definition of ai or the actual one? They are, surprisingly, very different.
@@shadowlordalpha I was referring neural networks and its automatic self-balancing based on lot of data. Useful tool, but I would say that naming it as AI was a bad idea. People got wrong impression. I was trying to say, that I consider program more intelligent if there was some intelligence used in process of its making rather than just crunch of a lot of data. Btw I don't think there is actual singular definition of AI.
@@lordook5413 Normally the definitions are simplified down to a program that can solve a problem in a "smart" manner, though this also includes most algorithms or anything that takes in and does something with external data, as opposed to always saying go left, though even that could be considered one as well depending on how its used. I do agree though its a dumb definition
For Neural Networks we don't define the algorithm itself but we do define what the inputs and outputs are as well as the rules that it is trying to maximize. So while you can think of it a lot like statistics its very different in the fact that it could get stuck at a local max instead of the global max or it could come to some outcome that you never expected. For example if you give it the rule to finish a SC2 game as quickly as possible you may think it will build an army and attack as quickly as it can, something like a roach rush or drone rush. However what you are more likely to get is the AI killing its own hatch right at the start of the game because that ends the game faster then building an army and trying to kill the opponent.
@@shadowlordalpha So I as "intelligent" being would press F10 then click surrender. :) Damn I would not like your AI drive cars :D.. I know it is tempting to use Neural Networks for anything that can be trained by data. It is something like fashion nowadays. Like it is a stamp of quality. I think it makes programmers a bit lazy. And data can be flawed as well. I watched a lot of Alfastar matches. It was not much intelligent. As time goes I am more and more afraid that this style of creating programs prevails so much, that we won't be able to fix problems, debug and humanity and IT will get stuck in this loophole. It is MY fear. Maybe unfounded but maybe not. So I give a praise for those who just don't blindly jump on this train. Keep a great work.
This was a great video, love how much different content you provide all centered around SC2, like it's fanatstic, way more effort and production value than any other SC2 player. You're really really good at this and with Hamster you're golden!
I think this bot hits the skill ceiling when it comes to micro, unfortunately not decision making though. Imagine if Alpha used tricks like this.
Alphastar is limited on how much APM it can use i believe, as the devs want it to be smart rather than brute force.
We like to say sc2 bots are like having each individual unit micro'ed by a pro, but all the pros are being told what to do by a bronze leaguer.
@@jli8225 Yeah but now that it is smart they could try and remove the apm limit. Would be quite interesting to see how it compares to these other bots.
@@nicolaunionsspezialfiliall3962 Alphastar would not be better with its APM unlocked in its current iteration/build. It "learned" to be good with its APM capped, it wouldn't know to use the extra APM unless they completely retrained it. That's just not how current AI work.
@@rayzerot Wasn't there a previous version that would save up apm and use it in fights? I'm pretty sure it would at least micro a lot better than it currently is. Maybe the strategies wouldn't be optimised for high apm but saying that it couldn't use them at all seems od.
As an extremely unskilled player, the idea of making a bot to fight other bots sounds really fun!
Cool tricks here but hammers home how important build orders and responses are. The micro can help a huge amount in an even fight, but the adept timing was all that was needed for a kill.
Really need machine learning for the bots to be able to match human pros.
@@N0noy1989 they would wreck them then... check how they used machine learning on Dota2. Bots smashed pros in 1 v1 and trashed them 100% od the time. even started baiting them with low hp. But was still loseing in 5 v5 comabt. PPls just was forced to roam as 5 and this is how they won.
If he didn't go disruptions, bot would annihilate him.
@@KrecikBobasek OpenAI trashed the 2 time TI winning OG in 5v5 xD
6 minuets in and the fact that the ai speed mines not just the main, not even also the expansion but also the 3rd base all at once. Yet still manages to pull workers from the 2nd base, micro roaches and burrow them is absolutely insane lol.
As someone who designed a fairly okay-ish StarCraft 2 bot as well, here are some couple of comments explaining certain things about its behaviour.
1. The thing that is clearly hidden to everyone is the fact that bots on StarCrat 2 AI Ladder are made to play against other bots. With that being said, I mean that they do not play on the "real time" scale, but all of these bots are configured to just plug in as many commands as they want inside a single frame, and when both bots finished putting all the commands, the frame passes (thus resulting in whole minutes of waiting just to process one frame of a game), unlike a multiplayer game, where the time passes regardless if you did a command or not. Since you were able to play against the bot on "real time", it means that you played against the much weaker version of this bot, which may indirectly skip certain commands that it would normally do.
2. The APM is really high, because in a normal game, you'd normally choose a group of units, and put them to do an action, this counting as one action being made. However in the API provided, if you do the same, it counts as "n" actions, where n=number of units selected, this fact alone increasing the APM drastically. However it is fair to say that the bot does a lot of actions as well, this being another reason why the APM is that high.
3. At some point near the end of the video you mentioned that you noticed the bot didn't split the army until you destroyed a creep tumor. In code, it is fairly hard to make a set of conditions where you explain when an enemy unit wants to attack you, so most likely it was set to a trigger which detects if you are are attacked inside a base, to go and defend that base.
4. You also mentioned about the part where you cannot really catch a zerling, unless you corner it in such a way that it blocks itself in the terrain. Even if the API does provide ways to check if there is terrain somewhere, it is hard to "evaluate" the shape of the terrain such that you have a better understanding of how to move the units.
5. The fact that the creep tumors weren't spreading might be either that they are programmed to spread only at maximum possible distance, or might be a reference to the point of 1., having too many other things to do, and basically skipping multiple frames the spreading.
6. One reason for why the Disruptor shots were that effective, might be that the units were coded to check "local" short-sighted danger, and when the danger of the shot was located, it was basically too late because it was not possible to outrun it anymore. (That's my intuition, so do not take it as a fact)
Anyway, great video! It was enjoyable to see the units kite so well and to have a formidable opponent.
My first experience with SC2 was actually watching some AlphaStar games, super cool to see more bots playing, even if it isn't the machine learning kind it's still interesting to see what SC2 will look like in 20 years when people figure out how to add cyber augments to their body! I for one am terrified of a world where Cyber-Serral exists and responds to the Overmind, unlike the current Serral of course.
Oh this seems to be an interesting video. Very curious about the bot tricks ^^
at about 11:05 you can actually see the bots main base double evo chamber sim city has bricked the high speed mining micro and is actually resulting in worse mineral mining, maybe you can tell the bot programmer about this?
29,000APM Dear God lol
15:17 Wow what a BM. Good thing the bot is very well mannered 😂😂
If I made a bot I would program it to respond in kind to BM, like if you say 'OUT!' or some variation is would spit back one of a few hundred pre-made responses like 'git gud, bitch-boi!' or something along those lines :P
Psh, I get 29,001 APM in my games.
I've seen 180k, however the game start to have trouble processing all the actions and starts to skip some around 100-150k.
Jesus, those changeling moves are beyond obnoxious. I love it.
When they bogged down that Disruptor, like wtf
You're a content MACHINE!
Macroing like a teabag XD
Very cool idea, hope we can see more of it!
having WAYY to much fun watching the drone chase that probe around in the beginning. RUN HIS ASS DOWN and you cut away on the first hit when he does all that work. nobody gets to recognize that epic chase for the first bite.
It's always cool watching stuff like this and seeing bots push the limits of what's possible. There are similar reasons for why TASes are often really fun to watch.
There was similar thing to the drone mineral micro in Warcraft III, back in very old versions of game, you could place buildings between the main base building and the goldmine such that it restricted the way workers could move, which somehow sped them up. Thanks to this, you could saturate a gold mine with 4 workers instead of 5. Then devs removed it, and for a long time you could micro the workers similar to the method displayed in this video (I've seen people do it, just without the queueing - shift button, and only one way, not bidirectional). I don't know if it's still in the game as I don't play it but I might have heard that they tried removing the micro and now it only works in certain specific situations where the goldmine is diagonally to the left and up in respect to main base building. Also, there is a weird thing where the workers traveling between a goldmine and a main base building have varying movespeed depending on distance between the two.
I want photos of the disguise you used to sneak into Bot rooms
this is excellent content - seems like its very interesting for you too. Please make more.
its really interesting from the consideration of human vs ai in starcraft; similarly resourced (at least relatively when considering resources allocated to AlphaStar) - would a human with unlimited financial resources + a team of highly intelligent people coaching and otherwise assisting be able to beat AlphaStar…
I was typing this before the end where you said most are diamond league, poor bots
Very interesting to watch. Two things come to mind:
Those disruptor shots were brutal. Every time you fired off a Disruptor, the bot lost a group of units. It might very well have lost due to that alone.
Why is is the bot so bad at using active tumors when they're ready? That seems like a relatively easy thing to program, especially compared to all the wacky splitting and stutter step micro that the bot can do. Real players struggle with tumors because they're using their APM on everything else instead, but that obviously isn't an issue for the bot.
If you ever wanted to play against SC1 (broodwar 1.16) AI check out SCHNAIL. It doesn't support Starcraft remastered on battlenet but does have very easy setup and custom keybinds.
I'll put the best bots to play against below. Team games don't work for 95% of the bots. Please play 1v1 and on the maps on SCHNAIL.
Terran: Krais0, adias, Hao Pan
Protoss: PurpleWave, McRave, Stardust, Locutus
Zerg: Monster, Steamhammer, McRaveZ.
Once again use only the maps provided with SCHNAIL in a 1v1 for the best results. Also these bot's play a different meta than SC1 human play. More of a bot v bot meta because they can be more cost effective with their units via micro/macro. Also another thing is they don't cheat, they can only "access" enemy player info from units they can see (Aren't in the fog of war) once they are visible they can be accessed. Some bots will "Lag out" for the first 0-10 seconds, just give them some time. Lastly most have "learning" and will decide the best builds based on previous performance against you. So keep playing them for different results. Bot APM can be ridiculous but have can have bad decision making. Happy hunting.
its impressive what the people did with this bot but i helped me just to realize how amazing the archivement from alphabet was to destroy the best players with a limited apm bot
10:27 & 10:33 are showing (new) masterful use of changelings to box in a disruptor, I gotta say this revealed more new tactics than anything i remember AlphaStar using. What a cool bot this is - is this Scarlett's bot or someone else that she helped make the rules for? Either way really really cool seeing all these new strats (burrow used as in-fight technique to dodge shots, changeling masked inside army entire time moving with army, etc)
Of course you did not show the ending, i whould also be embarrassed if i loose from that lead too
14:45 4 changelings are trying to obstruct of disruptor's movement
As a programmer (not a progamer lol), I found this super interesting. I wish AlphaStar and bots that use deep learning were more a part of the bot community though
Problem for us "hobby botters" is lack of computing power. Limited computing power and compexity of sc2 makes it extremely hard to make good ML bot.
accelerate is the right word for speeding up and slowing down. you had it right. Fun video!
I think I know how the computer gets more minerals. I noticed immediately how the number of workers on any hatchery is constantly fluctuating. When the number drops it must be the computer reassigning the drone to the nearest mineral patch that will be available exactly when the drone will arrive back at the mineral line. So it is not technically mining for a micro-second while it is reassigned. Every time it does this it saves that drone from walking from one patch to another. It could even choose between two patches that are next to each other based on which one will require the drone to wait an extra 1/100 of a second for the last drone to finish mining.
I didn’t know this existed, but that’s such like a cool thesis project for somebody studying ai or something. Normally I just watch for the hilarious things you say, but this is really ill
this is not the crossover I was expecting, but very cool to see the ai sc2 scene getting a bit of love!
This was so interesting! Please more 😁
Always love your vids!
What a sick view! I would love to see you cast a match between two top AIs in addition (with maybe some highlights to watch out for from the creators)
@harstem
at 8:40 ''this makes no sense''
its to do with the acceleration/deceleration values that the probe scv and drone all share, while the acceleration is 1000 like alot of non massive/non air units, giving them virtually instant full speed movement, the deceleration is the time/distance the workers/other units take to go from full speed to a complete stop, the harvest ability (what you refer to as return cargo) has a flag in the return cargo half of the harvest ability, which is actually 2 abilities, known as worker attack/harvest and return/return cargo, known as ''deceleration'' which allows an ability or attack to be used while in a ''deceleration'' this is most commonly seen with air units that appear to ''slide'' when given attack commands such as banshees and tempests shortly after a move command.
the ai is exploiting deceleration by using the move commands, to prevent the drones from ever truly stopping while avoiding the issue of the return cargo ability from being used while moving, since it cannot be used while moving, but since deceleration takes a certain amount of distance to finish happening, and since return cargo has a specific range, this allows faster harvesting, which is pretty cool, but would be fixed by simply removing or altering workers deceleration speed, or changing return cargos deceleration flag from negative to positive.
would it be possible to play archon with one of these bots? Imagine designing a macro bot to execute a sharp timing and then just controlling the units...
or a micro bot that always kites units/scouts/stops pushes while you practice a build order...
Nope. Blizzard would have to open an archon bot server, and not happening. Plus, the bots can micro at 10,000 APM. They just don't have strategic vision. Maru is a tiny bit like a bot that way, in that a Zest or SoS can outstrat. And machine learning over lots of games (SC2 is harder computationally than chess or go) needed. From my chess bot days, the challenge is to correctly "chunk" the data to learn many thousands of patterns. There are ways.
Some bots (mine included, named MicroMachine and open source) have an archon mode implemented in. With the arcon mode activated in MicroMachine, it will play normally but will not issue commands to units that are selected. So that way a player can control whatever he likes, but as soon as he deselects some units, then MicroMachine will do whatever it likes with them and might not understand the goal that the human had in mind. Still fun though, but can only be played against other bots unfortunately.
Interesting how the Changelings really mess up your micro. Essentially, the bot is using the Changelings to either stop you from getting closer or to help trap you from moving away
Y know harstem, they should make eris for terran and toss too...
I can imagine the pain of maru gameplay with eris micro...
Yet this is amazing.
The highest ranked bots are Eris (Zerg), NegativeZero (Protoss) and my own bot MicroMachine (Terran). They all have their own tricks and quirks and even the meta is different from the human ladder!
AI is so fascinating yet terrifying at the same time.
Interesting content! You're so good at keeping things fresh!
As you said "the bot didn't try defend that base" my phone registered it and tried to correct it as "book Uber to home" or whatever and I was like "come on. I wasn't even talking. Just let me watch, please :( "
When do we get Bot IODIS
"BEEP BOOP! Dear Harstem. I am bot! You will be assimilated! Disrupters are overpowered!! BEEP!"
"Multiple plugs in multiple sockets, hands tied behind my back"...
So when Blizzard makes the AI, do they limit the number of actions per minute to something that's possible for a human? It's interesting to see how it executes what it knows how to do, but then kind of falls apart when someone does something creative to it. You could imagine it targeting each interceptor with the exact number of units necessary to kill it, like focus fire but en masse. Warp in or do a BC flyby and they probably fall apart. In any case, it's obviously easier to maintain vision of the map when you have four mice to control your units.
"at this time he had multiple plugs in multiple sockets and my hands were tied behind my back" .... poor choice of words my friend.. poor choice...
10:37 the body blocking by the changelings on the disruptor......
It's 1600 APM with all the actions being done for a reason. Humans click multiple times when moving units etc. so you can't compare the two directly.
That was very cool my Captain!
That APM hitting 30k as the bot quite literally micros every single individual unit in it's army multiple times a second, while also microing all 70 workers is just... jaw dropping. One of those points where you really know you're playing a bot, not a player.
I never understood this thought process. Of course it can do that? That's like saying seeing someone put on a shirt is jaw dropping. It's something that is the baseline expectation of our capability.
@@queensharkeisha4479 because it’s interesting to see that and this person is excited about it.
how many ways are their to take out something burrowed or cloaked without detecting them? NUke, disrupter, what else?
Is there a way we can play with this ai?
1:34 I was already super impressed by scarlett, but oh my god
That changeling block with the immortal was so good.
Welp, it was an interesting match!
Thanks for the video!
That changling bodyblocking was insane
I would watch some of those bot v bot games. That micro is insane.
Harstem even BMs the bot.
Absolute #savage 💀
Play a game and set a random timer, which you cannot look at, and when it goes off, you have to all in your opponent with everything you have
9:27 😂
Queen walks straight up in front of your army and plops out a creap tumour in front of all the zealots.
Is that how a bot bm's?
i have no idea wtf im watching but im entertained :D
as a fellow Dutchie. our accents are so hard xD they give us straight away
04:45 - you forgot the most important uses: Tunnelling claws and banebombs. :P
Harstem, get some sleep man... :P
The body blocking from the changlings is crazy!
I would love to see you cast some alpha star games from back in the day
Watching the movement commands. It's like 1 pro vs 20 average players controlling the same team
Would the BGWSS strategies work against these bots?
I just discovered the AI SC2 Bots few days earlier and I was really impressed of them to be honest. Maybe due to the fact that I'm a developer ^^ But I'm really surprised that you have suceed to defeat it so easily. Could you provide the amount of kill of each disrupters please ? Excited to watch an other game like that after developpers patch the disrupters management to be honnest :) GG anyway
Beating grandmasters with stupid stuff idea: Fastest Possible Twin Adept Rush
I've terrorized diamond NA players this rush about 250 times over the last year and am 4000-4150 MMR with it, and 3500-3900 when playing standard or other cheeses. Build is:
12 pylon on low ground
13 gateway
15 assimilator, but prioritize mineral mining
16 gateway
17 cybernetics core
18 pylon
20 2x Adepts and 2x chrono (should have 50 gas as 2nd gate, pylon, and cyber core finish)
24 nexus
26 3rd adept
Adepts arrive at opponents natural at 2m 34s. The build works in all matchups but is best pvt as it can delay a low ground CC for over a minute. My most common follow up is a 2 base chargelot all-in with a prism and 9 gateways that hit at 5m 17s.
how do the bots handle cannon rushes or other cheese?
Roaches and hydras are cool and all, but why this bot didn't get any infestors? I feel like that's a major advantage of bots to controll individual units and what better than to control spellcasters in that case?
is there a way to change the difficulty of the bot Harstem ? could be usefull for a gold/plat guy for training ! thanks
I was waiting for you to play against sc2 ai bots!
como hiciste para ver la camara del IA demencial?
The opposite of accelerate is accelerate. Accelerate simply means to change velocity over time. Some people use the word "decelerate" to mean slow down, but this one of those commonly used yet improper words, as even "deceleration" is still acceleration.
Acceleration means to change velocity over time when it is used strictly in the world of physics. And deceleration is (from what I understand) not a term in physics.
But in general language accelerate means to speed up and decelerate means to slow down.
I guess this means that the opposite of accelerate is decelerate unless it is used in the world of physics where it has no perfect opposite - in which case the opposite of accelerate is accelerate like you said.
OMG you did it! yes!!! Bots being cast by my favorite absolute madman is the best thing ever
Where's that link you mentioned to play against it?
Computer left alone - they can snow ball quickly; good job
Really enjoyed this video 👍
I'd like to see more bot content. pls🙏
Is it possible to play vs ai bot im custom games to train a specific match up. I quitted sc2 because I could not increase my skills by laddering only. But I would love to play vs alphastar in custom games to train a specific matchup
I was curious about why the rar file is in Dropbox instead of GitHub and why there are not a README and LICENSE files so I checked the rest of the files to give it a watch.
I just went to the folder of the Ketroc bot's libraries and found it is using log4j library (last update of this files on June) and then get some Vietnam flash backs. So I suggest you not to run specifically Ketrock bot. Here is an scary and detailed explanation about that exploit: th-cam.com/video/BhEZieTl0TI/w-d-xo.html
That log4j lib is freaking everywhere, nightmare
it seems that the bot struggles with making drones at a certain point. or when to push or balance its gamble. nice stuff tho
Bots have to have special thought for (human easy) details like - can't see enemy unit - must not exist. Else the bot will charge up ramps and into tanks. But several ways to fake human memory.
great content man
Harstem, is there a way if someone can introduce alphastar into that bot community?
Link in the description is missing
did someone else notice both the evo chambers in the middle of main's minerals path ? lol
That was intentional. However the problem was is it broke the interaction with the manual drone mining, making the drones not mine anything at all. One of the two need to be disabled or modified/improved (or at least when they overlap)
Leave it to a bot to wait to give the GG til right before the last shot kills the last building.
Jesus Christ, that worker/mineral micro
There's another bot that mines gas through a bunker, and another than teleports probes back to the nexus through a warp prism.
Corner it in a corner you say. So you do not corner something in a open space?
5:56 no you couldn't go after the drones since the bot just researched burrow
We are the borg. We dont attack until we discovered a threat.
Did you see that? His apm is so high because he continues to do the worker trick throughout the whole game. That’s why the income is more because he’s doing it with the workers. Or not he, it.
Whatever happened to AlphaStar? And why is that Machine Learning-based bot excluded from the AI ladder?
Eris being the top ranked bot is fairly recent. For a long time, MicroMachine was the top ranked bot by a long way.
Just wait until serral or dark busts out that level of changeling micro :D
The reason it is mining better than you with less workers is because the drones are being rallied infront of the mineral patches which removes the deceleration. His drones are just moving faster. That would be my best guess from what I observed. I used to watch stephano do the same thing in the early game to get a few extra minerals. Doing this is impractical as the game goes on but it seems to maintain it on all the bases through out the game. would be crazy if humans could do the same things. Starcraft would be so different XD
Need an update rematch vs upgraded version