Legitimate Question about AI

[LoneCrow66 10]

Is anyone on their team actually working on it? Or is all right now bugfixes and working on the campaign? I'm talking mainly about the enemy AI and its behavior when under fire. Someone else is making a mod that should really be put into the game about how ai responds to sound long range. If it has gunfire coming in it should try and run to the nearest cover perpendicular or towards the direction (if cover is close) the fire is coming from.

If mortars and explosions are taking place around the enemy they should be running for cover even if nobody is hit.

I find as a sniper I can literally kill 50 guys within minutes because they all just run around in circles and lay down. Realistically after the first couple shots it should be difficult to find anyone after that. They all should be behind rocks and in buildings.

Are they actually going to improve this aspect of the game? Is it being worked on? Or are there bigger priorities for them right now?

If not is this something mods can actually truly improve? Do they have enough access to the code to improve the default behavior of enemy? It would be better if BI did it so not everyone had to use certain mods to connect.

[mistyronin 1181]

Development Branch Captain's AI Log :rolleyes:

[Tonci87 163]

See here

http://forums.bistudio.com/showthread.php?159710-Development-Branch-AI-Discussion

and here

http://forums.bistudio.com/showthread.php?159711-Development-Branch-Captain-s-AI-Log

Of course they are still working on it.

[Pek 15]

The AI is already inproved. Just take a look at A2's AI...

But I agree, some tweaking must be done, but I have full confidence BI or a community member will fix it :)

Edited October 8, 2013 by Pek

[LoneCrow66 10]

Thanks I see there was an update today. that is good. I hadn't seen any in a while that is why I asked.

It is good to see they are aware of some of these issues and are going to put some time and work into it.

[iceman77 18]

I haven't noticed any meaningful difference between the AI of OFP and the A3 AI. All of the important things are still broken. Ai still can't drive down a straight road without running into shit. AI still stand / loiter / lay around when i'm killing the living shit out of their buddies, that are dying right beside them. AI still go from one extreme to the other, either super commando leet god mode, or super retard mode. AI Still have problems dealing with concealment, they love to shoot well concealed players they can't even see.

[LoneCrow66 10]

Quote Originally Posted by AI Configuration Analysis

PROBLEM

AI configuration settings aren't very clear to players

AI 'skill' is quite abstract

Players can unwittingly 'break' the game by setting values that are too low

e.g. setting AI skill to low

Makes it difficult to reliably test behaviour of AI in playable content (moving target of AI skill)

Mission designers can't easily modify useful parameters of AI

e.g. Accuracy

Players can override what designers have selected (related to above point)

Mods are used to unlock additional parameters, but these are not adequately documented and tested in engine

DESIGN

Players have simpler, more predictable options

Some choices removed from game configuration (skill)

Only things that are simple to understand and test

e.g AI: HARD / DEFAULT / LIMITED

Value is an abstraction of limited number of predictable inputs (e.g. AI Accuracy and Turning Speed)

HARD: Default (current state) Accuracy; Default (current State) Turning Speed

DEFAULT: Lower-than-default Accuracy; Default Turning Speed

LIMITED: Lower-than-default Accuracy; Lower-than-default Turning Speed

linked to difficulty presets (recruit would be "limited", etc)

no slider nonsense, just combo-box options

Do players really need to be able to set ability of both friendly and enemy AI?

Why not simply abstract these settings into one value?

Curate the friendly/enemy settings to what we think is 'best practise' for each default

More advanced AI settings should primarily be the mission (official/community) designers choice

Configurable via the Editor

AI Skill

clearly validate and document what existing configuration options - unlock potential of parameters

Split into more options, if useful

----------

Ok so they are mentioning about the "turning speed" of enemies. I've seen this reference before where the turning speed affects how they can zero in on you. In reality sound at a distance is more of a determining factor for zeroing in on an enemy and tracers.

Almost all people would automatically head for cover and then try and peek their heads out and try and zero in on where it is coming from. Especially if it is only a few people. If it is coming from multiple angles it would most definitely head for cover. Some training in the military would tell you that the soldier would "double tap" in the direction of the incoming fire. So you might be able to get direct fire back at you instantly but it would be extremely inaccurate. It would be impossible for AI to turn and shoot at you that fast or a person without a scope.

This leads me to how I think soldiers should be in several "modes".

This is all for enemy AI in cooperative missions. Friendly AI is a whole other thing. Actually the friendly AI is not too bad.

Defensive Mode

Sub Mode A)Relaxed - There has been no gunfire or action in a long time. Routine patrols and defense situation. This takes a long time to get into readiness to zero in on who and what is shooting at them. This is not a state of readiness and will take them a long time to get organized and return fire.

Sub Mode B)Aware/Cautious - Expecting an attack and actively patrolling and scanning for targets. This it still limited by their own eyes and or any commanders that may be scanning using vehicles / binocs. I could see someone at 200M but doesn't necessarily mean they are an enemy. There is no way to know unless they start shooting. People just don't start shooting on sight. This is a more Cautious mode.

Sub Mode C) Under Attack - more prone to shoot at anything, run for cover immediately and try to narrow down where the fire is coming from. There is a quicker time to zero in on where the fire is coming from. They are more prone to put down suppressive fire to buy time to locate the targets. This effect is triggered by gunfire, explosions, detonations. Or soldiers being spotted within 100M advancing at rapid paces. If a vehicle is seen like an APC or Slammer it can go into "under attack" mode at longer range. Just remember just because the North Koreans and South Koreans see each other all the time doesn't mean they actually shoot. Sometimes you'll have a patrol going by and if this is a town they may not want to engage and let them dry past the town.

Offensive Mode (For AI Attacking and counterattacking us)

Sub Mode A) Casual Movement Mode -Traveling long distances like in a convoy not expecting action. This should take more time to zero in and find out what is going on. Like the ultimate ambush is to catch them in this state.

Sub Mode B) Imminent Danger Mode - Approaching target destination the AI will constantly scan for targets and the zeroing time will be significantly reduced. ie "heightened state of awareness". They are advancing with the knowledge there is enemy ahead - they will fire very fast and zero in very fast. They will advance to cover very fast and toward their objective.

These are the ideas I have, but some of these "modes" they go into will have certain profiles on how to act, and these are triggered by gunfire, sounds in the distance, overflight of enemy vehicles etc..

[instagoat 133]

You wouldn't believe how difficult it is to make an AI for a game like Arma. Every other game either has no meaningful AI, or tailors its environment exactly to the demands of the gameplay and capability of the AI. Arma does the opposite. It gives the AI a totally random, non-optimized environment, often community made, and has to work inside of that all the time, every time.

That they are able to judge corners 50% of the time is pretty astonishing to me. It is super annoying that they get stuck, but from a scientific point of view, they're pretty fantastic.

AI tech today even in the most advanced applications dedicated to AI science are barely above the cognitive level of cockroaches. We have seen advances, and we will continue to see them, but I'd not expect miracles.

What I'd -love- for them to do is work indoors, though. Movement indoors can be extremely regulated, which is why most shooters take place indoors: you can control the environment very exactly. I hope we will see advances in that direction, right now they don't know how to use buildings at all.

[jiltedjock 10]

You wouldn't believe how difficult it is to make an AI for a game like Arma. Every other game either has no meaningful AI, or tailors its environment exactly to the demands of the gameplay and capability of the AI. Arma does the opposite. It gives the AI a totally random, non-optimized environment, often community made, and has to work inside of that all the time, every time.

That they are able to judge corners 50% of the time is pretty astonishing to me. It is super annoying that they get stuck, but from a scientific point of view, they're pretty fantastic.

AI tech today even in the most advanced applications dedicated to AI science are barely above the cognitive level of cockroaches. We have seen advances, and we will continue to see them, but I'd not expect miracles.

What I'd -love- for them to do is work indoors, though. Movement indoors can be extremely regulated, which is why most shooters take place indoors: you can control the environment very exactly. I hope we will see advances in that direction, right now they don't know how to use buildings at all.

Agree re indoors. The AI has no concept of indoors and there is no tactical usage of buildings by the AI. Imagine how different the game would be if you approached every village wondering if there was enemy at the upper windows.

Of course there are mods that will make the AI behave with different tactics including building usage - albeit the building usage is all "look" without purpose as fundamentally a building means nothing to them. But compared to vanilla, Arma 2 for example with GL4 and TPW suppression is a totally different game.

And this is the frustration - modders have been writing scripts to do this sort of thing for many years. Why do BI not implement any of them, or at least draw some inspiration from them? How long was the danger FSM broken for in Arma 2?

To me it seems like BI have some deep rooted conflict between the idea of a sandbox, and the idea of AI with more of a "purpose" to it. AI which might ignore mission makers waypoints for example eg if you are using GL4.

It's the only explanation I can come up with for the neglect - it's deliberate, not because it's "difficult". Snakeman et al have shown there is nothing difficult about scripting the AI to hear shots and react, see smoke and react, see/hear explosions and react, be suppressed, garrison buildings, split squads to flank, call for support, simulate dead, call artillery support, call air support, use smoke for cover, use flares, rearm, detect bodies.

I hope that once the endless fiddling with prone turning etc is finished in Arma3, that a clear mission purpose and framework for AI with more strategy baked into it will be developed, but I don't think it will happen. The desire simply isn't there.

I sometimes wonder how many of the devs have played Arma 2 with something like GL4, just to see how different it is to vanilla to plonk down a few groups in the editor with no waypoints and watch them get on with it, it plays out differently every time.

Edited October 8, 2013 by jiltedjock

[LoneCrow66 10]

Maybe the lead in charge of AI needs to go? Move to another department? Let some fresher minds take over?

[jibemorel 10]

Maybe the lead in charge of AI needs to go? Move to another department? Let some fresher minds take over?

Or maybe not, apparently what you are suggesting is what happened the last 10 years, people replacing each other and leaving without documenting any of their work. See the mess...

Try harder.

[instagoat 133]

Maybe the lead in charge of AI needs to go? Move to another department? Let some fresher minds take over?

Doubtful that'd help. The problem is the inherent limitations of computational language, as far as I understand. There is a technological cap. However, I would be all for taking up a special testing and design programme at some point to completely build a new AI. Maybe find a european University that does AI research and work together with them, open up RV to education and research to try and build an in-world AI. Maybe that could also open new business avenues (such as using RV4 as a simulator for traffic education, as an example.)

[LoneCrow66 10]

Remember I'm talking from the perspective of enemy NPC AI. I mean the pathfinding seems well done, It is just when the AI decides to do something and WHAT it should do that is the problem. That can be simplified like I posted into probably 6 or 7 root scenarios that could be controlled by an overall command AI that will be able to flip them into different modes. (Defense vs Offense) etc. I think trying to get them all to think on their own is difficult. The only thing soldiers should do on their own really is find cover, fire in the direction of anyone shooting at them. The longer they can stay behind cover and have a good visual ID on the target (without seeing through trees etc) the faster they can zero in on the stationary targets. At a distance. If an AI detects something within a bubble of 50M or less visually it should flip into a higher mode that requires more processing. In other words this is sort of like the old bubble in Falcon 4.0. The further away the NPC's are away the more they act as a group, the closer that a player gets to an NPC the more individualized the AI becomes. I mean unless a commander calls over the radio "retreat" all the enemy ai could run back to a rally point unless suppressed.

I mean really this can be broken down into simple AI schemes/modes attack, defend, retreat. The commanders/squad leaders decide when the individuals flip major modes, and the individuals decide when to flip submodes. They also should all not aimlessly run around individually but to try and spread out and stay as a group behind cover. But stay with their squads. This makes a lot of "decision points" disappear, less processing if you only have 1 person commanding many while the soldiers are operating at a simple ai level like cover and fire. They are in low processing mode until you are in that close quarters battle with that other squad.

Anyone remember that Falcon 4.0 bubble?

---------- Post added at 07:03 PM ---------- Previous post was at 07:00 PM ----------

Hah this brings back memories! http://www.combatsim.com/review.php?id=451

Kevin gave me some technical insight as to how the AI runs that I found very interesting. He realized early on that a single level AI would never produce a "real world" feel, because in the real world, decisions in the military are made at many different levels. In his attempt to understand and then reproduce military (and human) thinking, he made a multi-tiered AI system. You can imagine a room full of 5-star Generals with a big "battle board" making "big picture" decisions, such as "We need to take Seoul". He calls these decisions "hi level".

Clearly, there are many hi level decisions being made simultaneously in a Falcon 4 campaign, following a plan of attack with an "ultimate goal" in mind. Each campaign has its "ultimate goal" and they are pretty well spelled out for you in the little movies that introduce each campaign. In fact, the concept of a "mini-campaign" in the TE editor also has its "ultimate goals". They are set up in the "victory conditions" screen. As you read on, you will begin to realize that the TE editor utilizyes man, but not all, of the multi-tiered AI subsystems of the campaign. Those parts that it does not use, the human mind is left to replace.

This is where things get very dynamic and very interesting. I once read on the Newsgroups one guy griping that the campaign couldn't be dynamic, because every time he starts the campaign, he gets the same set of missions. He couldn't be more wrong. A chess game between 2 humans is as dynamic as a campaign can get. It NEVER plays out the same way twice. It twists and winds around every move and countermove. And yet it still needs a standard starting point.

Every chess game ever played started with the exact initial setup. The 3 campaigns currently available for Falcon 4 are analogous to 3 different chessboard setups at the start of the game. And the 3 different ways to win (in one game, you get the king, for example. In another, you eliminate all pawns and rooks.) Even real wars start at some balance of power and position of troops. No matter how many times you look at the initial pieces on a chessboard, the game takes on a life of it's own once the initial moves are made. This campaign is dynamic in exactly the way its most vocal fans wanted it.

Once these "hi level" AI decisions are made, they are passed down to the mid-level AI. This is kind of like the 1 star Generals in the field meeting with their battalion commanders and airforce leaders. "OK boys, in order to take Seoul, we need to create a corridor of entry for our planes and ground troops. Let's start by knocking out their radar and softening up the defenses in the area." Now the mid-level AI actually starts making decisions about what resources to utilize based on those available.

And remember, as the war moves on, resources change. A lot of guys even make their first move or two in chess the same way, but if they suddenly get their bishop whacked, well then. Things start to change, eh? You start re-evaluating your resources and possibly allocate more towards defensive maneuvering. Etc. Dynamic.

The mid-level AI in F4 generates the actual packages and flights based on modern doctrine in order to take the objective, but as the friendly AI responds (and you tip the balance of things by making an AI flight that otherwise would have been a failure now successful), things change. Resource allocation, whether to be defensive or offensive with remaining assets in a particular area, and the actual package/flight/ground war activities are assembled and assigned. No scripted missions here boys and girls. This is the essence of a dynamic campai

The third or "low level" of the AI is analogous to the actual grunts and flyboys in the field reacting to the plan and their environment. Imagine the pilots in the briefing rooms, deciding how to lay out their waypoints for the flight they were assigned, and which actual targets (seen the target lists lately in TE?) to bomb at particular sights. What munitions to use. How to avoid detection based on current Intel. Etc. Also at this level, are the individual AI routines. Sure, they are brave, red blooded American heroes, but hey ! Everyone has his limits.

[instagoat 133]

They don't have playbook moves right now, and their behaviour is rather untransparent.

Also, if you want to make an AI, you need to turn your abstractions into logical arguments in the programming language, and FSMs. If you put yourself in a situation, you can immediately take in the situation, understand the bounds and fell a sound decision. Your perception is multi-channel, instantaneous and synchronized, filtered in your brain and turned into a chronological pattern of perception. Game AI is inherently limited in this regard because you do not have senses. You need to build -everything- they do. You are also limited by the nature of computers, which cannot do mulitple synchronous operations but can only emulate those synchronous operations by doing so many cycles in such a short time that they -appear- simultaneous. They are in fact done step by step.

If you put an AI in a situation, you look at it and ask yourself after felling a decision WHY you came to fell that decision. Make a list of the things you had to go through to decide not to cross that street and instead stick by that wall to take cover by that wreck. Now realize that you have to go through that cycle for every AI on the map, and take in special considerations like armament, nearby friendlies, nearby enemies, vehicles, terrain, buildings, quality of cover and concealment not only NOW but a certain amount of time ahead (impossible by the way in a nondeterministic scenario as far as I understand it, at least to any meaningful time.). Then you need to teach them to not just individually fell decisions, but fell decisions for each other (leaders) and work in coordination to execute their objectives (buddies).

Additionally you need to enable the AI to calculate how quick they can move over which terrain, under which circumstance, when to move and when to stay still, etc. In a tunnel shooter, this is difficult to achieve. There are unbelievable amounts of criticisms out there for the AI's of such great games as Half Life 2, and many games that in fact do have shitty AI -despite- being tunnels that could be hand-crafted to the abilities of the AI.

The AI in Arma does not have that advantage. They do not have tailor made terrains like the AI in FEAR had, instead they are thrown into the same sandbox as the player, and get by on equal terms. Pretty much no game I know does this. OFP:DR cut corners in the ballistics/spotting/firing departent, games like IL-2 made the AI ignore limitations of physics to make them competetive and still other games completely do away with AI because it is impossible to do. That we've gotten this far is brilliant already in my book. And, now that they seem to have figured things out, it will likely only get better.

I like to think that Arma's AI, as far as technology goes, is world class. Hopefully some AI experts will take a look at it at some point and confirm my belief. Meanwhile, I hope we can get them to work indoors at some point.

[timmo1982 10]

I think trying to get them all to think on their own is difficult.

When you start thinking really in depth about the problem though, it becomes vastly more complex and you realise that they do, indeed, need to think for themselves- Simplifying AI thinking by treating them as a squad entity makes sense on the surface but I don't think that can be applied to the actual scenario they face (unless of course no-one is looking at them). Say, for example, a squad needs to move across a field. In this simple case the simplified 'squad think' works fine....but what happens when one soldier in the squad encounters a tree: He cannot just walk through it in order to maintain the simplified/unified movement of the squad, therefore that soldier, every soldier, needs to think for himself to avoid the object.

If a Squad encounters an enemy squad: The whole squad can either target one soldier of the enemy squad or each solider can target another soldier in the enemy squad. If it's the former, then it's entirely possible that only one soldier in the squad can see/be seen by the whole of the other squad (i.e. Dumb AI). If it

s the latter each solider needs to think for himself: Can I see the target? Can the target see me? Do I need to move so I can see the target or move so the target cannot see me? If I need to move, how do I move without exposing myself and while avoiding the (many) objects. These problems increase exponentially at the larger ranges of conflict that ARMA is known for.

It's a bit like a centipede- A squad (centipede) can't walk without all the legs working correctly.

[maddogx 13]

Development Branch Captain's AI Log :rolleyes:

That thread should suffice for discussion of AI developments.

Closing.