Why games don't use AI HW?

[GreiverBlade]

i don't think it would be a good idea and developers too it seems.
because i remember reading that some games (iirc Tomb Raider for example) had to be dumbed down in some country (more for the puzzle difficulty, AI on a lesser extent)

although nonetheless HW driven AI would not play well with a difficulty slider (or selector)

That's what I see online pvp haha XD, no AI needed.

ahahahah that's why i mostly stick to solo games or play "lone wolf"in pve and avoid pvp in mmo, because .... "i do not like having opponent/teamate dumber than an AI" ofc it does not happen all the time ahah

Developers are just getting into it. It's a bit premature as the tech is pretty young for the gaming industry.

i hope they would not go into it further ... imho the closer to be "human" an AI is, the more boring she is, thus i pray for a good implementation if they do it

 

[BSim500]

DLSS, RT are all nice and stuff, but let's be honest, games have had good graphics for over decade now, meanwhile in-game AI hasn't particularly evolved much since early 2000s and often current games have worse bot AI than some games from even 2001. Why is that?

Games peaked years ago when developers were actually creative. I still think back to Monolith Lithtech Engine games (the flanking communication in FEAR, or how enemies would appear to react to a patrol dog that wanted to deviate from a patrol route to start tracking your footprints in the snow in No One Lives Forever games). And those are 20 year old games on single core Pentium 3-4 era CPU's, ie, believable AI is really not a lack of hardware problem, it's an effort problem. If it doesn't involve sparkly bits that can't be screenshotted, they will put in the minimum effort. Same with audio subsystems and how we went from the masterpiece that was Thief (1997) audio propagation physics in the Dark Engine on an Aureal 3D card to this 2014 train wreck after 17 years of 'evolution'...

 

[Mussels]

It's not that kind of AI
It's a "we ran a bunch of tests and found THIS hardware setup is really fast at specific tasks" and then they use it for those tasks (RTX, DLAA/DLSS, etc)

AI is a generic term that means a lot of different things, and they're often massively, completely different.

 

[The red spirit]

believable AI is really not a lack of hardware problem, it's an effort problem.

I just wouldn't be so harsh on developers. Believe it or not, but many game development companies aren't swimming in cash and in fact just barely break even. It's really tough to make money now and you can see that in their annual reports. It's true that some boring company like Procter and Gamble has bigger profit margins than most of game companies. Besides that games have to prioritize some things in games and usually only few things while ignoring something else. For example, Assetto Corsa has great physics, but visually it doesn't looks brilliant, nor it has any campaign and then you have Tomb Raider, which looks brilliant, but has poor gameplay. Thing is the game must have its niche and execute well in it, however there's another thing. Every game also can't be hilariously behind others even in less important aspect of it as that alienates potential customers from it. This is just how game market works.

I personally don't care too much about graphics detailedness much more and I think that games basically got reasonably good enough at everything they should do around 2005, but I'm a superniche of this market it seems and most people wouldn't be okay with that or at least that's what companies think. I think that gameplay peaked in early 2000s basically due to graphics cards facing largely diminishing returns in terms of improvements, market was also immature and nobody really knew how it worked, so focus was on gameplay. Also most dev studios were much smaller, with lower overhead, lower risks of going under and current concept of games basically being more like interactive movies just wasn't a thing (but Phillips CDi was a bizarre aberration and yes it did have Tomb Raider game which worked with every DVD players and I played it), therefore they had to offer something cool in action, strategy, story or whatever else that wasn't strictly graphics. The changing climate in game dev market definitely affected games and most AAA games now have intense focus on graphics, however, some "old school" games just simply don't benefit much form better hardware, but that's not to say that they disappeared, they evolved, but in different ways. So, tldr; is that effort is still being put, but focus of industry shifted and times have changed.

 

[Dirt Chip]

You can, given enough resource, to program AI game that send a log of battels\encounter that happend during the play, add it to every other play by other players out there to preform a learning network for the AI game over time. The game will change it`s respond in encounter to all players according to the best x encounters from all the pool and so on.
The game will get more difficult the more people are playing, so in the end you will have monsterus AI, like in chess.

 

[The red spirit]

You can, given enough resource, to program AI game that send a log of battels\encounter that happend during the play, add it to every other play by other players out there to preform a learning network for the AI game over time. The game will change it`s respond in encounter to all players according to the best x encounters from all the pool and so on.
The game will get more difficult the more people are playing, so in the end you will have monsterus AI, like in chess.

Since you mention chess, could chess bots be improved with those cores?

 

[dragontamer5788]

I think that's incorrect. It should be the most easiest thing for programmer to write is 100% accurate bot, meanwhile programming failures is much harder, as well as adding a lot of nuance or even some "personalities" to different bots. And I think that believable flaws that are typical to humans can create a very good bots in games. Also some games, like racing sims would benefit a lot from very good driver AI.

Unbeatable FPS bots exist, and have existed for decades.

Strong AIs to turn-based games exist. In fact, most turn-based games (Chess, Go), have been brought to superhuman levels with regards to AI, and it doesn't even take that much effort anymore to Alphabeta-prune and/or Monte-Carlo + Neural Net an AI of superhuman capabilities.

No one wants to play vs these. In fact, anyone using say... "Aim-bots" is considered a cheater in these games. And chess players regularly call each other cheaters if they suspect AI usage.

No one wants to play vs a superhuman AI. Its just human nature.

----------

If we're making a "flawed AI", then it doesn't matter how strong it is, and the game is balanced in other ways. See Fire Emblem, Wargroove, Halo, Command and Conquer. These AIs are designed to suicide-charge into the human player's units. And aside from rare "scripted" events, they "never retreat" and "always fight to the death". Because human players don't like enemies who run away and harass them like a Minuteman from 1776.

Since you mention chess, could chess bots be improved with those cores?

Very few people care about improving Chess AI because they went superhuman in the 90s... and open-source implementations of those AIs became cheap in the 00s. Who cares if a chess AI is 3000 Elo, 3500 Elo, or 3700 Elo? Its all superhuman.

Its so far into the realm of superhuman its not even a fair fight. No human would play against it seriously. The only purpose of these bots is to generate "study" material, such as opening theories or endgame theories, but few Chess players reach the level of skill where this is useful to improve their gameplay.

 

[The red spirit]

Very few people care about improving Chess AI because they went superhuman in the 90s... and open-source implementations of those AIs became cheap in the 00s. Who cares if a chess AI is 3000 Elo, 3500 Elo, or 3700 Elo? Its all superhuman.

Well, even those "superhuman" bots actually got beaten and i think one grandmaster specifically trained with those and can sometimes beat them. That's really nice achievement for human, but makes me think if chess bots aren't a bit flawed.

 

[dragontamer5788]

Well, even those "superhuman" bots actually got beaten and i think one grandmaster specifically trained with those and can sometimes beat them. That's really nice achievement for human, but makes me think if chess bots aren't a bit flawed.

The last time I heard of this rumor was maybe Stockfish 7 or so... maybe 2015-era chess AI had some flaw that humans could take advantage of in **RARE** games. Like a grandmaster would have to play 50 games, and one of those games the flaw would appear.

Ever since Stockfish NNUE came out, those flaws seem to have been fixed.

But these aren't "plays". This is basically masochism. Its not a game in the traditional sense.

You can, given enough resource, to program AI game that send a log of battels\encounter that happend during the play, add it to every other play by other players out there to preform a learning network for the AI game over time. The game will change it`s respond in encounter to all players according to the best x encounters from all the pool and so on.
The game will get more difficult the more people are playing, so in the end you will have monsterus AI, like in chess.

No. Anyone who creates a learning AI can generate games (and eventually superhuman games) using the learning AI alone. No humans needed.

There's no point to using human games as the basis. An AI is itself, the basis for learning new techniques and tactics. Version 95 of the AI will generate a bunch of positions, and you use those positions to generate AI#96. Then AI#96 creates a bunch of positions, use those to create AI#97. Etc. etc.

100,000 generations later, you get... LeelaZero / AlphaZero.

Honestly, the real benefit of the AlphaZero team was recognizing that you don't need to play a whole game to generate training data, but instead "Position + Win/Loss" information. Every time you win or lose a match, tag **ALL** positions of the game with "win" and "loss". Then the "training" is just the AI memorizing which positions eventually led to a win, and which positions eventually led to a loss... shortcutting the entire "generate games" situation and accelerating the AI's learning speed. For various reasons, this is much better for artificial neural nets.

 

[Dirt Chip]

I'm way off as an AI pro, but the date from heumen play (position) sound priceless.
But as said, it is good mostly as "masochism" level of difficulty

Edit: You can use ONLY heumen play to train the bot, so you get a much more reasonable hard difficulty that challenge but not crush you completely. Basicky fight against the pool of all other people out there.

 

[Aquinus]

Honestly, the real benefit of the AlphaZero team was recognizing that you don't need to play a whole game to generate training data, but instead "Position + Win/Loss" information. Every time you win or lose a match, tag **ALL** positions of the game with "win" and "loss". Then the "training" is just the AI memorizing which positions eventually led to a win, and which positions eventually led to a loss... shortcutting the entire "generate games" situation and accelerating the AI's learning speed. For various reasons, this is much better for artificial neural nets.

That's also where the whole performance versus accuracy issue comes into play. The advantage of AI is having a complex problem broken into a simple one where data can guide outcomes. The quality of the data determines the quality of the outcomes, assuming the premise of the algorithm is sound. The more data you have, the more computational effort is required to make these determinations. So you essentially end up with performance and accuracy being inversely proportional. Basically you can't have your cake and eat it too.

 

[Mussels]

Since you mention chess, could chess bots be improved with those cores?

No. You'd have to design hardware specifically optimised for the type of equations used for chess


They could let a server farm crunch away with emulated hardware. Get an existing core design and emulate lots of variants of it with different caches sizes, pipelines, shader counts, etc etc and then when the simulation is done "our program found that this particular config was the fastest, this was the most energy efficient and this was the most cost effective" and then tada, that's your "AI" designed GPU

You can do the same with how DLSS and such worked: It's video encoding, but making a half-way in between frame. They did it in software, then designed hardware to do that software - because it's constantly using one of many methods and techniques its more than a simple stretching algorithm so they glorify it as "AI hardware"

You cant get something made to render an in-between shot of two frames and use it to play chess, it'd just give you an image of chess pieces not moving to go between the existing images.

 

[AusWolf]

Yeah, I'm aware, also within smaller labs with workstation cards too, however nVidia has been very vocal about AI on RTX hardware since RTX 2000 series and those AI capabilities which they talk about almost don't have any use case. DLSS is probably the only one thing that came out of it and it leaves me wondering "that's all it can do?".

Agreed. I'm still on the opinion that DLSS (and upscaling in general) is only for people who aren't happy with the performance they're getting at the level of detail they're expecting. There is no reason why it should be a selling point on anything more powerful than a 3070.

Besides, Intel CPUs come with some AI crap as well, or isn't the GNA something similar? I know that my Rocket Lake i7 has one, but I have zero idea what it's for.

I think that's incorrect. It should be the most easiest thing for programmer to write is 100% accurate bot, meanwhile programming failures is much harder, as well as adding a lot of nuance or even some "personalities" to different bots. And I think that believable flaws that are typical to humans can create a very good bots in games. Also some games, like racing sims would benefit a lot from very good driver AI.

Racing is probably the only game type where I'm missing better AI from time to time. In any other genre, upping the difficulty level usually does the trick. When it comes to friendlies, voice acting and animations give them personality, not the way they shoot. Or if you're thinking about bots as in games like CS:GO, if players cared about personality in the game at all, they would be playing something else altogether.

 

[Mussels]

I thought of a better example: Siri, Hey google etc. These are all "AI" but it's literally just hardware designed to convert sounds into written speech, for as low a performance cost (CPU cycles/power consumption) as possible.

That could never run a games AI, or play chess, or upscale graphics engines. It's a buzzword for "hardware acceleration" in the modern era since stuff like "Yeah this phone supports AVX-512 and SSE2 with MMX instructions" wouldnt market very well vs "quantum AI magic voice controls"

 

[Calenhad]

Sure you could use "AI" hardware such as the Tensor cores in RTX GPUs to train game AI as you play. But a significant part of balancing a game is knowning how the bots behave. Constantly evolving bots can have highly unpredictable results on playability. So it would be a significant task to implement something like this without the possibility of adverse effects. Meaning development costs will go up. Depending on the quality of the algorithm, you could in fact very often end up with "smart" bots that behave like total morons. Just get one of your starting variables wrong by 0.0001 and that could be the result in 1 of 1000 playthroughs.

No. You'd have to design hardware specifically optimised for the type of equations used for chess

No, just no.

Training AI bots to play chess is a student level task (making them play well is another matter). Any bot algorithm with CUDA support can use those tensor cores. Training an algorithm to play chess is considered a solved problem by now. There is a reason we have AlphaGO not AlphaCHESS.

You can do the same with how DLSS and such worked: It's video encoding, but making a half-way in between frame. They did it in software, then designed hardware to do that software - because it's constantly using one of many methods and techniques its more than a simple stretching algorithm so they glorify it as "AI hardware"

You cant get something made to render an in-between shot of two frames and use it to play chess, it'd just give you an image of chess pieces not moving to go between the existing images.

Saying DLSS is video encoding is over-simplifying things. There is a reason the game need to implement support for DLSS. And it is not so the driver can take frame n1 + n2 from the frame buffer and do "video encoding" on them. Otherwise DLSS would be just another toggle in the driver panel next to Image Scaling. This is not a dumb upscaler, it is a rather clever dynamic upscaler which require integration with the game engine. This was not possible in real-time before tensor cores. And since they sit idle in the GPU while you game anyway, that is a perfect match.

"AI" hardware just means it is designed specifically for that task, and is therefore faster than general purpose hardware. E.g. CPU cores vs Tensor cores. Tensor cores was a big speed improvement coming from pre-RTX hardware for ML tasks. Generally speaking most ML tasks can be run on any hardware. It is just a question of how soon you want your results. Want to run DLSS on a Pentium II? Sure. Just don't expect to be alive when you've calculated two frames. (This is also blindly ignoring other requirements such as memory, storage, ...)

 

[Vayra86]

I think that's incorrect. It should be the most easiest thing for programmer to write is 100% accurate bot, meanwhile programming failures is much harder, as well as adding a lot of nuance or even some "personalities" to different bots. And I think that believable flaws that are typical to humans can create a very good bots in games. Also some games, like racing sims would benefit a lot from very good driver AI.

Dunno man, look at how the meta-game develops in certain games, you know, the ones that are more than a simple shooter.

Its a crowd based thing and humans are nearly self learning AI in that sense in how they adapt. We recognize patterns, especially if we're into a specific activity all the time. And we have the data too. There are oceans worth of spreadsheets and online build tools to stay on top of the meta in numerous games.

Do you want your AI to be like that? You'll bring competitive online gameplay into the single player experience, while the extended control you have in SP is exactly what makes it tick: its your world. And if the AI only learns based on your meta, it'll be the perfect opponent, not the weaker one. And if it teaches itself to be weaker, you'll have an easy time.

Nah, good gameplay in non competitive singleplayer (any single player) relies on good gameplay elements. Perhaps small parts of a programmed AI could be truly adaptive, and they already are today. But even those adaptations are pre-cooked even if they seem dynamic and responsive. The moment you make it fully dynamic, is the moment you lose the control you need to keep the experience as described.

 

[The red spirit]

Agreed. I'm still on the opinion that DLSS (and upscaling in general) is only for people who aren't happy with the performance they're getting at the level of detail they're expecting. There is no reason why it should be a selling point on anything more powerful than a 3070.

I tested it in Horizon 5 and it tanks visual quality so much, that it's basically the same as just using lower resolution anyway, but with more overhead. Speaking of that, AMD's RIS is still borked in most games. I only seen it working in Genshin Impact, but nowhere else. And if you crank it generously, it can make lower resolutions more bearable, more importantly it just works better than FSR. Here's a fun thing, my secondary computer's monitor has "upscaling" like RIS already built in. That makes me really question why AMD just doesn't get their shit together.

Racing is probably the only game type where I'm missing better AI from time to time. In any other genre, upping the difficulty level usually does the trick. When it comes to friendlies, voice acting and animations give them personality, not the way they shoot. Or if you're thinking about bots as in games like CS:GO, if players cared about personality in the game at all, they would be playing something else altogether.

No, it's mor eabout aiming, tactics and bots not feeling like they only have one accuracy slider cranked or lowered in them. CS:GO has an absolutely awful bots, that can be worse than 1.6's. I also disagree that bots are already good enough in many games, hell even Call of Duty struggles with bots and there are games, where team work is important or game just has many players. And I don't see really well made bots often, they are rare and plenty of games just put them as an afterthought.

 

[AusWolf]

Dunno man, look at how the meta-game develops in certain games, you know, the ones that are more than a simple shooter.

Its a crowd based thing and humans are nearly self learning AI in that sense in how they adapt. We recognize patterns, especially if we're into a specific activity all the time. And we have the data too. There are oceans worth of spreadsheets and online build tools to stay on top of the meta in numerous games.

Do you want your AI to be like that? You'll bring competitive online gameplay into the single player experience, while the extended control you have in SP is exactly what makes it tick: its your world. And if the AI only learns based on your meta, it'll be the perfect opponent, not the weaker one. And if it teaches itself to be weaker, you'll have an easy time.

Nah, good gameplay in non competitive singleplayer (any single player) relies on good gameplay elements. Perhaps small parts of a programmed AI could be truly adaptive, and they already are today. But even those adaptations are pre-cooked even if they seem dynamic and responsive. The moment you make it fully dynamic, is the moment you lose the control you need to keep the experience as described.

I couldn't have said it better.

SP games are defined mostly by atmosphere. Well-written scripts and visuals are a more helpful tool in creating that than AI. MP games are made (or ruined) by human interactions, so AI is not generally needed.

I tested it in Horizon 5 and it tanks visual quality so much, that it's basically the same as just using lower resolution anyway, but with more overhead. Speaking of that, AMD's RIS is still borked in most games. I only seen it working in Genshin Impact, but nowhere else. And if you crank it generously, it can make lower resolutions more bearable, more importantly it just works better than FSR. Here's a fun thing, my secondary computer's monitor has "upscaling" like RIS already built in. That makes me really question why AMD just doesn't get their shit together.

I played Cyberpunk with DLSS Quality. It didn't bother me much, but anything lower really destroys image quality.
That aside, my point is: if you have enough FPS, you don't need DLSS or FSR or any other black magic. Why it is a selling point on high-performance Nvidia cards is beyond me.

No, it's mor eabout aiming, tactics and bots not feeling like they only have one accuracy slider cranked or lowered in them. CS:GO has an absolutely awful bots, that can be worse than 1.6's. I also disagree that bots are already good enough in many games, hell even Call of Duty struggles with bots and there are games, where team work is important or game just has many players. And I don't see really well made bots often, they are rare and plenty of games just put them as an afterthought.

I see what you mean, but I'm not sure it would give story-driven games too much extra. As for CS:GO and COD, they're multiplayer games, so adding bots is naturally not the main focus.

 

[The red spirit]

Dunno man, look at how the meta-game develops in certain games, you know, the ones that are more than a simple shooter.

Well I play Victoria 2, which is grand strategy game and it simulates whole world. You get a simplified simulation of all countries, cultures, economy, politics, wars, religions and some other things. Also every single human is simulated with their own specific preferences, issues, disatisfactions and etc. The game sis really easy to modify and while normal games lasts 100 years, you can set it to anything in configuration files. Obviously performance sucks as even after 60 years, as it is 32 bit app and only supports 1 core or 2 cores tops and crashes if it exceeds 4GB RAM usage. It doesn't run well in late games on any hardware. But anyway, I had patience only for 200 years so far and developments were interesting on map, but they didn't deviate from game's predisposed biases for bots, like Germany forming or UK staying number 1 in stats or Africa ending up colonized. So that just sounds like programmabnle thing and yo ucan avoid some weird game corruptiosn from bots doing weird shit. My point is that more advanced bots could be more amusing to play with and the more random they get, well the more fun I hope. Game like OpenTTD has community created bots and creativity there is insane and community managed to pull off som really sophisticated bots, despite the fact taht game itself is ancient, uses very little resources and etc.

And RTX card should be more than capable of runing bots and it could randomize them on the fly too.

Do you want your AI to be like that? You'll bring competitive online gameplay into the single player experience, while the extended control you have in SP is exactly what makes it tick: its your world. And if the AI only learns based on your meta, it'll be the perfect opponent, not the weaker one. And if it teaches itself to be weaker, you'll have an easy time.

Do you play shooters? Bots there universally suck. Do you play Horizon 5? Bots there suck too and the game has awful rubberbanding. And if it would be possible to bring MP like experience to SP, I would say "shut up and take my money".

Nah, good gameplay in non competitive singleplayer (any single player) relies on good gameplay elements. Perhaps small parts of a programmed AI could be truly adaptive, and they already are today. But even those adaptations are pre-cooked even if they seem dynamic and responsive. The moment you make it fully dynamic, is the moment you lose the control you need to keep the experience as described.

Most games are to some extent competent and some require smart bots, not the ones with just cranked accuracy or power stats.

 

[Vayra86]

Do you play shooters? Bots there universally suck. Do you play Horizon 5? Bots there suck too and the game has awful rubberbanding. And if it would be possible to bring MP like experience to SP, I would say "shut up and take my money".

I played shooters before they turned into a yearly shareholder's feast, such as UT'99, which had outstanding bots with multiple difficulty levels. Compared to today, yes, I agree, current day shooter bots are an afterthought, because the games are MP focused and cost is kept to a minimum.

That's a benchmark for me, and unfortunately very few games managed to reach that after it.

What it shows is that you really don't need AI to make interesting NPCs.

Randomization in AI is not AI by the way, its randomization based on a set number of parameters. Its as simple as: RNG, a roll of the die, which is effectively as simple as a hit rate percentage for anyone. You can hit or you can miss, boom its randomized. You can add situations, but its the same thing.

If you want an AI to be better (more dynamic!) than that, you'll be forced to make it act without those rules/limitations. Now, imagine a game with Fog of War, or limited knowledge of enemy positioning. AI can either cheat by knowing more (escaping limitations you have, it happens all the time and makes an AI 'feel unfair' to us, its the same thing as giving it more resources or infinite resources); or it can be on equal footing which means it will extract advantage by conservative positioning. Games might just go completely borefest instead of being a new challenge, the AI will be evasive because it wants to know more. Unpredictability doesn't mean a higher fun factor

 

[The red spirit]

I played Cyberpunk with DLSS Quality. It didn't bother me much, but anything lower really destroys image quality.
That aside, my point is: if you have enough FPS, you don't need DLSS or FSR or any other black magic. Why it is a selling point on high-performance Nvidia cards is beyond me.

At least DLSS does upscaling better and does it smartly, meanwhile FSR is basically just sharpening and bad one at that. DLSS is also cool to play with. So perhaps it does reduce quality in games, but nV pushes it, becasue it brings awareness to it, not because you should use it everywhere. Basically anythign nVidia says or shows themselves is tech demo, which is meant to inform buyers on what they have been working on and what stuff is. Also nV pays reviewers or at least sends sample cards awith instructions of what to say, more precisely what to mention, what's perhaps buggy and should be hidden in videos and etc. Launch day "reviews" are basically just a massive pro-nVidia propaganda with a lots of bias to talk positively about them. Some reviewers handle them badly, like JayZ, who days before launch said it's crap in million ways and after review it was the best shit ever. Yep, he was a shill and that's the way all media works, you say good things, you get free samples and moneys, you talk crap and be real, no more free samples for you and fuck your revenues. Not to say that media doesn't try to talk more realistically about products, but they basically have to skirt around various rules, restrictions set by nVidia to do so. In fact nVidia sometiems does better job on their channel for being unbiased than some reviewers, because nV just says "hey we made this new tech, this is how it works, oh we also made things faster and that's cool and this is how much it costs", meanwhile media can be "Holy shyet, performance is insane, you basically have to buy it now, else you are a retard, also fuck wattage and crank it way more than from factory, becasue only fps matters to drive your overpriced monitorz". In the end finding independent media is hard and the best idea is to just find media that doesn't try to push anythign on you and just shows objective results and discloses lots of stuff. That's why something like Notebookcheck is an excellent review site (unfortunately they don't cover GPUs or CPUs much), by just showing things and just letting you decide for yourself if they matter to you. Anyway, Ithink that DLSS is jsut fun to play around with and unlike FSR doesn't feel so much like a crutch to achieve playable framerates, more like a cool tech for superfulous stuff to try.

I see what you mean, but I'm not sure it would give story-driven games too much extra. As for CS:GO and COD, they're multiplayer games, so adding bots is naturally not the main focus.

??? COD is mainly SP game. CS:GO is obviously not, but bots are essential, when lobbies ar ehalf empty and their quality matters a lot. Meanwhile many games are like Battlefield, basically a big maps and somewhat tactical gameplay with mates to achieve simple predefined goals (overthrowing outposts). Arma is like BF, but with more accurate simulation and more serious gameplay. And more shooting games today are open field tactical shooters, than strictly on rails shooter like CoD, so good AI would be great.

I played shooters before they turned into a yearly shareholder's feast, such as UT'99, which had outstanding bots with multiple difficulty levels. Compared to today, yes, I agree, current day shooter bots are an afterthought, because the games are MP focused and cost is kept to a minimum.

That's a benchmark for me, and unfortunately very few games managed to reach that after it.

I actually played a lot of UT2004 as kid and offline. I never played UT99 much, since it is PITA to make it run on modern hardware, hell it's PITA to even run it on XP with Athlon 64 and ATi X800 Pro card. I just gave up on making it run. I eventually got to play it, but it felt like older version of UT 2004, so it's just more bare game than radically different one. Either wya, it's an arena shooter, this whole genre was dead for a darn long time and now has changed much.

What it shows is that you really don't need AI to make interesting NPCs.

Randomization in AI is not AI by the way, its randomization based on a set number of parameters. Its as simple as: RNG, a roll of the die, which is effectively as simple as a hit rate percentage for anyone. You can hit or you can miss, boom its randomized. You can add situations, but its the same thing.

I strongly disagree, good AI isn't RNG, or at least not so unsophisticated RNG like in those UT games.

If you want an AI to be better (more dynamic!) than that, you'll be forced to make it act without those rules/limitations. Now, imagine a game with Fog of War, or limited knowledge of enemy positioning. AI can either cheat by knowing more (escaping limitations you have, it happens all the time and makes an AI 'feel unfair' to us, its the same thing as giving it more resources or infinite resources); or it can be on equal footing which means it will extract advantage by conservative positioning. Games might just go completely borefest instead of being a new challenge, the AI will be evasive because it wants to know more. Unpredictability doesn't mean a higher fun factor

Unfair AI is waht we have how and it sucks, I don't see how improvement to it could make it worse, becasue current AI isn't really that good to begin with. Most racing games have cheating AI, many shooter games have clearly cheating AI too. I would like to see something that is fresh and isn't that poor.

 

[Vayra86]

At least DLSS does upscaling better and does it smartly, meanwhile FSR is basically just sharpening and bad one at that. DLSS is also cool to play with. So perhaps it does reduce quality in games, but nV pushes it, becasue it brings awareness to it, not because you should use it everywhere. Basically anythign nVidia says or shows themselves is tech demo, which is meant to inform buyers on what they have been working on and what stuff is. Also nV pays reviewers or at least sends sample cards awith instructions of what to say, more precisely what to mention, what's perhaps buggy and should be hidden in videos and etc. Launch day "reviews" are basically just a massive pro-nVidia propaganda with a lots of bias to talk positively about them. Some reviewers handle them badly, like JayZ, who days before launch said it's crap in million ways and after review it was the best shit ever. Yep, he was a shill and that's the way all media works, you say good things, you get free samples and moneys, you talk crap and be real, no more free samples for you and fuck your revenues. Not to say that media doesn't try to talk more realistically about products, but they basically have to skirt around various rules, restrictions set by nVidia to do so. In fact nVidia sometiems does better job on their channel for being unbiased than some reviewers, because nV just says "hey we made this new tech, this is how it works, oh we also made things faster and that's cool and this is how much it costs", meanwhile media can be "Holy shyet, performance is insane, you basically have to buy it now, else you are a retard, also fuck wattage and crank it way more than from factory, becasue only fps matters to drive your overpriced monitorz". In the end finding independent media is hard and the best idea is to just find media that doesn't try to push anythign on you and just shows objective results and discloses lots of stuff. That's why something like Notebookcheck is an excellent review site (unfortunately they don't cover GPUs or CPUs much), by just showing things and just letting you decide for yourself if they matter to you. Anyway, Ithink that DLSS is jsut fun to play around with and unlike FSR doesn't feel so much like a crutch to achieve playable framerates, more like a cool tech for superfulous stuff to try.


??? COD is mainly SP game. CS:GO is obviously not, but bots are essential, when lobbies ar ehalf empty and their quality matters a lot. Meanwhile many games are like Battlefield, basically a big maps and somewhat tactical gameplay with mates to achieve simple predefined goals (overthrowing outposts). Arma is like BF, but with more accurate simulation and more serious gameplay. And more shooting games today are open field tactical shooters, than strictly on rails shooter like CoD, so good AI would be great.


I actually played a lot of UT2004 as kid and offline. I never played UT99 much, since it is PITA to make it run on modern hardware, hell it's PITA to even run it on XP with Athlon 64 and ATi X800 Pro card. I just gave up on making it run. I eventually got to play it, but it felt like older version of UT 2004, so it's just more bare game than radically different one. Either wya, it's an arena shooter, this whole genre was dead for a darn long time and now has changed much.



I strongly disagree, good AI isn't RNG, or at least not so unsophisticated RNG like in those UT games.



Unfair AI is waht we have how and it sucks, I don't see how improvement to it could make it worse, becasue current AI isn't really that good to begin with. Most racing games have cheating AI, many shooter games have clearly cheating AI too. I would like to see something that is fresh and isn't that poor.

I'll give up, either we have a language barrier, or you just want to disagree/misunderstand what I'm saying. I understand what you want, I'm explaining why what you want isn't feasible (or at least: why self-learning artificial intelligence is the wrong tool to get there).

 

[The red spirit]

I'll give up, either we have a language barrier, or you just want to disagree/misunderstand what I'm saying. I understand what you want, I'm explaining why what you want isn't feasible (or at least: why self-learning artificial intelligence is the wrong tool to get there).

Your arguments are basically "because current AI is made like this", which is okay, but AI HW opens new possibilities and new ways to create it in games.

 

[Vayra86]

YOur arguments are basically "because current AI is made like this", which is okay, but AI HW opens new possibilities and new ways to create it in games.

No, its pitting the philosophy of current AI against what you're proposing and showing you the problems that has.

There aren't a whole lot of ways to go here, its all about NPCs using the game logic, those are the boundaries. Or perhaps you and I have a radically different idea of what is possible with 'AI HW'. Enlighten me

 

[The red spirit]

Or perhaps you and I have a radically different idea of what is possible with 'AI HW'. Enlighten me

I won't but do yo ureally think that thousands of devs with budget wouldn't come up with anything better tha nrubberbanding or caranking one stat up, when bots lose? Devs hav ealready utilized weird tech like voxels, PhysX and etc. I just think that RTX AI HW could achieve something fresh, without taking CPU performance and having a lot more resources than ever before.