14 replies to this topic

[Subhntr]

Your lancemates will yell at you if you accidentally clip one of them with your shots, but I've noticed that they happily fire if one of your other lancemates is in the way. I've watched in horror as these idiots proceed to put volley after volley into each other if they're in the way, tearing off arms and coring out lancemates. It's gotten particularly egregious as the tonnage and firepower have gone up and the speed goes down.

I shouldn't have to micromanage my lance commands just because 'don't shoot friendlies' was too much effort to code.

[Ilfi]

Let me know when they stop walking directly through our Defense objectives. It still feels like they'll never not unless I forcefully tell skirt them around cities with NAV points.

[Zerstorer Stallin]

Console AI abilities detected.

[Vellron2005]

Subhntr, on 19 December 2019 - 05:26 PM, said:

Your lancemates will yell at you if you accidentally clip one of them with your shots, but I've noticed that they happily fire if one of your other lancemates is in the way. I've watched in horror as these idiots proceed to put volley after volley into each other if they're in the way, tearing off arms and coring out lancemates. It's gotten particularly egregious as the tonnage and firepower have gone up and the speed goes down.

I shouldn't have to micromanage my lance commands just because 'don't shoot friendlies' was too much effort to code.

Not only that, but your lancemates are perfectly content to plow their mechs through buildings of a base you're supposed to defend, or better yet, try to take out a VTOL out of the sky with a PPC shot straight through a skyscraper.

You just gotta love em' right? I wish I could take it out of their pay

[Dee Eight]

So they behave exactly like real humans... especially ones who play MWO...

[Daggett]

Zerstorer Stallin, on 19 December 2019 - 08:49 PM, said:

Console AI abilities detected.

Empty claim detected.

[7Zebras]

I have to agree. When my teammate is blocking my shot I see it change to a big green "DON'T SHOOT" reticule. Which suggest that there is a mechanic in the game that can tell the difference between friend or foe. So I guess we are figuring out how to tell the AI teammates that red=good and green=bad. I can see how AI would have trouble with it. Odds are the AI has been driving on roads for years and is having trouble going from "green means go" to "green means don't core out the back of your teammates."

I am sure this is being works on. Actually, I am not. I am hoping the AI continues to be horrible so that co-op play is encouraged.

[Nyte Kitsune]

For those defense missions I find If you send them to stand at a specific position they rarely stray too far from where you sent them and attack anything that comes within their field of vision (i.e. Not hidden behind and object). I generally have one pilot stationed in the N,W,E,S locations around (including myself) the objective area and cover them as needed. My best at defense so far is 91% and worse being 57% (Most of which was my fault when I decided to have them join up on me and damage I caused myself)

[overheal]

Zerstorer Stallin, on 19 December 2019 - 08:49 PM, said:

Console AI abilities detected.

As if there is some fundamental difference in how AI acts reliant upon which platform its run on

[Sabertooth1966]

They aren't completely oblivious to the buildings when on defense. I have seen them trying to maneuver around the small buildings when the enemy gets inside it. Though they will still sometimes clip the buildings or hit them with fire when shooting at an enemy near it. I've just been keeping them with me and moving around the perimeter of what we are trying to defend as much as possible to avoid collateral damage to the buildings.

However if you step infront of them they will usually end up shooting you at least once.

Edited by Sabertooth1966, 20 December 2019 - 10:26 AM.

[Daggett]

Sabertooth1966, on 20 December 2019 - 10:23 AM, said:

Though they will still sometimes clip the buildings or hit them with fire when shooting at an enemy near it.

That's exactly what i do when i hunt enemies between buildings, assuming that i will prevent more building damage overall by removing enemies quickly even if it means some collateral damage.

[LordNothing]

Daggett, on 20 December 2019 - 06:44 AM, said:

Empty claim detected.

to be fair game ai hasn't really changed much in the last 30 years. no matter how much attention to detail your decision trees get, they are still just decision trees. game ais have yet to benefit from advances in artificial neural networks over the last 10 or so years. there is that neural net trained anti-aliasing that nvidia tried to do with their rtx line. but things like that are trained on server clusters in non-real-time. in theory there is no reason you cant use those very same tensor cores to do ai. but to the best of my knowledge, no dev has done that kind of thing. would come with the caveat that you need more expensive hardware to have smarter ai performance.

Edited by LordNothing, 20 December 2019 - 12:08 PM.

[Koniving]

Subhntr, on 19 December 2019 - 05:26 PM, said:

Your lancemates will yell at you if you accidentally clip one of them with your shots, but I've noticed that they happily fire if one of your other lancemates is in the way. I've watched in horror as these idiots proceed to put volley after volley into each other if they're in the way, tearing off arms and coring out lancemates. It's gotten particularly egregious as the tonnage and firepower have gone up and the speed goes down.

I shouldn't have to micromanage my lance commands just because 'don't shoot friendlies' was too much effort to code.

Those of us that "beta tested" noticed this the very first day.
The issue is that the AI does not recognize AI units, does not have a concept of "team", and every other AI unit friend or foe is recognized as an "object."

The world is a nav mesh with either "Object" or "Structure".
Both objects and structures can be targets.
A target is simply a designated object or structure that gives off a red Dorito.
Both objects and structures have a clause that if it is in the way, the AI should try to go around it an unspecified percentage of the time, with the remaining percentage being to go through it. Or, if the object persists in being in the way, go through it.

As such, objects (allies) are ignored when passing the line of fire, and the firing continues. If the obstacle is constantly in the way, they will actively target and destroy it until it is no longer in the way or until they teleport (gotta get over 1,500 meters away for them to be able to teleport). So if you ever find a lancemate actively trying to destroy another one it's because the two collided for over a minute where no matter what one did it couldn't get around the other.
(To note it is specified, but I haven't found it yet. However the behavior for obstacles is told to us by the developer that coded it in this video.)

But passively shooting through teammates will happen all the time. Don't worry, the enemies do it too.

Edited by Koniving, 20 December 2019 - 12:10 PM.

[Koniving]

LordNothing, on 20 December 2019 - 11:56 AM, said:

to be fair game ai hasn't really changed much in the last 30 years. no matter how much attention to detail your decision trees get, they are still just decision trees.

They have -- changed in the last 30 years.
They gained decision trees in that time.
Prior to that, they were running as Finite State Machines. Which some games still run, such as Batman Arkham Asylum.
Trees started to gain momentum closer to the later half of 2000s, as many PS2/Xbox games also ran finite state machines.


(Half-life's FSM is truly impressive.)

(There isn't "much" of a difference between FSM and trees, other than trees can run multiple decisions at once, do not need to check as often for realistic responses [so like Goldeneye N64 where the AI is doing something and you shoot him dead but he has to finish what he's doing before he dies is a result of it not checking its health until after it is done shooting, where now the instant the health is zero for a tree it's done right away without the need for a check because it's an event rather than a request to check status), are a lot easier to debug, visual, and so easy a child can do it instead of a programmer

The rest is accurate. There's no neural nets or actual intelligence. To call game AI as AI is fundamentally incorrect as there is no intelligence there, only simulated virtual reaction to stimuli.

Edited by Koniving, 20 December 2019 - 12:23 PM.

[LordNothing]

Koniving, on 20 December 2019 - 12:13 PM, said:

They have -- changed in the last 30 years.
They gained decision trees in that time.
Prior to that, they were running as Finite State Machines. Which some games still run, such as Batman Arkham Asylum.
Trees started to gain momentum closer to the later half of 2000s, as many PS2/Xbox games also ran finite state machines.


(Half-life's FSM is truly impressive.)

(There isn't "much" of a difference between FSM and trees, other than trees can run multiple decisions at once, do not need to check as often for realistic responses [so like Goldeneye N64 where the AI is doing something and you shoot him dead but he has to finish what he's doing before he dies is a result of it not checking its health until after it is done shooting, where now the instant the health is zero for a tree it's done right away without the need for a check because it's an event rather than a request to check status), are a lot easier to debug, visual, and so easy a child can do it instead of a programmer

The rest is accurate. There's no neural nets or actual intelligence. To call game AI as AI is fundamentally incorrect as there is no intelligence there, only simulated virtual reaction to stimuli.

thats what i mean, you got a lot of deterministic stuff, some rng, and some algorithms for path finding and the like. run of the mill game "ai" in a nutshell. therefore i'm not surprised or even bothered with mw5 ai is as it is. after all its one of the things pgi is trying to improve.

what im interested in is what comes next. eventually we are going to have enough parallel processing in our machines to be able to run pre-trained neural nets. actually thats kind of what DLSS does (running a neural net that is, for a completely different purpose), so at least for some people the hardware is already there. you have all these specialized cores designed for really fast 4x4 matrix math. this can be used by various neural net topologies. synthetic sensory data in, control data out. the neural nets themselves would be pre-trained on high performance hardware.

actual usage in games would probibly be as boss units (for example the mechs), with classic game "ai" running the fodder units (tanks, hovers, helis). you might even have your ai's learn as they progress. they might be trained against a generic player archetype. its loaded from a file at the start of the game. as you play more games it learns from your play styles and how to counter you, forcing the player to try different strategies to keep it off guard. multiple ais might start developing their own little personalities as you progress.

i have a feeling that within the next 10 years we will start to see proper neural net ai in games. core counts only keep going up, and video hardware is always getting faster. early tech demos will likely have minimalistic graphics because a lot of the video hardware is going to be occupied with ai tasks.

Edited by LordNothing, 20 December 2019 - 04:23 PM.