Jump to content
rawalanche

Suggestion about AI for future ArmA games

Recommended Posts

I believed that they tried a more autonomous AI in the past (ArmA1 I think?), but result with more complains. Player and AI often have different perspective about situations.

AI often broke off from the formation to engage enemies, refuse to move while under fire, or open fire too early and ruined ambush.

Edited by Lugiahua
  • Like 1

Share this post


Link to post
Share on other sites

im willing to bet a hundred bucks that arma AI is a literal clusterf*ck of code accumulated over the years and BI couldn't untangle their own mess even if they wanted to

Share this post


Link to post
Share on other sites
im willing to bet a hundred bucks that arma AI is a literal clusterf*ck of code accumulated over the years and BI couldn't untangle their own mess even if they wanted to

Please show us what proof you have of this, because I'm willing to take you up on that bet. Obviously you don't realize how much most AI mods rely on the vanilla system, nor do you check the AI changlog (albeit, that only gets updated once in a blue moon). The fact that people can use the current AI system, and mod it using scripts is testament enough that BI is doing something right for the AI.

Share this post


Link to post
Share on other sites

I don't think many of you realize just how complex the AI is.

 

The AI makes a threat assessment and then evaluates its options and picks the best one it is programmed with. I assume this is based on their own studies and their long experience with various military forces. 

 

There isn't really any game you can compare this with as you can literally place the AI in any terrain and any environment and it will behave much better than any other "AI" and many humans as well. The problem is that it is very CPU heavy, so on weak CPUs it doesn't have the time to assess and/or evaluate and falls in to a default mode which I assume is hit the dirt. Now I agree if you argue that the AI isn't good enough because of this, but it isn't because it's dumb - it's quite the opposite, it's too smart (or rather tries to be).

 

As for the idea of just having people randomly distributed in a circle, it's seems like a very bad idea. I doubt fire and maneuver has changed much the last 15 years since I was active and it's all based on experience and truly tried under fire so to throw that away for something that you think works better on the ground with (what I assume is) very little experience is not something I see as an improvement. Either you line up for fire and maneuver, be it attacking or pulling out, or you set up a perimeter in an area where you have cover and concealment. 

 

I think many people who object to the AI either has too weak CPU or lacking military experience/tactics and place themselves or the AI in a very disadvantageous position where they are very exposed. Even the most experienced and well trained soldiers would have a terrible day if caught in an ambush in the middle of an open field. 

  • Like 1

Share this post


Link to post
Share on other sites

I actually think Arma's AI performs extraordinary, compared to other games out there. (Controlling vehicles aside, their driving is just retarded.) They're capable of navigating and fighting in any terrain without the need for additional scripting or pre-programmed paths. If you observe AI squads they're actually quite smart, depending on how the mission designer set them up. Their weakness is that their behaviour is designed for fighting in open terrain, dense urban areas or closed spaces are what troubles them. But what really breaks the AI in my opinion is when you mix players and AI. I for example tend to break the formation or run too far ahead when being the only player in an AI-led squad. From the AI's perspective I must look like a complete moron. And it only gets worse when I'm in command, also due to the horrible command menu. I can't issue orders as fast as an AI commander, thus I get my squads killed on a regular basis. Plus I don't have any military experience and I'm not really interested in reading myself into it (but I watch Dslyecxi's guides, that's something ;)), so most of my tactical decisions are probably complete failures anyway.

 

The human-computer interaction in Arma hasn't changed much since OFP and BI seems to lack real concepts or solutions for that problem. The command interface is just a complete clusterfuck. Even this "quick interface" tends to be inaccurate. I lost track of how many times I accidentally issued a movement order instead of an attack order just because the enemy I was targeting moved away right before I clicked.

 

TLDR: I think the AI is only part of the problem. Instead, the interface to control them needs to be completely reworked.

Share this post


Link to post
Share on other sites

I concluded some time ago that the best (only?) way to sort stuff out is to have 2 kinds of AI.

 

Type A is more autonomous and will act according to the FSM/scripts wishes, with its own tactical goals and timeframe. Don't tell these guys what to do, just put them in your scenario with a few objectives, and leave them be. Give them high level commands but don't try to micromanage their behavior. "Go to that town and search the buildings"

 

Type B is more responsive, and will respond more readily to human interaction. Break contact the second its told to, go directly where its told to go, disregard its own safety, etc. Zero tactical 'brain', but will do what a human tells it to do immediately. Note, agents do not fill this gap. Give these ones low-level commands. "Go to that window and watch west, suppress any seen targets"

 

Full credit to the guys like oukej working on the AI, but trying to straddle those two types of AI and have them both autonomous and semi-responsive isnt the best of both types, its a murky, uncertain road leading into a quagmire.

  • Like 4

Share this post


Link to post
Share on other sites

im willing to bet a hundred bucks that arma AI is a literal clusterf*ck of code accumulated over the years

 

Sounds like the human genome :)

Share this post


Link to post
Share on other sites

You are right rekrul about the Ai in Arma3! ^_^

And in the past there is no other game that can compare with Arma 3 but i think its not about how complex the AI is but improving on it to make it more human like. As future games are heading towards this goal. But the AI is very complex and challenging already in the game. But always good to keep on improving it so you will never really know if your playing a player or not! Its not about making it so it become unbeatable and hit perfect shots. Like thing such as personality in the AI, ranking just creates more realism for the AI. The AI learns from all its experiences!

It remembers who it has played and improves! Players learn from the AI and adjust so it only would be fair so the AI lean and adjust to the challenge.

Instead of trying different way to beat the Ai an added challenge is put forward.

The challenge now is how long can you play with each other and keep winning? As you both are learning and adapting! Who will beat who? :) Is that not something so exciting to look forward too? You need to keep trying new things because old ways will be useless as the Ai will already be prepared for that! Sure its already exciting at the moment but with the thought that each game play is a new learning experience it something more exciting!

 

You can begin lessons to train your army so you really need to keep commanding them every step of the way. They would know how to follow you because they have been training with you for sometime!

Anyway if they can learn like a human with realistic sensory inputs instead of going up in percentages then perhaps learning will be more realistic and fair.

They can have ranks too and this is all from exp from the game so if they have good head shots its only because they have learnt how to make good head shots and done lots of it to earn their position. These experience could be pass onto other AI individuals to speed up learning process.

Each Ai will run as an individual instead of a collective but still obeying orders. This way each learns on their own and go up in rank. So you can have a group of AI with all different styles of play and experience. And this will really diversify the experience even more.

This has not been done in a tactical game before.
 

The AI could have a very bad dislike about you and if it sees you will try and do all it can to hunt you down if you annoy it too much!

Players could have an option on this so if some dislike it they can play with out it.

And with the AI in ARMA 3 RTS games are taking the lead on this and fps games are catching up! Their Ai can learn and store infomation in the cloud server or perhaps on your own comp and use it up when needed! Their Ai is the most exiting compared to most past and current FPS games! They really build tactics and will take you down and learn and adapt!

 

Like there was a situation of this stealth sniper at night gunning down a few AI in ARMA 3. I am not sure what version of the game it was but the AI did not stand a chance with one lone sniper! It seems the Ai were not coordinating properly but in an RTS this would never happen.  

 

They could be situations where you need to escape with a few soldiers in the forest where a few elite soldiers are also hunting you down! Be some mission to do! Tracking you down till they kill you because you stole something of theirs!

 

So there is always room for more improvement! :D And you wont need two types of AI if the Ai becomes very realistic like other players. They can carry out commands on their own because they have their own squad. And if you have Ai with your group and your leader or higher ranking you can command them to do certain things.

 

You wont need to micro manage them also when AI know what to do because of training as explain above. Just like in reality how its done.

 

 


Here some facts about DNA!

With odds standing at 1 chance in 10164 of finding a functional protein among the possible 150-amino-acid compounds, the probability is 84 orders of magnitude (or powers of ten) smaller than the probability of finding the marked particle in the whole universe! - Stephen Meyer

Share this post


Link to post
Share on other sites

Is that not something so exciting to look forward too?

Not exactly.

 

Computer game's AI should not be complex or hard to beat — it should be moderately challenging and interesting to compete against (noone likes to compete against a superior opponent).

AI behaviour should be predictable — otherwise it will be useless for campaigns/co-op missions (for it's main purpose that is).

AI should be cheap to develop — it's unlikely that Billy Gamerson's parents would give him $3000 to buy a new installment of his favorite computer game.

AI should be undemanding towards PC's resources — there is no point in AI if most of playerbase can't even play with it.

 

I'm afraid that advanched self-learning AI fails in all of these categories, especially in first and second. Too hard, too advanced or too autonomous AI would not be fun and "not fun" is basically a death verdict for a program that is designed to make people entertained.

  • Like 1

Share this post


Link to post
Share on other sites

 

Look how fluid this battle looks. I'd like an analysis from people better informed on AI than I. There are more AI vids on Youtube as well.

 

One aspect that I think will increase AI immersion in Arma is the animation. Soldiers still move like humanoid cyborgs. When they run there's no momentum to it, it's like a mech walking and ducking. No dynamic cover, still primitive ragdolls.

  • Like 1

Share this post


Link to post
Share on other sites

My suggestion about ArmA AI ? Go back to OFP AI model, simple, efficient and predictable, which has a believable dammage system, fleeing when defeated, which i can script the way i want. Get rid of the intricate and uselessly complex ArmA2/3 AI, unable to perform simple things, apart sniping me. The only thing missing in OFP AI is a FSM tool.

  • Like 1

Share this post


Link to post
Share on other sites

 

 

advanched self-learning AI

 

Frankly, personally I would right away buy game containing such a miracle just to see it, and most likely I would be more than satisfied just watching it in action. This however does not exclude the layer of human-like errors/behaviors/flaws. The more human-like AI, the better for me. This of course means also degree of unpredictability, and IMO it doesn't hinder any mission making or scripting if under strict control, modular/optional (not Arma's status quo case), and when mechanism of unpredictability are understandable/documented, thus may be taken into account during scenario designing. Keeping it modular, where each aspect/level is switchable BTW allows to adjust software computational weight to the hardware. Of course such beauty has to be carefully designed from the scratch and anyway scenarios assuming given level of sophistication may not work, if user will reduce it or vice versa.

 

Nice to dream about. 

  • Like 1

Share this post


Link to post
Share on other sites

 OFP AI bummed me out once I realized I could snipe the whole town from an upstairs tower without them coming for me -ever. Also Micropathfinding for Infantry is a major plus as they can now actually get behind cover rather than just get near it.

 

Computer game's AI should not be complex or hard to beat — it should be moderately challenging and interesting to compete against (noone likes to compete against a superior opponent).

AI behaviour should be predictable — otherwise it will be useless for campaigns/co-op missions (for it's main purpose that is).

AI should be cheap to develop — it's unlikely that Billy Gamerson's parents would give him $3000 to buy a new installment of his favorite computer game.

AI should be undemanding towards PC's resources — there is no point in AI if most of playerbase can't even play with it.

 

 

 Why in the world would a consumer and PC gaming fan ever espouse this? This sounds more along the line of a Developer vs Publisher board meeting notes.

 

 Of course we should push for the advancement of AI and of course we want tough, challenging AI. Notice I say AI not Aimbot. Id say quite a few older PC gamers were pretty excited when Swat 3/4 , Rainbow 6, No One Live Forever, Goldeneye (nintendo) came along as this seemed to bring about a whole new level of interesting AI. Now some 20 years later, its pretty shocking how little advancement has been made since then -save a few titles and id include Arma in that list. Multiplayer is far cheaper and less bug prone than any AI and that is the exact trend we are witnessing which is tragic for the SP'er.

  • Like 1

Share this post


Link to post
Share on other sites

Realistic Ai that acts like other players and are not super head shots bots and not scripted 100% to do things. If we like playing with human player then playing with realistic human like AI is the way to go. And you dont need a super computer to run it either. Again games are heading down this road of making complex challenging and fun AI.

Share this post


Link to post
Share on other sites

Why in the world would a consumer and PC gaming fan ever espouse this? This sounds more along the line of a Developer vs Publisher board meeting notes.

Whereas I'm actually in agreement with Semiconductor's stance, though I'll go with "too autonomous" as the biggest problem.

Of course we should push for the advancement of AI and of course we want tough, challenging AI.

Correction, the goal should be AI that can be scaled (I'm looking at you, AI settings) to feel challenging to players.

Notice I say AI not Aimbot. Id say quite a few older PC gamers were pretty excited when Swat 3/4 , Rainbow 6, No One Live Forever, Goldeneye (nintendo) came along as this seemed to bring about a whole new level of interesting AI. Now some 20 years later, its pretty shocking how little advancement has been made since then -save a few titles and id include Arma in that list. Multiplayer is far cheaper and less bug prone than any AI and that is the exact trend we are witnessing which is tragic for the SP'er.

The second sentence is probably because developers and publishers realized the third sentence.

Share this post


Link to post
Share on other sites

Why in the world would a consumer and PC gaming fan ever espouse this? This sounds more along the line of a Developer vs Publisher board meeting notes.

Thing is, our favorite computer games are products of meetings between developers and publishers, they aren't simply materialized underneath project lead Christmas tree. :D So we have to consider the harsh reality before making requests, otherwise they might happen to be unreasonable or infeasible.

 

Of course we should push for the advancement of AI and of course we want tough, challenging AI. Notice I say AI not Aimbot.

As a matter of fact there is a game (albeit with simple rules and world) with tough, challenging AI already — chess. There is no need to buy Deep Blue to compete against it and basic version is free, just navigate "All Programs > Games > Chess Titans". Difficulty 10 AI have won against me in less than 5 minutes without cheating, shooting through grass and spotting me from the other side of the map. It was a completely fair victory for AI but there's a problem — it wasn't entertaining for me to lose, especially so quickly. Result: none of us were entertained and one of us was unloaded from memory. Thankfully, there is an ability to make it less advanced but more entertaining, just like chortles have said, so the game still remains playable even for total amateur like me.

 

The point is: advanched (most effective in achieving its own goals) AI is not really suitable for a game. General-purpose AI and game AI are like a commuter train and rollercoaster train. Former should deliver maximum amount of people from point A to point B with maximum speed, comfort and with minimum expenses, i.e. it should be cost-effective. Latter doesn't have to be fast, comfortable, cheap or easy in service within reasonable limits, it even doesn't have to deliver people to point B. Instead it must somehow force its passengers to feel emotions, preferably positive ones.

Share this post


Link to post
Share on other sites

 

 

advanched (most effective in achieving its own goals)

 

 

Not so important to argue here about words and meanings, but still for the sake of conversation clarity - no. Advanced doesn't mean that. "Advanced" may mean effective in achieving goals, creators set for the AI. So there may be advanced chess AI, if playing chess is its purpose, but also AI advanced in simulating human-like behaviours, means doing this job very well. Or "advanced" may indicate complexicity of its algorithms. If so, most effective combat AI doesn't need to be advanced at all - perfect aimbot is much less advanced in that meaning than more "human" AI. I think, when someone says, he would like to have more advanced AI in Arma, then he doesn't mean aimbot. 

 

One thing more - behaving like avatar controlled by a player is not the same as behaving, like human would in real world in given situation, because players' avatars often doesn't behave such way (in game they may do things, they wouldn't in RL). If I would like to have advanced human-like AI, it would be about advanced simulation of real world behavior, not human in-game behavior. 

 

 

 

otherwise they might happen to be unreasonable or infeasible.

 

If numerous enough, such requests have a power to make unreasonable reasonable, means profitable (many declared buyers can make business risk worthy of taking). To the reasonable extend of course. Absurdal requests will stay absurdal, no matter, how numerous. Still I think, potential customers should be bold and demand challenging things from businessmen, otherwise we'll stuck forever in the world of cheap, financially safe trash. Maybe AI in games AD 2015 would be more advanced/complex/better, if more people would loudly care about this? Let's be loud then about ambitious dreams to improve chances, they become true some day. 

  • Like 2

Share this post


Link to post
Share on other sites

I think, when someone says, he would like to have more advanced AI in Arma, then he doesn't mean aimbot.

Neither do I, I was talking about self-learning human-like AI as described by DrJamesWhite. Chess was an example of a AI that is advanced enough to win against player without any cheating whatsoever and that was mean to show that it isn't really fun to play against an AI so smart that it can defeat majority of players without breaking a sweat. In other words, no, it won't be really interesting to play against an AI with skill and experience of Spetsnaz, Seals, Delta Force, Commandosi, [   fill in the name of your favorite SF unit from Wargame   ] combined, especially if they will act autonomously, following the commands of their algorithm and won't be affected by skill settings due to their self-learning nature.

 

One thing more - behaving like avatar controlled by a player is not the same as behaving, like human would in real world in given situation, because players' avatars often doesn't behave such way (in game they may do things, they wouldn't in RL). If I would like to have advanced human-like AI, it would be about advanced simulation of real world behavior, not human in-game behavior.

Sure, even current Arma AI tries to behave like real-world soldiers/squads. An AI who learned how to fight on a public server would be a human kind worst nightmare. Imagine it: an insults towards your family members that are original and in fact so witty that you can't even think about an adequate comeback! :(

 

Still I think, potential customers should be bold and demand challenging things from businessmen, otherwise we'll stuck forever in the world of cheap, financially safe trash. Maybe AI in games AD 2015 would be more advanced/complex/better, if more people would loudly care about this? Let's be loud then about ambitious dreams to improve chances, they become true some day. 

I'm not disagreeing with it. But I think that an extremely challenging, unpredictable and probably expensive AI is just as bad as dumb, predictable and very cheap AI as far as computer game is concerned.

Share this post


Link to post
Share on other sites

 

Look how fluid this battle looks. I'd like an analysis from people better informed on AI than I. There are more AI vids on Youtube as well.

 

One aspect that I think will increase AI immersion in Arma is the animation. Soldiers still move like humanoid cyborgs. When they run there's no momentum to it, it's like a mech walking and ducking. No dynamic cover, still primitive ragdolls.

Those are some nice animations and transitions but damn that ai played out pretty bad. Enemy soldiers just standing next to enemies totally unresponsive or very late to act. Wonder why he places them so close together from spawn as that can spazz out ai as its just unnatural.

Share this post


Link to post
Share on other sites

 

 

 self-learning human-like AI 

 

Well, all depends, what it would learn. If shooting/reacting faster in time of practice, then this would be at best only a very, exactly, advanced and expensive way to "educate" an aimbot. In fact relatively easy to emulate such learning (meaning training) skills, doable even now with not so complex scripting. Dead end IMO and nothing so thrilling in that, unless this wouldn't be a cheap emulation, but actual human training process simulation - result similar, but mechanisms behind as fascinating as human brain. Not sure, if such thing still could be named a game.

 

But... If such AI of only moderate and/or fully tweakable combat skills could learn by recognizing player's tactics, he uses against AI and adapting, seeking the ways to counteract, finding weak spots im my tactics, improvising, maybe taking initiative and surprising... Fully autonomous and dynamic, planned and executed from A to Z by AI ambush effective against AIs or players without using any superhuman AI combat skills, just smart tactics - how it sounds? At least to optimize, even if only via semi trial and error method forcing also me to change/enhance/adapt my tricks against it, all on individual but also multi-AI cohesive force tactics levels, not like some tireless, flawless war machines, but like real thinking, but flawed human could perform - that I would name an extraordinary achievement alone worthy my money. Even if in the result such AI would eventually outsmart me most of the times. I would just watch how it kicks my ass time after time in admiration - quite opposite experience than watching, how AI kicks my ass by happy aimbotting, which IS frustrating.

 

IMO better experience than any usual, standard gaming entertainment, which sooner or later becomes in gamer's eye repetitive, flat, tiring and unrewarding, thuss less and less entertaining each new standard game with standard AI, he plays.

 

And if on the top of all that being human-like AND self learning and adaptive like human (but not a superman), I would get full control over all its aspects to activate/deactivate/enhance/reduce/modify - wow! Really nice dream. Possible? If so, no idea, how. 

Share this post


Link to post
Share on other sites

Thing is, our favorite computer games are products of meetings between developers and publishers, they aren't simply materialized underneath project lead Christmas tree. :D So we have to consider the harsh reality before making requests, otherwise they might happen to be unreasonable or infeasible.

All the more relevant when one considers how much the back-and-forth between developers and publisher has been nigh-explicitly admitted to be what drove Arma 3 development after the reboot following the winter-of-2012 collapse.

It was a completely fair victory for AI but there's a problem — it wasn't entertaining for me to lose, especially so quickly. Result: none of us were entertained and one of us was unloaded from memory. Thankfully, there is an ability to make it less advanced but more entertaining, just like chortles have said, so the game still remains playable even for total amateur like me.

Nailed my thinking.

 

@Pulstar: don't count on a fix, because the project lead's claiming that even the currently 'bad' animations are "reaching the limits of our animation tech (some would say we've breached them a while ago ;)). While we do still plan some animation work and tweaks, they are very costly due to the limitations we have to work around. The Enfusion engine seeks to take animations forward a lot, but it's not going to be used wholly in Arma 3."

Share this post


Link to post
Share on other sites

You could certainly put machine learning to some good use in AI decision making - so exactly the level (and up) Rydygier mentions. This is not about basic movement, spotting, aiming and shooting stuff. And it's also not only about group-AI (such as flanking), but also about inter-group communication and interplay (think guard waypoint).

 

Lot's of fancy machines could be trained (rnn, decision trees/forests, svn, hmm, ...), and for a start maybe rather offline(-ish) and not online, while actually playing the game. Training might be very costly (even incremental), but querying such little blackboxes isn't that expensive (well, there is the memory, still...). I can imagine a rather tiny, but very high level feature vector could already work reasonably well (getting clean and potent information in a computer game is rather easy...). But encoding answers/reactions, seems to be already way more difficult, especially in a generic/general way.

But sure, I could see mission designers train some AI for their missions in a bunch of sessions: setup a scenario, do your stuff, and repeat a bunch of times, while slightly altering the scenario (to model different stimuli and the reaction you'd like to encode/train). BIS could train stock AI behaviours and incrementally update them, maybe just based on some special server logs...

 

Well, the time will come. I expect to see indy games go first in this direction with games based on ML as core gimmick. For mainstream games it's probably still way too early: all those fancy little blackboxes are hard to understand, train and integrate (in a meaningful way/as you want) and debug ("debug"). That's quite some risk to put such a thing into your game (i.e. highly experimental/not recommended).

Share this post


Link to post
Share on other sites

You could certainly put machine learning to some good use in AI decision making - so exactly the level (and up) Rydygier mentions. This is not about basic movement, spotting, aiming and shooting stuff. And it's also not only about group-AI (such as flanking), but also about inter-group communication and interplay (think guard waypoint).

 

But could be. Machine learning could be employed to such low level decision processes too, actually it should be easier to start there, and only then increment upon with higher level features, be them handcrafted by the coder or latent and "discovered" from the training provided.

One issue is that the decision space and dimensionality is infinite. We have two possible directions: firstly, we let the AI train randomly and explore that infinite space by itself and it will most definitely diverge from human behaviour since it will figure its own features/decision criteria, secondly we control, as you suggest, the training (the context) we will provide the AI, in the hope that the converging happens around features which are human intelligible.

Another issue: AI inputs and outputs, their constraints are slightly different than that which the player deals with (data available and decision vectors respectively). We can't exactly evaluate one to one and gauge the balance between them, unless we impose some evenness. Take enemy identification for instance, the AI is provided an hardcoded threshold of identification based on a set number of variables, players rely on visual perception and a contextually tied, relative importance of on screen information. Take the shooting decision on the other hand, here the AI and the player similarly follow a disparate pattern.

Third issue: how to evaluate these gaps if not via the end result? No matter how close we make inputs and outputs similar this will never be sufficient since only a very small difference in inputs will have catastrophically different end results (given current machine learning solutions). Consequently this will require a broad evaluation of competing behaviours. We must use the human behaviour as a "ground truth" instead of the abstract solution figured by the trained AI (despite it likely being more effective if given the training time), in order to serve the players an enjoyable experience.

Fourth issue: Arma hides plenty of AI decision processes deep down. To control the process above the behaviour evaluation cannot proceed if there are intervening variables influencing the end result there being no means to observe. The only alternative left is to totaly nullify current AI and build it up from scratch, task which will involve coding the AI and the concurrent systems which are hip-tied to it (ie. animation, communication, etc.)

 

Summary:

 

- Behaviour evaluation

- Data collection

- Magic

- Balanced AI

Share this post


Link to post
Share on other sites

But could be. Machine learning could be employed to such low level decision processes too, actually it should be easier to start there

Maybe. :)

Then again, ArmA is no Super Mario (I'm sure you've seen plenty of AI playing 2d-sidescrollers). The environment can't be learned, since it's not static (or deterministic), but highly dynamic. We can't just use a "simple" fitness function optimizing some score (like distance reached in a 2d-level).

 

firstly, we let the AI train randomly and explore that infinite space by itself

The question is what the AI is supposed to be optimizing/maximizing? Combat effectiveness? Some combined score? Or should some error (deviation to "human behaviour") be minimized?

We probably do not want to end up with AI that finds new and effective ways of combat, if it looks silly and wrong (e.g. exploiting glitches/bugs). The goal, afterall, is to model soldiers (and real, or believable, behaviour/tactics).

But you make good points: there are indeed many low-level decisions that could be driven by ML: spotting decision, shooting decision, and such stuff. And maybe for such small problems (or subsystems) ML could be employed in a much less restricted fashion (i.e. unsupervised). I just doubt that ML is the right tool to model AI on a holistic level, since again; we do not want to breed a new species, but rather mimick something we already know.

It's not about the abstraction level where ML can be introduced (first). It can be low-level or high-level - what matters is that it's a confined subsystem where we can formulate reasonable goals to be optimized.

 

We must use the human behaviour as a "ground truth" instead of the abstract solution figured by the trained AI (despite it likely being more effective if given the training time), in order to serve the players an enjoyable experience.

Exactly. But there is no single "human behaviour". There are many different roles that could be trained. Training data thus should be plug'n'play (per unit or group). And there needs to be a streamlined workflow to let mission designers easily train new roles as they see fit (besides having a set of pretrained roles by BIS). Imagine if missions would (partially) come with their customly trained AI? Bananas! :D

  • Like 1

Share this post


Link to post
Share on other sites

Please sign in to comment

You will be able to leave a comment after signing in



Sign In Now

×