AI disaster + tanks

[Second 0]

I don't understand this fuzz over improving AI to level that it can be called smart.

Because ArmA's AI heavily relies on waypoints and instructions given by mission designer, so AI can't perform very much better even if it's given "intelligence of human". Waypoints and mission designing has the biggest responsibility. If AI hasn't been given good instructions then it's performace is bad. And AI in armA can't be very reactive to surroundings because of the fact that it has to follow it's waypoint. ArmA has big squadsizes which somewhat makes things easier for bigger units, but it still it doesn't make some smart tactics/technigues very easy, like tanks moving on road and infantry securing the forest that surrounds the road. Or syncronicing tasks for units in better manner. War-things needs lots of coordination which lacks in ArmA and result in ArmA is that cooperation of different groups doesn't work very well if at all

For Game 2 (or modded in ArmA) best thing BIS would do would be to create alternative way which can be used also (waypoints still exists for cutscenes and such "unnimportant" stuff). It would let AI to work more in dynamic manner with set of instructions given to it.

-Unit organization (it's subunits, superunits etc...), from where to request reserves, air, and firesupport. And who will do what:"my squad suppresses, your squads charges and their squad can take a break" and so forth.

-AI would have given it's objective(s) which it has to reach. Objective is given to unit (AI) who commands the unit (batallion etc) that needs to reach that objective. Subunits then follows.

-what it's expected to do in it's way to objective (attack, patrol, defend etc..)

-in what manner it tries to behave in it's way to objective and in objective (in aggressive or cautious manner. choosing evading, quickest or shortes route etc...)

by these kind instructions let AI to do it own judgementions (from something like tactical pool) acording those instructions.

Hard for my kind complete noob to say it's it doable in FPS (i've made one bit like it for ArmA but it's severly limited), but in strategy games it's in use (Airborne Assault-series anyone? the system is unbelivable flex and efficent) up to corps level i've seen it.

EDIT: Sorry about my bad english and other strange conclusions. Had to take care of my kid in mean time as i was writing this.

[ShrubMiK 0]

Agreed a hierarchical decision-making structure seems like the best approach - I Was thinking along similar lines when looking at the sort of things that typically went wrong with the decision making process of individual units in Rome:Total War, amyn units would do stupid things because they were making their own decisions without regard to what ther neighbouring units were doing.

It will be interesting to see what comes in Game2.

Quote:

"Unfortunately, software programming has reached its limits; its impossible to design software that can emulate sentient life.

IMO, the only solution to our AI problems will be AI cards (similar to graphic or physics cards); emulating a human brain."

That's an interesting statement...IMO in theory anything you can achieve in hardware can be simulated in software, assuming the computer hardware the software is running on is powerful enough.

Conversely, if you can't work out a good approach to emulating intelligent reasoning in software, it's going to be hard to build dedicated hardware to achieve the desired effect (except perhaps in a very limited, well-understood situation)

.

There are plenty of interesting approaches to making software learn and appear intelligent...none of which are magic bullets and I haven't ever come across anything that really made me go "wow!" yet, but who knows how they will develop in future.

[Hostilian 11]

I thought that the AI constantly using RPG's on soldiers was a bug.. It didnt happen anywhere nearly as much in OFP (if at all).

It looked like AI were actually using their RPG/LAWs as sights..

Quote[/b] ]

1.08..

- Fixed: AI no longer uses RPG/LAW for looking at unknown enemies.

I'm not so sure that it has been fixed.

#C

[Easy_tiger 0]

my point is this: programming a convincing AI is the most difficult - if not impossible - part in every game and I doubt that we will see any significant improvement anytime soon.

Not really...

Its just a matter of time...

You would need like Millions of "what ifs" and "then do"

in the code. And especially the recognition code is verry

important to AI. Like "This is a Road", "use as" and "use to".

But like i said it takes A LOT of time to write these

codes...

oh come on thats just silly

Neural Nets or Finite State Machines are the only sensible way to model artificial intelligence, a huge block of coditional IF's, switches etc are inefficient,hard to maintain and would probably lead to an infinity problem.

Even when using the above proven methods it is very hard and time consuming to emulate "Intelligence". If you examine a real life action and break it down you will see hundreds of of very small decisions, definitions and variables which are dependent on other decisions and variables along with the situation at hand.

Though it may be possible to create a "Uber" intelligence for Arma i doubt it would be computionaly efficent enough to be used in the game, i dont think that there are many home PC's out there that could handle the computations anyway!

[Whargod 0]

And why not have soldiers use RPG's against soldiers? It's realistic!

Geist, I don't get what you mean by programming reaching it's limits? At the moment it's mainly hardware that's the problem. As for OOP and no conditionals, you will always have conditionals.

And as you said, we need better hardware, and the only way to interface and use this hardware is with better software. There is no such thing as a general purpose AI on a chip, not gonna happen. Most likely it's going to be a specialized CPU and someone is going to have to code for it from the ground-up. So software still plays a monumental role in all of it.

And as was pointed out, it is possible to create software that can emulate sentient life, we just can't do it yet.

[mr reality 0]

And why not have soldiers use RPG's against soldiers? It's realistic!

Yes it's realistic, i dont think anybodys denying that. It's just that the AT AI soldier will target infantry even though there is OPFOR armour nearby. They only carry 3 rounds and they generally shoot all three at the infantry target. This leaves them completely defensless against the armour. I would prefer if they just didn't target infantry.

[Whargod 0]

I would rather they prioritize, take the armor and then hit the infantry with rockets while the other soldiers suppress them. Everyone has a wish list I guess.

[mr reality 0]

The problem with ArmA AI is that they dont have self presevation. That's why you see the AT soldier run towards the tank that you've just ordered him to target. Even though there maybe opfor infantry units nearby, he'll still run towards the tank until he reaches the target/weapon effective range.

AT soldiers in ArmA= joke....

[NoRailgunner 0]

AT soldiers should fire their AT-rockets at Main Battle Tanks and Armoured Fighting Vehicles! Thats the main task for them.

Maybe BIS can implement at the next patch two warheads for AT units (HE/HEAT) and better AI....

[Second 0]

Like NoRailgunner and Mr Reality said their behaviour in fullfilling engage-order isn't very optimum in sense of self preserving. But i think that from viewpoint of getting the job done before whole squad gets killed by enemy vehicle, the current way of duing things is good. It relies on speed and to fact (or wishful thinking) that enemies wont notice that AT-guy is running towards them. They survive from task alive if vehicle is on it's own and it gets destroyed or enemy vehicles are in target rich enviroment where single AT-guy is no biggie.

I'm not sure is it possible to make it much better. Keeping low and in cover would be one possible thing to improve (but then it leads to other issues like: they fulfill their task very slowly)... About rest i'm not very sure. Quality of firing position has big effect in survival of AT-guy. Quality of path towards the target is another. Quality of path seems to be quite okay in ArmA, but guality of firing position (if we even can use that kind of term) isn't very good.

[Sand Flea 0]

ArmA is the only game I know where the AI have moments of extreme stupidity and moments of extreme intelligence. They can shoot down A-10's with the main gun of a T-72.

[goodguyswearblack9 0]

if ArmA's AI could move around and react in an 'intelligent' way we would have had robots in our households, on the battlefield and everywhere else for quite some time

my point is this: programming a convincing AI is the most difficult - if not impossible - part in every game and I doubt that we will see any significant improvement anytime soon.

I couldn't have said it better myself.

[M. D. Geist 0]

IMO in theory anything you can achieve in hardware can be simulated in software, assuming the computer hardware the software is running on is powerful enoug

Not quite. Something as complex as human behaviour cannot be concieved with mere programming. You simply cannot forecast every possible problem/question that might come up.

Quote[/b] ]Conversely, if you can't work out a good approach to emulating intelligent reasoning in software, it's going to be hard to build dedicated hardware to achieve the desired effect (except perhaps in a very limited, well-understood situation)

I'd would say its the other way around. Robot engineers are now moving from CPU assisted decision processing to pure hardware processing.

Quote[/b] ]There are plenty of interesting approaches to making software learn and appear intelligent...none of which are magic bullets and I haven't ever come across anything that really made me go "wow!" yet, but who knows how they will develop in future.

AFAIK, these programs are merely storing data in strings/integers/arrays and loading said data on demand. Imagine how many strings/integers/arrays you need to create/allocate in order to achieve any decent level of 'intelligence'. This, IMO, is a very dangerous approach due to the high possiblity of memory leaks (especially if you're using pointers). This danger can of course be eliminated with garbage collectors (i.e. managed code (ex. C# but who's programming AI in C#?). On the other hand, managed programming languages are slower than unmanaged. I'm not an AI guru, but I can see a lot of things that can put you in a world of hurt.

Geist, I don't get what you mean by programming reaching it's limits? At the moment it's mainly hardware that's the problem.

Half-Life 2 has close to 150,000 lines of AI code; and most of the AI is actually scripted on level-basis. You can multiply the lines by ten for non-scripted, dynamic AI, and still not get anywhere near the level of sentient life.

Hardware is not dependent on software. A simple robot consisting of a few transistors already equals 1,000,000 lines of code. In fact, its even superior. Not only can it make decisions in a fraction of CPU processing time, it can also learn and "remember" an incredible amount of decisions.

Quote[/b] ]As for OOP and no conditionals, you will always have conditionals.

Conditional statements and conditional programming (procedural/structured programming) are two completely different things. Before OOP came along you only had the conditional/structural approach. Nowdays you won't find it anywhere but at the lowest level.

Quote[/b] ]And as you said, we need better hardware, and the only way to interface and use this hardware is with better software.

No, that's completely false.

Quote[/b] ]There is no such thing as a general purpose AI on a chip, not gonna happen. Most likely it's going to be a specialized CPU and someone is going to have to code for it from the ground-up. So software still plays a monumental role in all of it.

Again, completely false.

http://www.robotroom.com/RBFB.html

Quote[/b] ]And as was pointed out, it is possible to create software that can emulate sentient life

No its not. Software engineering is too limited. Simple transistors are already superior.