Will the AI bubble make single player RTS great again? For years, RTS games have promised to feature an "AI that learns from you", yet it always turns out to be bullshit. RTS AI has always been stupid, even among the latest titles. Will we finally get good AI now?
The current AI bubble will do literally nothing for in-game AI. People use the term "AI" for a lot of unrelated things.
You probably technically could have something like that, but you'd need to code the game to specifically use that kind of machine learning. But if there's one thing you can count on most game companies to be, it's risk-averse. So I doubt any execs would greenlight something so cutting-edge and uncertain. That would only interfere with their plans for the next DLC-riddled, battlepass-enabled mess for the zoomies.
Your only hope is some obsessed indie or a severely autistic modder who'll put in crazy amounts of dedication into making it work.
It'd take some smaller company that's desperate to make that risk, it'll explode and they will be the OG pioneer and original and then every major AAA homosexual game company will copy that grift as its the next big thing
but the ai might use all kind of meta (more tryhard than multiplayer) to beat the shit out of you
>they finally make smart AI
>it metagays you
it's over
this is what Im afraid of, AI so good, its like youre playing multiplayer
the AI bubble is about algorithms shitting out pretty pictures and snippets of working code for ranjeets in india jobbing at $0.25/hour, it won't do anything to create highly complex AI behavior trees that correspond to RTS gameplay. We will never get that as long as the industry is in stranglehold of globohomosexual investors and investment funds, focused on *surprise surprise* financial profit at lowest cost possible ONLY. Thank the dodge brothers for setting precedence in the US, it's not just Ford seething, it's the entire world populace.
There are games that have that but even if it had perfect human learning you'd still have to play hundreds of hours long matches against it to train it.
OpenAI made a DotA 2 AI back in the day by the way, but this has nothing to do with regular consumer strategy game AI that doesn't need to be that powerful to be very very good but also more importantl needs to not be very very good but instead fun to play against which is much harder.
I agree that the AI should not mimic a pro APM microgayging meta player. But it still needs a much more advanced level of decision making that we get.
Wouldn't it be possible to develop one that mimics an RPing player that holds back from cheesing? Give it awareness of the game's mechanics, block cheesing actions, and give it the ability to have an over arching strategy. It would be cool to play an RTS where the AI doesn't just spam units in your direction to "win" but tries to be a complex enemy that mimics the strategies of its historical setting.
>the missions have objectives and situations beyond a traditional symmetric battle which play into this behavior rather than having it as a limitation
This is a good point. Missions are fun because they're scripted - imagine an RTS that writes a background script for itself for every skirmish and modifies it according to player actions. The script would be not how to simply beat the player as fast as possible, but to build an army and move it around the map to surround and surprise the player, and build its base and outposts to slow the player down. Shit would be so cash
What advanced decision making? All of /vst/ is basically a giant optimization problem splashed with randomness.
More to the point, "waaaa make better AI" gays never actually specify what is it that they want AI to do. "Be like a human player" - in what way? Human players do moronic bad plays all the time, or act irrationally, or metagay, humans play for different reasons and in different ways. Which of these behaviours do you want to mimic and why? There's no objective or benchmark, I'm 100% sure that most of these whiners would fail a blind test to detect whether a player is AI or a human, and the AI that would hide the best wouldn't be the smartest one, but the one that obfuscates behind the smoke and mirrors.
It's like knowing that it's a computer algorithm somehow makes the game less fun for these people. Just play MP then, because no advancements in AI will make it feel better for you, considering you have 0 actual concrete complaints.
>"Be like a human player"
The post you quoted literally says the opposite.
>Wouldn't it be possible to develop one that mimics an RPing player that holds back from cheesing?
That's the point, you don't want an AI that is actually good or learn from its error.
You want an AI that pretend to be a strategist with ingame "doctrine" and only lose if you don't unravel the weakness of the strategy.
At most you can have the AI learn but only to avoid a ridiculously easy defeat when the weakness is exploited
>you don't want an AI that is actually good
"Good" AI in a single player game means engaging and fun, not mimicking a pro meta player (since that would be the opposite).
> or learn from its error.
If it learns you have a go-to move, then it shouldn't learn a hard counter, but to react in making it more challenging to pull off. That way you always have to find new approaches and be unpredictable instead of relying on the same move to win.
Why couldn't they let it play out millions of battle simulations before making the game around it?
>"Good" AI in a single player game means engaging and fun, not mimicking a pro meta player (since that would be the opposite).
That's the point.
The AI isn't created to be "good", it is created to look like the current enemy and offer you a challenge, if it's not built like a tutorial (you must use artillery to defeat this enemy, click here to research artillery).
In multiplayer a "good" AI is one that can pretend to be a type of players (rusher, turtles...etc) and change only if that's a possible dynamic ingame (hard to rework the entire strategy midplay).
>If it learns you have a go-to move, then it shouldn't learn a hard counter, but to react in making it more challenging to pull off. That way you always have to find new approaches and be unpredictable instead of relying on the same move to win.
In a single player no one have a "go to" move unless the game is lazily built like a MP game against AI.
Even in those game, the FACTION with its different dynamic is what is supposed to bring variety.
And good luck designing multiple faction so each are "equal" against each others, yet have different ways of using their units that change everything.
The first criteria require both side to be equal
The second crit require both side to have rock-paper-scissor gameplay
They cancel each others
AI in RTS games isn't bad because it's limited by the technology, it's bad because the programmers are incompetent or the game designers think the players are incompetent. Or they're catering to the lowest common denominator that wants an easy AI even on hard difficulty.
I hope they do something like that again in the future, afaik it was a 1v1 bot only, I wanna see if they can somehow train a full team of bots.
They later made a 5v5 AI that stomped the world champion team. And then let people on the Internet play against it, it was pretty good.
It was also a based shittalker dropping messages like
>probability of victory: 78.6%
in chat as the game starts and every 10 minutes
>It was also a based shittalker dropping messages like
That is unironically a winning strategy in dota if you can do it in a way that makes the enemy team fight each other and not you.
Oh frick yeah, I didn't keep up with it after the first showcase of the 1v1. I gotta look up some videos.
As far I'm concerned DOTA 2 is an example of eSport-optimized game with simple DPS and AoA type gameplay, no physics whatsoever, standard hitbox, speed is just a value you change as is HP and other.
Change the graphics and it's a game you could have played 20 years ago.
So no surprise at all an AI can play it. It's the most simple environment an AI could have.
My only question is how does the AI "see" the game.
And I don't mean the Fog of war that is apparently correctly followed
https://en.wikipedia.org/wiki/OpenAI_Five
>Thus, playing Dota 2 requires making inferences based on this incomplete data, as well as predicting what their opponent could be doing at the same time
I'm talking about wether the AI directly obtain all elements position or if it do image analysis.
From the wiki it look like it read directly from the API.
So once once again, that's the easiest job for an AI.
Finally given what entertain human I'm not ruling out that game like DOTA are built on models that are mathematically easy to solve for AI.
Still impressive, but to me the challenge is in image analysis and robot walking in the real world.
>game with simple DPS and AoA type gameplay
You never played dota. It isn't like league.
DOTA 2 does hate DPS and AoA.
All you did here is insist there's other mechanics that somehow change the result of what I said.
Sorry but as long as it can be treated with simple logic/math without complex image analysis it's still easier for AI.
>broken abilities
I hope your just meant "high impact" because broken mean unbalanced shit and this is even easier for AI if it can exploit a winning strategy much faster than human, or even too fast to be accessible by human.
That's why I'm not surprised by
saying some players won by playing safe and analyzing what this AI is still bad at.
If DOTA 2 is balanced then it's not broken and it's just a matter of someone/something finding the meta.
Like a mathematician long ago who kept ruining Military wargame and winning by abusing their model. Like spamming small boats to overcome economically balanced fleets. Funny how it look like recent warfare.
And what I mean by this is the game isn't always about fighting. The enemy team can be objectively stronger than you and will 100% win any fight for the rest of the game, but you can still win by pulling something weird out of left field by abusing some heroes ability. The worst losses happen this way. One team gets arrogant because they will win any fight easily, but when they initiated a fight there was a guy teleporting around on their side of the map killing all their creeps so even if they win they can't do anything.
Dota is the only game I know of where it doesn't come down to straight forward meta. It's everyone abusing broken abilities, and then responses to those broken abilities, and then responses to the responses to the broken abilities. Then some dude will pull a weird ass strategy out of left field.
OpenAI was incredibly good at winning fights even from behind, it was much better at teamfights than the world champions because of inhuman precision, teamwork and reaction time.
What it wasn't good at was long term strategy so teams of pretty good (but not pro tier) players could beat it by avoiding fights, playing the map and outfarming the AI
>use my AI to beat Dota fans
>become famous player
>sponsor shitty products
>profit?
i can bet OP is a homosexual who doesnt even play rts and loses to beginner difficulty when he tries.
Be careful with what you ask for.
You do not want a "good" AI. You want an AI which is flawed and tell you its move in a predictable way but just complex enough you felt it challenging to unravel it's tactic.
For reasons pointed out by other posters, there are bigger hurdles to good AI in gaming than just the implementation, which has been within reach for some time now. I think a better solution is for rts vs AI to be more asymmetrical, since the issues with bot performance arise from the need to mimic the player's abilities without simply playing perfectly or randomly. Look at games like AI war or campaign missions from warcraft/starcraft: the AI is simple and predictable but the missions have objectives and situations beyond a traditional symmetric battle which play into this behavior rather than having it as a limitation. Separating or otherwise accounting for the differences in PvE vs PvP in RTS is, I believe, key to innovation in the genre.
ruski devs made a neural network ML AI for blitzkrieg 3, it was a big part of the marketing for the game
nobody played blitzkrieg 3 though because it was shit
>6 years ago
Believe it or not AI evolved quite a bit since then.
Either a bait thread or a very moronic OP.
Neural networks only gets good at something after millions of try and failure sessions.
>RTS games have promised to feature an "AI that learns from you"
First they would have to code an AI that is not mechanically inept. The only game I know where the AI is mechanically good is Zero-k.
Only after that can we try to make it "smart".
I'm going out on a limb here but I think most people want to feel challenged while not being completely overwhelmed. I think the perfect kind of 'AI' would be one that adapts to how well the player is doing and either slows down or ramps up the challenge based on how they're doing.
If it's way too easy players learn bad habits
If it's way too hard players just give up
I know some other games already have dynamic difficulty but I'm not sure if any RTS do.
You won't see this until at the very least every GPU comes standard with AI hardware acceleration.
Otherwise it would run like shit because you'll need a pretty big neural network for this to actually work. A small neural network is effectively moronic and traditional AI would perform better.
"The best (that is, the most compelling) AI isn't one that can win decisively, but lose gracefully."
Not an exact quote. Wish I could remember the source, it was specifically about AI in RTS games.
Neural Nets will be overkill. Devs just need to put effort into AI and iterate on it. I don't think they ever update the AI in patches for any game or its rare.
I don't want to deal with some turbo autistic AI with instant reactions that turns the game into a singleplayer pro match, but I also don't want to deal with a moron that cheats like crazy.
>I don't want to deal with some turbo autistic AI with instant reactions that turns the game into a singleplayer pro match.
That is not even fun, unless heavy restrictions are put on the AI it could outmicro any human player without effort because a human player can only look at one computer screen and have one mouse and can click on just one thing at the time. AI has no such limitations unless it is deliberately programmed into it. It would easily be able to micro every unit in the game simultaneously and pay attention on every single pixel of the map simultaneously it isn't even a fair game. But once you put all the limitations into it then it will still be annoying to have instant reactions you would never land any skillshot or click because it can react instantly.
Just as annoying as fighting an AI opponent in a fighting game where it is very clear to notice that the AI can react way faster than you an enter the appropriate movement command to counter your move as soon as the first frame of the move animation plays. It is just not fun because every time you manage to hit it you know it deliberately allowed itself to be hit.
>It would easily be able to micro every unit in the game simultaneously
Age of Empires 1 original release had that. Just watch a post-iron age hard ally microing 20+ elephants all around their base.
i predict in the future we will all have customized personal AI's just like our smartphones today that we will instruct to take over primitive AI's from older 'retro' RTS games like Starcraft 1 or Warcraft 3 and just use them as smart opponents.
>and just use them as smart opponents.
AI: "should I pretend to be more stupid this time sir? During our last 76 previous match my fake 50 IQ mode was still too high to allow you to win."
*Sigh* "Yes Ashley, please allow me to win again and don't remind me about difficulty settings until our evening ERP session is over."
unorinically pretty much 95% of sessions will be like this. you will beg the AI to let you win like a soggy b***h.
which is why they should train the AI during dev and not vs me, bonus is you could retrain it after any balance change once you have the framework in place, this sounds like a startup idea
Did you seriousy miss the joke?
Joke is future AI will be too smart for you to win, but not user-friendly to protect your ego from the harsh truth.
also read
, even without going "aim-bot" they did make AI who would act as a team better than players.
So it's more about making sure the AI have weakness that aren't too obvious.
This is pretty much how AI already works but instead of saying "How smart should I be" its "How much should I cheat to even stand a chance"
>For years, RTS games have promised to feature an "AI that learns from you",
>Turns out marketing talk is choke full of lies and bullshit.
Gee Anon, what else is new?
People don't really want an ai that can learn in a strategy game. It'd get too good and you'd only fight losing battles.
What they want is an ai that can alter its strategy to keep things fresh while having the occasional win but ultimately fools the player into thinking it's competitive but really just provides a catered experience for the player.
If a better AI would revitalize RTS, how did the RTS-es of old get so great to begin with? The AI in those games would by your statement be shit, but they were great anyways.
Also presuming that the recent boom in transformer AIs would lead to a betterment in game AI is like saying that power-tools should be a shot in the arm for classical painting due to how quick they make work. Sure painting is a form of work, and a painter certainly uses tools, but a screwdriver is only useful in a very small subset of artwork.
>Also presuming that the recent boom in transformer AIs would lead to a betterment in game
I ignorantly assume the development of GANs & graph models could also help.
I really want to do more RL however I can't actually code lmao so even if I had the right model I wouldn't have the right inputs.
>RTS games have promised to feature an "AI that learns from you", yet it always turns out to be bullshit.
If you think you want that, you're dumb as shit. Nobody is gonna want to play an unbeatable game, which is what every game that has that AI will become.
> "AI that learns from you"
I don't see how it isn't possible, however optimising it would be difficult.
1. Game saves your replays against the "ai" and either uploads them to a central database or locally and trains an agent against it
2. Game gives you a new ai that is slightly different, still within *reasonably similar skill range*
3. Allow options to limit play types & handicaps or else it will always cheese turtle gays and its apm would be in the thousands as it perfectly micros each unit.
4. People will complain about the "ai" always knowing how to beat them because they don't know shit about "ai"
5. People will complain that their "ai" that only turtles is either too boring or too good at macro.
compstompers don't want good ai, they want to feel smart.