The Highest Ranked StarCraft II BOTS Are INSANE
Вставка
- Опубліковано 21 лис 2024
- How to play vs the bots: www.versusai.n...
How to watch replays: aiarena.net/wi...
Where to download replays: aiarena.net/re...
Make your own bot::www.versusai.n...
Patreon: / harstem
Daily Stream highlights: / @harstemdaily
My second youtube channel: / @harstemcasts
Twitter: / harstemsc2
Twitch: / harstem
Instagram : / harstemsc
#harstem #SC2 #StarCraft2
20.000 APM, defeated by burrowed lings. The irony.
ZvP free for zerg at all bounds
Harstem: drones should be pulled at this point
Eris:Bro I have 29.000 of apm
The text tags are to help the developers see how the game is going without watching the entire game. Since these are not usually AI, there behaviors, although complex, does not self-adapt. It is possible to have multiple builds, and dump a build once another bot adopts a good counter to your bot build. True learning takes enormous computer resources, and then may just do blink stalkers or other 10K APM strategy. For a while the top bots were coded in C++, but Eris and Deimos are both python, the same as most basic bots. There are 699 bots uploaded to the site, but only 75 are in the main competition. Mine is currently ranked 39 of 75, even though it's been untouched since July 2019. (Top ranking 3 - it's just a zergling rush bot).
But the true learning is what makes it interesting. AlphaStar invented its own strategies, like playing mass Disruptor, nobody else ever played this. I found it so interesting to watch, too bad they are no longer working on that AI. The normal bots I wouldn't call AI, just a script, since they're not reacting to enemy strategies (like building Phoenix upon scouting Spire, they react only to enemy movement like staying outside of an adepts range with the workers).
How much is needed to set up everything so you can make a bot? I'd love to try, but maybe improve someone's bot, I hate it when I have to go trough a long setting up process only to look how something works.
@@nightmareTomek Its not too hard to get started. If you have some programming knowledge you could get started less than an hour. If you code in Python or other popular language alot of frameworks to get started with that handle alot of the heavy lifting. The description of this video has some links to resources. If you didn't want to start your own from scratch I do a Mob Programming Coding Night on Mondays for new commers who just want to jump in and help!
@@Vers-AI Well, I work with Java and JS, learned C++ once and recently learned GDScript which "looks like Python". I know something.
I would love to jump in and help, but I also have 12398234234 other hobbies and don't wanna neglect them any further, it's bad enough as it is. So I'd like to avoid starting from scratch... xD and I wonder whether I'm gonna need hours just to select a unit or make a default pathing.
You have that programming coding night online somewhere to watch? Dunno if I'm gonna be free mondays.
@@Vers-AI Well, I work with Java and JS, learned C++ once and recently learned GDScript which... looks like Python.
I would love to jump in and help, but I also have a bazillion other hobbies and don't wanna neglect them any further, it's bad enough as it is. So I'd like to avoid starting from scratch, just help someone and learn about ai... xD and I wonder whether I'm gonna need hours just to select a unit or make a default pathing. I mean I'm already losing to the algorythm on this site here.
You have that programming coding night online somewhere to watch? Dunno if I'm gonna be free mondays.
@@Vers-AI I work with Java and JS, learned C once and recently learned GDScript which... looks like Python.
Well, I have a bazillion other hobbies that I am already neglecting as it is, so I'd love to help but not start from scratch. I don't know the scope either, I'd love to make the bots smarter and have ideas, but if I have to code 1h to select a unit, I might get discouraged.
You have that programming coding night online somewhere to watch? Dunno if I'm gonna be free mondays.
My guess is the tags are there to timestamp for the bots on small executable strategies.
For example you can see the Bane Drop was executed regardless of the state of the game, which makes sense because as a designer you'd need to code the baneling drop separately to anything discreetly necessary to win (injects, making drones, expanding, etc).
So if the bane drop is like an executable for Eris, having it say when it starts the executable will make it much easier to debug
When a developer changes a strategy, or introduces a new one, the tags help looking at the part of the game where the play changes. The new code may be buggy, or it may never run in a particular game. Once a bot gets to a certain level of complexity, it's impossible to know how it will actually play, and it's very easy to add code that looks good, but drops the rating 100 points or more. Or add code that never runs, or only runs when it doesn't really matter, etc.
3:11 Holy shit Im laughing my ass off at the drones staying always a perfect circle away from the adept it looks so fucking silly, love it
Like a school of fish
16:41 lmao those stalkers on the right upper side avoiding the lurker spikes like it's nothing
You can't do that? Pathetic! :D
Lurkers are crap for smarter bots. Buut that's only about 12 of the 75 bots running. Game theory suggests that a small number of strategies should mop up, as long as you're willing to get creamed in the games you lose. A bit like Protoss v Protoss for several years was rock, paper, scissors, except Stats and finally when Hero broke the mold.
@@Galahad54 I would just rush borrow, block the expansion and hold on to dear live until the AI runs out. Would probably still lose.
all you need, is 26k APM! ez
2024 and a bot does not get a fourth base because of a burrowed zergling. Furstrating
true
Would like to see the rematch where they update the bot to solve for that
The issue is probably much more complicated to solve then it seems
Especially if people think "just updating" is the solution like that Idea is some kind of insight 😂
@@TheGahta It is not so hard.
The bots uses predefined strategies and build orders. You could put in another rather simple strategy to always have a single detector roam around your base or potential bases to scout for burrowed and invisible units regardless of what he plays against.
That is usefull against this strat but also against burrowed roaches, DT or Ghosts.
There is even no need to check for a specific "OH i could not build here maybe there is something hidden"
@@1337Jogi "not so hard" then why isn't it solved?
Unless this is a new thing and they just didn't get around to it, which sounds even less plausible 🤣
@5:43 I like how the mutas swing and swerve around while assaulting a base, looks funnily enough like a Starcraft cinematic animation.
When you have all the micro in the world you can do sht like this. Its fair cause its bot v bot. Pushing 5000/7000 apm at the start stage of the game lol. I bet most of it is effective apm, not just spamming for the sake of pumping tne number like humans do.
I played some games against Eris and it's insanely hard to beat. The speed mining just means it has so much more stuff than a human ever can and it uses that to also harass a lot with burrow at expansions and tunneling claws roaches among other things. My only wins were from camping a bit on 6 bases and starving it out with late game compositions.
Your're a g gg ga gi girl..... :O
how you play "em?
Hi. What do I need to do to play Eris? (I watched the video to the end and saw the link in the description :)). And what MMR should be to have chance to beat Eris? 🙂
@@markjedigorian7489 the bots play at around diamond 1 level so you're going to be fine if you're a diamond 1 or masters player
@@Hagar239 I don't know if I can link it but search "take on the best sc2 ai"
From my experiences I'd say that you should be at least M3-M1 to be able to win a game or be somewhat competitive against various bots, maybe GM to beat Eris. It looks like Rotterdam played against it and said it felt like playing against reynor 😅
Love how this makes the matches feel more "realistic" in the sense that with an APM this high it's basically as if each unit is actually acting as an independent actor in a coordinated army instead of a mindless blob just following direct orders or doing nothing otherwise.
I think my first harstem video was a bot video about 2 years ago, and almost a daily viewer since. hope this one is as popular and brings in new folks for you.
Wow, even galaxy-renowned Donny Vermillion watches Harstem!
@@gehirndoperUNN won’t let me watch myself at home anymore so I settled for this guy.
Dude, roach burrow micro is so legit, I don’t understand how that isn’t utilized at high level. It’s not always applicable in every matchup or scenario, but there has to be opportunities we are missing. Burrow in general is underutilized, whether it’s banelings, burrow micro to save strong units, spotter lings, denying bases, burrow harass…there’s so much room for burrow to be a bigger part of the game in the future.
Burrow was used a lot more in the early days of SC2, especially with roaches and banelings. I miss those days.
Serral does it. Especially with early roaches.
Rogue in recent games have been burrowing lings for great vision in middle game
@@ThaSPAWN he used to, it’s been a while since I’ve seen him use it tho..
in big fights it dosent really matter because your opponent will almost always have vision, or get it the second burrow becomes a problem.
Tags are literally tags for the bot replays, we have a system that takes these chat messages and adds the tags to the replays for download so we can search a set of replays for specific tags and such.
Haha i love how the protoss boy has an optimization to utilize shields even if they arent fighting (the phoenixes) and the zerg has an optimization to take out individual units which is more important than using all its units to fight. Really fun video.
It is really interesting to see a bot game. It would be interesting to see how they improve in the future if other games become available.
www.youtube.com/@LaughNgamez for consistent bot game casts, he drops one almost every day from the bot ladder
is the deepmind still a thing? i remember them implementing as "alphastar" or like that years ago to get the sense of strats etc. btw im so hyped for the rank roulette coming up, love the series ^^
As far as I know they achieved their goal and therefore the project is completed, and sadly alphastar isn't available :(. Sadly these teams don't have google level resources, alphastar was insane and ended up beating serral and ended up being stronger than anyone in starcraft with a few *.
Deepmind is still a thing, but they stopped developing alphastar about 5 years ago. The project they did after "alphafold" won a nobel prize this year.
@@UberKrangThe ai was definitely not the best. It did beat serral 2-0 but serral played poorly. Harstem even beat the bot when it was at its peak. And so did many other pros. It was easy to beat by players in mid gm to top gm
I remember it was around 4.8-5.2k or smth like that before they took it down.
Alphastar was an AI trained bot, this series is a bunch of independent programmers making bots for fun. Some use AI, some use a set of pre-programmed rules, lots of very fun bots. My favourite is Ketroc's (anyone who watched LagTV will remember Ketroc was a player infamous for doing mass ravens or planetary fortress rushes). Ketroc's bot will do weird stuff like mass ghosts or ravens, try fill the map with planetaries, and if it gets well-established will begin mule-mining the enemy's mineral patches to deny the enemy minerals. The other one I love is Sharkbot which is also a chatbot :D
@@noobzerg1990 If I remember correctly (and that's a big if) Serral was playing the initial version of alphastar, while on the ladder it was a scaled-down version - the initial version had no APM cap, while the one on the ladder did, since they wanted to measure intelligence instead of mechanics.
So happy to see bots again! I love these casts.
there are YT channels who just post AI games. They're pretty mad.
Bots will replace human pro players, and later also casters 🤷♂️
@@leonnunhofer3453 What makes you say that?
If both bots have unlimited api and max out that function, it effectively turns into a pure strategy game as mechanical skill is no longer relevant.
If both bots have a perfect strategy, it becomes a game of pure mechanical skill.
@@dinastyx7414 I don't think they can have perfect strategy though, there's a constant chance of scouting being late or things not being discovered in time.
If the bots had maphack then I agree, it would be simply about mechanical skill.
Idk about pure strategy, because you can still have one good tactical advantage and win the game off of that. A single adept harass routine that exploits a vulnerability in their queen defense routine could get like 40 workers, ya know.
Not a strategy game, but a money game.
The bot running on the better hardware can just spam enough APM to lag the other bot out. Of course, once their hardware are both better than blizzard's server, then it gets tricky.
@@toast_recon yes tactics/strategy, my point is that in your example they exploited a strategic vulnerability to get those kills not that the adept control was faster than the queen could keep up with because of physical limitations.
Legitimately, wouldn't a competitive hybrid human/bot competition be fun? Like a player trains with an bot in archon mode or a mod of it, it gets tailored to that players style, and then the player and their bot play together in a tourney.
I think the bots would need way too many matches to ever be reasonable partners to a players style.
I agree.@@luckygozer
I suspect that this would quickly become "the bot speed mines, the pro player does everything else themselves".
I think it'd be neat to have professionals with access to the micro and multitasking of the bots. I think it'd be most interesting if each human player is able to customize/program the bot that works with them ahead of time, so you can get it to do what you want. (Unfortunately I imagine you'd have to do a lot of work to make the bots accept commands - not sure how easily this could even be done with existing tools.)
In a similar vein I've always thought players with access to scripts would be a fun game to watch. Automating creep spread and larva injects would be a dream lol
Shout out to LaughNGamez (www.youtube.com/@LaughNgamez) who does casts on alot of bot games over the years, starting with the AlphaStar ladder games from a while back.
Some context on this game: Deimos is actually a very new bot, Eris is much older and has more fleshed out scripting, thats why you see Deimos just sit on Stalker/Immortal all game, that's the only thing it has right now. AS mentioned by Harstem, rushing burrow to block all expansions is indeed something almost every zerg bot does and its difficult for even the higher tier bots to respond to, also Roaches with burrow/tunneling are an absolute terror on the bot ladder, and requires specific responses that many bots dont have.
moar botz. I watch him alot too.
Is there not some way to detect burrowed units? Couldn't you just have a bot send burrow-detection units to sweep the expansions periodically? If they had some "cleanup crew" units with them, they could clear out the expansions easily. It shouldn't take that many units to dislodge a few lings.
Sure, the cleanup crews would tie up some supply. But the more expansions they had, the less ground that the bots would have to keep clear. So the tactic would become less and less expensive over the course of a match.
@@tbotalpha8133 Some bots do have a clean up crew for various things like creep tumor clean up, zerglings blocking expansions, etc. but its not easy to do that and do it effectively while doing everything else that the bot has to do, detection units tend to be expensive and important, so they can be pulled in multiple directions sometimes and one thing is going to win, or sometimes clean up can be ineffective (I've seen bots have ravens circling expansion locations for an entire game trying to clean up creep tumors whose creep is blocking the expo but the raven simply didn't have scripting to effectively get the outlying tumors). At the end of the day the bot scripter has to directly code all this behavior into their bot and make it mesh with all the other stuff going on, its crazily difficult.
Cool to have an update on bots! Thanks!
Feels like a different game from 1v1.
And I think tag messages is when bot notifies about his plans what would be done next.
Deimos is a much newer bot, and not generally as well rounded as Eris.
The hard part about making bots like these is, they don't think and adapt to things they haven't encountered before like a human can. They either have programming to respond to certain things, or they don't. And if you want a well rounded bot, you have to specifically program for a lot of different edge cases.
In Deimos's case here, because it hasn't been specifically programmed to clear out burrowed zerglings so that it can expand, it just doesn't do it.
Of course, the main thing these bots can do that human players can't do is individually control every single unit on the map at the same time, for the whole game, without getting tired.
This is just beautiful. Like imagine what could happen if you could play at the speed you could think it?
10:02 here you can see that Deimos is actually throttling his commands for some reason. Since there's seem to be no limit to rate of commands that bots can issue i dont really see why this is happening. It is obvious that phoenixes eventually go attack floating queen but with some surprising delay and one by one. Very strange.
Maybe because the queen attack goes out, hits the Phoenix and the bot cancles the lift because of the incoming dmg.
There is a reason to throttle: the match server is a bit of a black box for bot developers (we cannot query the server metrics), but with any server there is an upper limit to throughput so it becomes necessary to throttle. This is especially important in real time matches as they are exactly that: real time. To accelerate development bot developers can run bots in a fixed step mode where the game loop will wait for all issued commands to be processed by the server before advancing. Writing bots which perform correctly in both fixed step and real time matches is a well documented challenge. With respect to real time matches, if the bot does not throttle commands the commands issued by the bot will fall behind what's actually happening in real time as the message queue for the bot is saturated and responses begin to lag.
It seems to me like they are waiting for a minimum amount of shields to attack a lifted unit.
We can see that there is no danger but the bot doesn't realize and is abiding by some past experience or training and trying to maximize the Units Hp over anything else. Same reason why the phoenixes kept dropping and picking back up the same queen later on. The damage of the attack from the queen is delayed and because of the bot using the phoenixes on a knife edge the delayed damage from the queens attack triggers the lack of shield alarm, so the phoenix disengages the lift and runs away and another phoenix takes over (even though there is no threat).
I think there is a 100k apm limit where the game just crashes.
@@triple_dread_bonus_wave I guess a server side standardized maximum throughput configuration make sense here? If you had a upper limit you would at least know not to pass it and throttle yourself.
The HUGE differences between this and the machine learning AI:
Bot AI isn't usually "trained" - it's programmed to interface with the game, keep track of all the game variables, _decide_ the next move to make using heuristics (numbers) and a LOT of "if" statements (if drone almost in range of an adept, move away) and executes them. This is how "tool assisted speedruns" and cheats like "aimbot" are made in other games. Now these guys have an another step because you have to battle another bot so you keep running matches and then a _person manually_ looks at it, figures out why their bot lost and addresses it like fixing a bug, to keep improving strategy. The two top bots are practically bound to always go toe-to-toe because the losing bot will keep getting updates by a programmer to fix an issue and hopefully win - a self-balancing equation involving humans (a few friends battling bots for fun, in this case)
There's a really cool game all about making bots called "Gladiabots" - you don't even need to know programming to play it.
Neural Network AI refers only to the "decide the next move part" but the programmers don't tell it how to make any decisions - the programmers basically give the AI "eyes and muscles" and connect that to the "brain" - but only decides the size of the brain, not what it thinks). It's inspired by how neurons in an animal brain work.
Machine Learning means it is set up to constantly evolve and "learn" by constantly making slight changes to the neurons via complex algorithms - these algorithms _do_ get updated by programmers but only to fix the _learning_ process, not the gameplay decisions made by the AI - after changes, it's usually _completely_ reset and is a dumb baby again, going to the "learn by failing at it a million times" school. In this case, the "grades" are based on performance against another player, so there's various algorithms for battling bots against themselves millions of times OR, as Alphastar started out, its "grades" were decided on how well it predicted what a human player's next move would be in a replay (but only at first, to have a starting point - battling against itself is better)
Alphastar was specifically designed to be as generic as possible and notice how I never mentioned "Starcraft" or anything from it in this part and even the "battling" part can be changed - the whole point is to do something a human can do, to as great a capacity as a human. That's why they limited the APM to be closer to human limits (they overshot by a mile but... eh, it was interesting!) It was never meant to play Starcraft specifically - maybe the development team just liked the game and were very familiar with how competitive it is - a perfect environment to develop something really really complex. By the end, I think hundreds of programmers contributed to Alphastar.
It was never meant to compete with bots and it wasn't meant to compete with humans either, actually - the matches with professional players were mostly a show. They didn't care if they actually beat the pro players fair and square, they cared that they made a Neural Network that could actually fight humans - which was never done before - and probably also so that investors and internal higher-ups would notice and gib money. That's why they stopped there and went on to make Alphago and later Alphafold - that was their real goal.
I watched the Bot games pretty often and definitely enjoyed the cast would love to see more of this
The big question is, did the bot exploid the other bot, knowing that it would be overwhelmed with the burying?
The way they box in and flow zerglings with a stalker was just amazing
Fun fact; Eris a dwarf planet in the solar system named after the Greco-Roman goddess of strife and discord.
Deimos is one of the moons of Mars, and it is named after the Ancient Greek god and personification of dread and terror.
The other moon of Mars, Phobos, is named after the Greek god of fear and panic.
Deimos is a new bot with a very strong mid-game push that it sees a lot of success with but it doesn't have the depth of some of the more mature bots. It doesn't win many games that reach the end-game and it's susceptible to certain cheeses. Still given how new it is it shows tremendous promise.
For the protoss its the local optimum problem. It is not going to stumble into the solution and reinforce that behavior. Not being able to warp because the ground is blocked should also give the a.i. info to work with. It cannot formulate a constructive response.
Hardcode it to scout expansions with an observerbefore the probe goes there and a units to squash that ling.
2 observers covering the same ground does not help either.
There is no "reinforcement", this is a hard coded AI, not one that trains itself. The 2nd type is much more advanced and it would require an insane number of games as training data.
Just goes to show these bots dont really understand what theyre doing. Its just learned scripts that are constantly running. A total noob could in 5 min of reading about the units realize that oh yeah the only reason I cant build my base here must be because of borrowed unit -> send observer there with a couple of fighting units -> send them to other expansions too
@@Leonhart_93 ok then my criticism makes little sense. A solution can just be coded.
I think a hybrid model of a.i. could work. But the machine learning requires to many games as the number of iterations to learn increases exponentially with the number of variables.
@@Leonhart_93 there was AlphaStar who learnt from 65,000 baseline games and then thousands of ladder games in 2019. But that's an exception and was a hugely funded specialist project.
Looks like there just was an oversight on building space being blocked by invisible units. Cause you wont get a red restriction building marker, it just will fail at placing building. So there just wasn't a decision to handle burrowed unit put in place in code of the bot. Just needs a check if building had space but failed to be placed. Then send detection and some units.
7:22 There it is, the first time I've heard someone call an AI a person nonchalantly.
High APM, balancing games for bots would be interesting
I remember watching Alpha Star erase some top rated grandmasters several years ago. Interesting to see they created new ai/bots to play against
Would love to see you cast more bot games!
The ultimate combo seems a human making executive decisions and AI making heavy lifting in micro and tuning.
Crazy idea after hearing how the AI micros the drones: SC2, but each unit is controlled by a different person from the player who controls all the buildings. It would be the most insane viewer battle between streamers ever
Native A.I on easy difficulty is programmed to detect & kill a burrowed unit, this 12,000 apm megamind hasn't grasped the concept
To be fair the built in AI has direct access to the game's state. Something akin to perfect maphacks. And it's programmed to pretend to not be cheating. This as I understand has access to largely the same information a player would.
@@stonium69 Seems not so hard to give a check to the worker if building in a space is possible and if not if he sees anything blocking it + adding a flag to scout that position.
If that is too hard the AI could just always patrol a detector around over your bases and potential bases to see burrowed or invisible units.
With his APM it is totally worth it to constantly let a single detector roam around.
@@1337JogiI mean it's probably an oversight. Eris is at least four years old and I remember similar bots from many years ago having this capability. Deimos is a very new bot (like 3 months old) likely made by one person who built it with a bunch of micro optimizations/build strategies made specifically to counter the other bots and bring it to the top of the ladder but which doesn't yet have coding for every niche situation.
@@1337Jogi well the main problem is how these A.I learn things really... it's kinda not a trivial task to teach them to move detectors towards the base they struggle to take
@@MadProgger These AIs don't learn anything, they're hardcoded.
actually the only bot cast to properly describe the mechanic of speed mining and why it's faster 👍
That was interesting to watch!
Thanks for the video!
whoa im less than a few mins in and this is crazy with the mineral micro. thanks for the explaining it.
its so cool to see the micro. especially from the mutas and pheonix. Wow.
Bummer about the burrowed lings being the deciding factor. Hope the toss AI learns a solution. Crazy to see all the chaos with everything being microed individually. Thanks for sharing
These ai are fascinating. I don't know much about the script files, The ai i've attempted to wright have never done any thing different, so I can never exactly tell either. I may need to ask these guys a few questions, Good pointers they might have for me perhaps. Cool video harstem well done.
7:40 burrowing lings at expansions is very much the meta right now, and has been for a while. Burrow move roaches are practically immortal in the early game, so the bots have burrow anyway, which means the cost of delaying or denying bases this way is very low.
I love Ai vs Ai more then anything and not enough people post about them!
i think more people should watch bot vs bot games.
they are very interesting to watch and are very different from just normal pro matches =)
more bots!... oh sorry wrong channel :) anyway I love to see your commentary on those bots viewing from the perspective of the professional player comparing their moves to what actually humans would do - or do not-
Maybe the protoss bot is programmed to get upgrades after getting a 4th base? That would be a huge flaw
Thats crazy, I really believed my 3v3 teammates were the winners of the latest major bot competition.
Zerglings, in terms of pure stats, can beat about 3x the resoruce cost of stalkers if all units on both sides are constantly attacking.
With individual unit control, you get something close to that. That's why they seems so oppresively OP.
The key being "if both sides are constantly attacking. There's lots of room for improvement here, especially with limitless APM. You'll never see a pro match where this happens unless the Protoss is distracted. They'll fight in a corner or a choke point to limit surface area, and blink back isolated/ damaged stalkers, which the bot here hasn't figured out yet.
This is just awesome xD
over 500 actions per second and super percise
im just baffled
Are there any Games with Ai and APM limiter? To see which Race is the best up to 100, 200, 300, 400, 500 APM.
Would like to see more of this.
Will you cast more or can Tell anyone where to see the matches ?
first time I see these AI games like 3 years ago from the Alex007 channel. well done that you decide to have such playlist also
11:35 - Not surprised AI dont use spore crawlers,they pby try to micro the after each shot
Ai playing Starcraft is like a real war. Every unit moves like they have their own mind.
.....i have a fantastic idea, how many pros (in archon mode) does it take to out-do one of these bots? Do it Harstem, this'll be such a Man vs Machine
those few burrowed lings pretty much won the game for zerg, because the fights were going pretty well for protos with his colossus being at zergs mainbase. Protos just had no money.
I would love to see AlphaStar vs some of those bots.
I'm shocked we didn't see warp prisms? Surely the micro would be ridiculous
I am an AI developer, so at first i have say how impresive it is. Makeing AI bot for SC2 is super hard, and even google found it challenging.
Second of all, i do realy believe, that Eris is not a just 1 AI model, but severals. There is some kind of "overmind" model that is in charge of the game for Eris. And there are other models, that are specificaly trained to do certain strategies, like bane drop, or ling harrassment. Those msg on chat, are information for devs, that "overmind" made a decision to activate a "bane drop" model.
Then during improving the bot, then they sort of "debug" their overmind bot, knowing when they made certain decision. They can mark for example that this ling harassment was a mistake, and instead the bot should use a different strategy there.
I could also be completely wrong about this, because there are multiple approaches you can take to create such AI. To my sense it seems pretty natural approach, and also pretty cool tbh.
i don't have enough information to know completely but I am pretty sure that Eris just like other bots on AIArena is not AI and is just a pre-programmed bot.
Are you sure you're an AI developer? Comparing google's neural network machine learning with starcraft 2 if-statement bots is like a mathematician comparing variables to signs.
Eris does have "managers" and modules which contribute different aspects to the bot. I can't speak to its specific design but it can get pretty complex and I have personally been admirer of how Rasper (the author of the bot) lays out his bot's code. You're correct that the chat message is for the dev but its also for the replay system to help filter out later
What about Alpha Zero? I thought he was the best robot.
I think it would be far more interesting if they found a way to limit the APM to human speed (maybe creating maximum actions per second as well as maximum actions per minute). With Neural Nets, the result should be some quite spicy tactics that humans could also implement.,
Enjoyed the bot cast
basically the IA didn't took fourth base cause underground zergling, the rest of the video after was useless until the IA get fixed
This looks so beautiful, humans can't micro like this... it almost looks like actual war where every unit has its own mind. I like to imagine a mode where there are like 200 human players each microing their owm unit, it would look something like this :)
It's interesting how lopsided the skill-profile of these bots is: sometimes they are utterly ridiculously superhuman, but then, there's staggering inefficiencies: around midgame there's a point where the game is roughly even and where even I might well be able take over either one of the players and possibly win, since e.g. mere 5srrrrrrrrrrrrrrrrrrrrr with floating resources would give a major boost in fighting power, and some of the micro seemed counterproductive.
Imagine if AlphaStar-like AI (which has far fewer exploitable weaknesses), or an expert human player, could get additional button in their controls: "assign AI control", and e.g. start the game by box-selecting workers and pressing that button, or pressing that button when they are facing mutas while having phoenixes, or pressing it when enemy banelings are rolling at their marine clump.
Yes!! Now you have to play them!!
If protoss bot could have warp prison tactics,that will be huge
Would love to see these bots play Alphastar with uncapped apm
I feel like rather than the lack of APM, a lack of decision making is apparent here. But these AI will only get better and better with time.
@@ddenozor _These_ AI won't, they don't improve over time. The programmer might get better or code in solutions to cases they previously missed, but I don't think that's what you meant.
their decision-making is not the greatest since 90% of the things they react to need to be programmed by a person. so unless we get a bot made by a pro gamer, I don't see how the bots will beat an actual AI or a high ranked player.
I would love to see the bot matches on the Evo mod.
It is sadly impossible to do =(
Im not sure how the bots are programmed but they should definitely be attacking the mutas until they have had their shields damaged. Microing this so ultra safely means the phoenixes can't deter them from still doing damage
the burrowed lings is kinda cheezy lol
I'd like to see some bot developers try this for WarCraft III . . . .
need an API :P
Would love to see Serral or Clem go against one of these super bots.
these bots have no chance against Clem or Serral since they have perfect micro but bronze player decision making.
these are not intelligent AI these are pre-programmed bots.
They have work to do on deimos to clear burrowed lings, and get upgrades.
We need 4v4 AI top tier right away. This is nuts.
This game was won by five units that never attacked.
I thought that you were talking about yourself when you said that the bots are insane.. My bAD!
I’m wondering how these bits would do against some of the higher level players. Like, the micro is obviously really scary, but it’s clear that the strategy behind it just doesn’t measure up against human sensibilities. Have there been any high-level human/bot exhibitions?
it's weird in a lot of senses the bots seem very hard to beat with the apm and micro intensity at the start of the game, but the longer things go on the weirder the bots start to behave.
@ShiDWa they aren't. The builds are optimal, they don't know how to properly engage and never make the correct tech. They might be plat level.
It seems Deimos lost pretty much because it can't deal with expansion blocking which seems like a fairly big bug.
*micros 4 mutas all individually with 13k apm* and our Captain drops a "nice control" with the straightest face I've ever seen
that APM is nuts
at one point 40k apm, /60 = 666.66 actions per second.
Remove the auto slow for mining from drones so everyone gets minerals at the higher speed ?
burrowed lings were the mvps of this game
If the protoss bot figured out the burrowed lings, then I think protoss would have won
I know not StarCraft, I was thinking to make bots that learn for WarCraft, would not be difficult to make them about 1500 MMR.
Oh, you HAVE to play against them!!! And cheese them!
How is the zerg scouting places on the map that protoss couldn't have possibly gotten to yet?
do the bots change thier strategy in between games? or is it the same one?
The game is (relatively) balanced around the current pool of human pros. I'd guess it's unlikely there would be competition across the races if playing with 2k APM bots. I wonder which race dominates in the case of perfectly developed bots. I wonder which race dominates on the current bot scene
Is Alphastar in the ladder as well?
So the lings just stop the other bot from playing? How do you not program a simple fix for this once the bot notices it cannot build the 4th base
As @Knowbody42 pointed out, Deimos is a new bot programmed to specifically make a very strong mid-game push. It can't do anything else - it might not even be able to deviate from the build order one bit. Eris is a more developed bot with answers for every situation - and, sure enough, it answered Deimos's push.
idk thats this exist but make sense i love to watch stockfish against leela and now i have a new content i can enjoy thx
So are the bots limited like the google bot a couple of years ago or they are of the leash ? Does they cheat? Like map hack ?
Its odd that with these uncapped APM's terran bots arent dominating. I heard that with infinite apm marine rush would beat everything since marines are extremely strong if you could individually control each of them in a 30 marine blob
i know that some terran bots are getting to division 1 on AIArena no problem with 5 rax 1 base marine push lol
Maybe that's why there's no terran in this video
@@benismannactually no. That Zerg AI is actually the rank 1 on AI ladder. For some reason, the Terran AIs are not dominanting. Could be maybe they are getting cheezed/exploited, like the Zerg AI burrowed zergling to prevent the Toss AI from getting bases.
If I remember with uncapped APM also Blink Stalkers were extremely powerfull as they could individually and perfectly attack and walk or blink away each stalker.
Since Stalkers have higher range and speed than Marines and higher range than Roaches they can technically fight either of them without taking damage unless Stim is activated or the Roach is on creep (or upgraded) even with that fact they can take minimal damage and blink away.
@@atifarshad7624 the Zerg AI is actually the rank 2 on AI ladder because Deimos recently took over the first place. (I know what you're talking about and Eris was #1 for a long time, and will probably take the #1 spot back next season)
but if we look at AIArena ladder leaderboards right now it's
1: Deimos (Protoss)
2: Eris (Zerg)
3: DominionDog (Terran)
4: Xena (Random)
5: Caninana (Zerg)
6: EvilZoe (Zerg)
7: MicroMachine (Terran)
8: Ketroc (Terran)
9: BenBotBC (Terran)
10: sharkbot (Protoss)
bots per race:
Protoss: 2
Terran: 4
Zerg: 3
There are pretty much the same amount of bots for every race. there's no "dominating race" really.
Is there a bot to cannon rush me ? for pracitce
i know that sharkbot loves to cheese.
No. (You could make one if you have a lot of spare time. Some bots do it but very bad)
moar botz(welcome to seening the sc2 bots. They do some cool stuff.)
There's a very good reason why deepmind limited alphastar's APM (to like 200 iirc) when they were trying to sell it as truly 'intelligent.' If the bot is allowed to win games microing with 40k apm it isn't actually getting smarter
Alphastar didn't really do enough for limiting micro imo. It didn't restrict itself to microing units only visible on a screen for example so you'd see it microing far apart units in ways humans can't. And from the games I saw it never really learned much strategy. It learned how to execute strong all-in pushes so it never needed to alter its composition much depending on what the other player did.
@@asterpw Interesting. I remember when it was revealed it played a bunch of games against TLO and another I don't remember (Naniwa?). And it was playing super safe/careful, always scouting obsessively, notably it clearly oversaturated its initial bases before expanding and shrugged off all of the harassment/trolling the humans threw at it as a result.
It would start moving units across the map and then immediately pulled them back and send out scouts when it saw something weird like the humans' army not being where it expected them to be (and lost opportunities to kill them for it). And eventually it did always win with a crushing macro advantage.
But I haven't played in many years and I understand many of its perceived advantages were disproven in the following years.