AI bug

[Predator555 0]

good video and yes, the scientific approach is always the best one, im glad you said that

but just to clarify one thing, the AI does not appear to turn its head by more than about 60 degrees to the left or the right, but from that video it appears you have to turn your head well over 90 degrees to the left to see me (looks about 125 degrees or so?), something which the AI does not do on its patrols

am i mistaken or were you simply proving that i was visible from my position, if that is the case then it does not explain why i get spotted especially so late in the patrol.  

EDIT - perhaps a better test would be to simulate the AI patrol by actually doing it and scanning in the directions the AI do and then see if you can see me

i did it and found it was near impossible to spot the hidden soldier.  Im not familiar with how you are able to set one soldier in a position and then switch to another so perhaps you would be better doing it and recording it, if you have the time of course, i could only put a static soldier using the editor to a fairly close position to the one i took up in my video.

[walker 0]

Hi Predator555

In terms of the control of the unit.

Method

Put 3 soldiers on the map

Set all units in special to "None" (so that they are not in formation)

Place the command camera unit in the gated bay. Make this unit player.

Place the subject unit near the gate and alcove. Make this unit playable.

Group the command camera unit and the subject unit

The subject unit needs to be under the command of the command camera unit.

Place the target unit ungrouped at the approximate AI position. Make this unit playable.

Preview the mission.

Team Switch default [T] key to the subject

Move the subject to the aproximate position

Take a knee default [x] key

Team Switch back to the command camera unit (the subject AI unit will STAND UP this is the problem Predator555 describes above.)

Select the subject unit default [F2] key

Order the subject unit to crouch default radio commands [7] [7]

Team Switch back to the subject unit

Readjust the subject unit's position to get it correct

Team Switch back to the command camera unit (the subject AI unit will REMAIN CROUCHED/KNELT this solves the problem Predator555 describes above.)

Team Switch to target unit

Perform observations

Kind Regards walker

[Predator555 0]

thanks for the method Walker, ill look at it tomorrow, any thoughts on the rest of my post?

[KeyCat 131]

Here is another approach...

I took Predator555's mission and added a script that gives you an eyecam and let you see thru the patrolling AI's eyes.

I haven't verified that the cameras FOV is exactly the same as the AI's actual FOV, really don't know how to do that?

You start the eyecam by pressing 0-0-1 in game (or via the radio on the map). I made two versions of the same mission, one with a single AI patrolling and one with a group of AI's patrolling.

You can download both test missions here:

http://keycat.no-ip.com/files/ArmA_114_Peripheral_vision.zip

By running them a few times it revealed that the squad leader clearly spots you in the door when he turns his head plus a couple of other interesting things...

1. The "spasm effect" is clearly noticable in this view but only occurs on the squad leader when he his leading a group of soldiers. If you try the mission where he is alone he pan his head nice and smooth and have no spasm.

After giving this some thoughts I actually think this  "spasm effect" may be by design. Having the squad leader scanning the area faster and wider "simulates" that a group of soldiers usually sees more than a single soldier and that way the CPU only needs to calculate one AI instead of every AI in the group (or the code checks/updates the SL at a faster rate and the other much slower) and thus saves resources .

The above probably explains why you are only detected when the group patrols and not by a single soldier.

2. When used in combination with my previous posted AI_No_HeadTurn test addon (where I set minHeadTurnAI = 0 and maxHeadTurnAI = 0) it seems that I only limited the visual headmovement on the 3D model and not the actuall "AI eye scanning", the head on the model doesn't move but the AI still scans the area and thats why they where as good as without the test addon.

Anyone knows if it's possible to tweak the actual AI eye scanning parameters?

Since we don't know how it really works it's pure speculation from my part and I could be completely wrong. Give it a try and post your thoughts... It's easy to use the script for other scenarios/missions as well.

Quote[/b] ]

I congratulate Predator555 for setting a proper tone in this thread and presenting a proper scientific case with video and example mission rather than just a rant as some in other threads have.

I second that! Some people claims that BIS don't listen to feedback etc. but I know for fact that if you come up with solid information (RPT-files, dump files, videos, screen shots etc.) and even better, can show the issue/problem in a reproducable way (preferable a simple mission in the editor) BIS more than likely gonna look into it.

/KC

[-ASA-ODEN 0]

It's class ViewPilot.

Setting minAngleX etc. to 5 deg the patrol strolls past me.

zip archive with odenAI.pbo and edited keyCat mission:

http://web.comhem.se/~u83610294/odenAI.zip

<table border="0" align="center" width="95%" cellpadding="0" cellspacing="0"><tr><td>Code Sample </td></tr><tr><td id="CODE">

class CfgVehicles

{

class SoldierEB;

class odenAI_SoldierEB : SoldierEB

{

vehicleClass=VehicleClassOdenAI;

displayName = "Handsome Odan";

/* Headlimit is not the issue

class HeadLimits : HeadLimits

{

initAngleX = 5;

minAngleX = -5;

maxAngleX = 5;

initAngleY = 0;

minAngleY = -90;

maxAngleY = 90;

};

*/

class ViewPilot : ViewPilot

{

initFov = 0.95;

minFov = 0.4;

maxFov = 1.1;

initAngleX = 8;

minAngleX = -5;

maxAngleX = 5;

initAngleY = 0;

minAngleY = -0;

maxAngleY = 0;

};

};

};

[walker 0]

Hi all

KeyCat's experiment is far more refined than mine.

My Experiment established the baseline of what a player can see.

KeyCat's experiment Established the answer to Predator555's next questions "Ah but what can the AI Target see? And did it see the subject?"

He then went on to speculate as to why the group leader looks around so much. And concludes that CPU reduction is a good explanation.

I will add to the above explanation that in reality point man is supposed to be looking around a lot. His life and that of everyone else in the column depends on it. So there is an additional simulation validity point there (needs a military person as subject matter expert to weigh in on this being correct SOP though)

And finally [ASA]ODEN comes up with the theoretical background as to why it happens and discovers the factors that can altered to change it.

This is how it should be done from this we discover things and take the the simulation forward. From these things we have a sound basis on which to make new mods.

This thread is how bug tracking and should be done.

Maybe the moddies can pin this thread up as with some additions to the title.

AI bug (how you test a bug)

In the end not an AI bug

Kind regard walker

[Predator555 0]

ok,

so just another question

how does the amount an AI soldier has turned his head correspond to his view in 1st person

this is an image of the 1st man in the patrol taken fromthe eyes of the 2nd man in the patrol using Keykats script

now i would expect him to be able to see the barn but actually his view is this

so he cannot see the barn but can see much further than he should be able too, try it yourself and see

this is a big problem imo, if i see the AI's head turn left 60 degrees i expect him to see what i would see by looking left 60 degrees, instead he can see what i would see by lookiing over my shoulder, so in effect i do not think i have been spotted but i have been which is unfair.

do you all know what i mean? or am i just completely wrong?

[walker 0]

Hi all

In Reply to your point Predator555

We also have to add in this is not just eyes in line with the head you gain about another 30 to 45 degrees with eye movement.

I admire your tenacity Predator555 but if you also add that this is limited to matching what a player can see we have established the AI is not gaining an undue advantage.

One tries in a simulation to get close to reality but there are physical limits such as screen size. Effectiveness of interface. Effects on the user. Pressing keys on the keyboard is not the same as running up a mountain. What you see on a computer screen is less than what you see in reality. We are getting better at these things. Trackir and other new interface methods will improve things drastically.

But in the end...

All simulations are just that simulations. They are not reality. They cannot match reality, nor would we want them to; otherwise all the ArmA/OFP players would be long dead. In my case somewhere on the Hill above Houdan less than 50 yards from where the Deuce and half dropped us off in the Demo game almost a decade back now.

The only questions are is the simulation Valid?

Can the results be Verified?

We then come to questions of certification but that is outside the scope of ArmA. Though not VBS.

Kind Regards walker

[Predator555 0]

so you agree that it is a flaw in the game that when the AI is looking left 60 degrees it sees objects that would possibly be in its peripheral vision in real life yet cannot see objects that in real life would be at the centre of a real persons view like the barn?

seems a little crazy dont you think and allows for situations to arise where you dont think you should have been spotted but in fact have been, remember if this is BIS's way of modelling peripheral vision which may i add i doubt then then we must remember that peripheral vision is unreliable and unclear especially when we consider the distances involved which while small are pretty significant when it comes to how it would effect peripheral vision in real life, peripheral vision is better than nothing but that is all it is and certainly not as fullproof as the AI's appears to be.

there is still something wrong here, whether it is as small an issue as the AI having to good a peripheral vision or not there is something wrong.

if this were carried out in real life how many times out of 100 would you be spotted if the soldiers only looked left and right 60 degrees, my guess is that it would be a  very small amount of times that you were spotted with this "peripheral vision", certainly nowhere close to 100% of the time

[walker 0]

so you agree that it is a flaw in the game that when the AI is looking left 60 degrees it sees objects that would possibly be in its peripheral vision in real life yet cannot see objects that in real life would be at the centre of a real persons view like the barn?

Hi Predator555

No. I do not agree that it is a flaw in the game. Your three question invetiably spill over into each other.

As to your second paragraph. You are sort of repeating an earlier question. This is the glass half empty versus glass half full one. You think it was an unfair spot I think it was a fair spot. You placed your video and demo mission as evidense and in one of my first posts in this thread I answered it and gave the solution and I followed it up with a demonstration video.

The question is not is the simulation realistic? I have answered that, if it was real you would be dead. The question is always; is this a valid simulation given the constraints?

I answer your third point thus:

We are dealing with physical limitations of hardware; monitors do not do peripheral vision, graphic cards cannot model depth of field etc. In the end such things have to be compromised.

Secondly the AI does not see at all. We do. So what is on your screen is not what the senses of the AI percieve nor is the neural net perceiving as we do in our brains.

With modern systems and as our understanding of physiology and psychology improves and as computers become more complex and powerful our ability to simulate improves; we draw closer to what people would now classify as The Matrix or Star Trek's HoloSuite but you can go back further to Des Carte and his evil demon and that lead to Cogito Ergo Sum. But just because it would fool you does it make it the same as reality?

Thirdly from the last point you are up against all those deep philosophical problems here. From Plato's Cave and the Veil of Perception, through to Cogito Ergo Sum, the Mind body Problem, the Tripartite Theory of Knowledge, Turing Tests etc...

These are the questions that spill from any Turing Machine and the software that runs on it.

I can sum it up most easily with some seemingly simple questions.

"What is it like to be an AI? Is what I call blue the same as what you call blue and do you even exist in order to call anything blue?"

Do a search on VV&A and you will start to understand this area of simulation.

And finaly I leave you with this: When you kill an AI in ArmA are you commiting murder?

Kind Regards walker

[Predator555 0]

ok i think it would be best to keep the philosophy out, lets keep it to questions about realism of the simulation and how accurately it portrays the real world rather than questions about whether or not the AI is alive or not

simply we want to know,

1) is the AI behaving realistically as humans would when put in a similar situation?

my answer is no, and i think anyone who looks at the videos and tests themselves would agree, the AI will detect you a significantly higher percentage of the time than a human soldier in real life. I accept that it would be possible for a human soldier to detect you, you have changed my mind on that somewhat, however i would assess that risk as being medium to low, certainly not every time.

2) can this be fixed

my answer is i think it would take a small fix, i accept your comments about limitations of computers for modelling peripheral vision, i think it probably could be represented on monitors with some innovative thinking but for now it cannot. i think it would be better if it cant be done well then either just dont bother with it like most games or come to a compromise on the view and reduce the ability of the AI to spot you some, that would probably satisfy me.

[Johan_D 0]

if the AI view angle can be reduced, then it is all solved.

can we?

[walker 0]

Hi all

First off I believe that we have already adequately proved that this is not a bug. As I showed in my reply to Predator555's demo mission and video. I can avoid being seen in the example by the AI 100% Every time.

No Reducing the angle that AI can see is not the correct solution. Would you then put blinkers on the human players so that they could not see as far to the sides because they keep spotting you? We have established that AI sees just as much as we do.

Humans already have a big advantages over AI. Our brains are phenomenally faster and more experienced. We think in ways we cannot even begin to program into AI.

The easiest solution is to alter alertness essentially make the AI dumber and this already exists it is in the editor and the user config and difficulty settings. So the solution is an easy player and mission fix. Play at lower difficulty settings on Cadet rather than veteran and reduce AI Skill of Units in the editor.

THAT SAID!

I continue to contend that the errors are in improper SOP for the Ambush. ArmA is and has always been a game that punishes you for transgressions in your battlefield Tactics, Techniques, and Procedures.

It has always amused me to see the BF, CoD4 and CS crowd staggered that they cannot just stand up in the middle of the street blasting away. In ArmA you learn to crawl real early. You learn that a bush is concealment not cover. You learn that if you fire a shot and bust your concealment, you had better not miss. For ArmA player cover is thick walls, preferably several and mountains; mountains make excellent cover. You learn that real rifles have bullets that take time to reach the target and that they are not lasers unaffected by gravity so at long distances you can even drop one onto a reverse slope always a rewarding kill.

I would contend that you learn these things precisely because ArmA is so unforgiving. War ain't forgiving and neither is ArmA.

Kind Regards walker

[Predator555 0]

a solution would be for the characterss head to be capable of turning so he can look over his shoulder, this applies to both the human character of the AI character

if in the center of your screen you see a door then in 3rd person the characters eyes should be focused on that same door, its not that hard a concept and i doubt it would have been especially hard to implement either, i dont know why they chose to limit how far you can turn your head yet still give you the full view

Quote[/b] ]First off I believe that we have already adequately proved that this is not a bug. As I showed in my reply to Predator555's demo mission and video. I can avoid being seen in the example by the AI 100% Every time.

if not a bug, then a design flaw

Quote[/b] ]Humans already have a big advantages over AI. Our brains are phenomenally faster and more experienced. We think in ways we cannot even begin to program into AI.

thats not a good justification to give the AI unrealistic abilities, and neither is the fact that we have the same abilities in game, we should all be looking for realism in a simulation like Armed Assault and the current setup is not realistic, if i see an AI has turned his head a certain angle i dont expect him to be able to see beyond what i would expect a similar human person to see in real life.

Quote[/b] ]I continue to contend that the errors are in improper SOP for the Ambush. ArmA is and has always been a game that punishes you for transgressions in your battlefield Tactics, Techniques, and Procedures.

It has always amused me to see the BF, CoD4 and CS crowd staggered that they cannot just stand up in the middle of the street blasting away. In ArmA you learn to crawl real early. You learn that a bush is concealment not cover. You learn that if you fire a shot and bust your concealment, you had better not miss. For ArmA player cover is thick walls, preferably several and mountains; mountains make excellent cover. You learn that real rifles have bullets that take time to reach the target and that they are not lasers unaffected by gravity so at long distances you can even drop one onto a reverse slope always a rewarding kill.

I would contend that you learn these things precisely because ArmA is so unforgiving. War ain't forgiving and neither is ArmA.

agreed, armed assault is a very good simulation in a lot of respects but i am dissapointed in you that you continue to try to justify a clear design flaw, like i said before this could be easily remedied if BIS had designed the heads as i said so that the centre of their "screen" corresponds to the position that the eyes focus on and also BIS should have allowed the AI to turn their heads to look over their shoulders.  Had they done this i could have had no complaints, the AI soldier would have looked straight at me.

EDIT

also it should be noted that again errors in my SOP arent relevant, we are testing the game here in a particuar scenario, but the aim is to test the levels of realism in the game, just because you could have completed the scenario in a better way is not important and would not justify the design flaw, although i agree with you that i could perhaps have been better concealed. What you are saying is no different to me saying that if a game has bugged weapons that cannot fire that it was the players fault for not using a grenade because in that particular case it may or may not have been the better choice even though the bugged weapon is the issue, do you know what i mean?

[KeyCat 131]

...if this were carried out in real life how many times out of 100 would you be spotted if the soldiers only looked left and right 60 degrees, my guess is that it would be a  very small amount of times that you were spotted with this "peripheral vision", certainly nowhere close to 100% of the time

I think you have a valid point here and I agree that (at least in this particular test scenario) that the AI shows a touch of "un-human capabilites" when it comes to detecting a static target in their peripheral vision. IMO toning this down a notch or two would make current AI feel a bit more "human".

I ran this particular test again several times and with different skill levels on AI (0.00, 0.25, and 0.50 in editor and with skillEnemy=0.820000 set in my config), laying prone but the squad leader spots you 100% of times, skill level didn't seem to make any difference here?

Again remember we are laying prone, completely still, in the shadows with no sillouete and with camouflage clothes on and on top of that the AI have the sun in their eyes when looking in that direction in this particular test mission wich should make it even harder.

I think it's a good thing having hard ass "super human" capabilities for the AI as long as it's scalable and we have the possibility to adjust it from dumb AI to super AI and anywhere in between (like we now can do with AI's gun dispersion) so it would be great if BIS somehow also tied the AI's peripheral vision capabilities to the skill level slider or some other config setting.

I really don't see the need to argue if this is a bug or a design flaw but I do think this "nitpicking" and our small experiments shows that the AI's peripheral vision capabilities is something BIS could spend some time on to improve further, if not for current ArmA then defiently for ArmA 2 (and please give us working stock smoke grenades for AI in ArmA 2 while your'e at it ) !

Also wan't to say thanks to Oden for sharing the info. I will try fiddle with the values and see if I'm able to come up with something that tones down the AI's peripheral vision globally but if anyone beats me to it please share....

I'm sure BIS are as interested as we are to have proper AI behaviour and we saw major improvements to the AI's hearing capabilities with the 1.09 Beta so who knows what the future brings, until then I guess all we can do is to adapt and overcome

Again, just my 0.20 SEK!

/KC

[lugiahua 26]

actually one problem is

This game is designed for very wide screen(something like triple screens)

So the AI has same looking angle as triple screen player

but no every player has this type of screen....

anyway, if there are any major improvement on AI

I hope they are

1.pathfinder on bridges

2.AIs take cover properly behind rocks or vehicles...

[walker 0]

Hi all

Lugiahua yes we have mentioned the peripheral vision limit of monitors several times. Track-IR more or less reduces the problem and a Triple head system fixes it. For everyone else you have to use the the BIS peripheral sensors. Some people like them some do not. You also have the number pad direction keys but in all honesty you have to be triple handed to do get the best out of that; any mutants want to lend us all few genes? Some have mapped the

Like I say I use Track-IR like Predator555 that is the best solution for head movement. And Triple head solves the peripheral vision problem.

But that is all human side and we are really talking about AI side.

Predator555 your Anim point is accepted

Predator555 I accept that the anim could look a little further over the shoulder as a visual clue to the fact you have been spotted BUT KeyCat's experiment established the AI is looking at you and can see you.

I have already established that what is shown in either KeyCat's experiment or your screen shots above is what a human player would see and matches the baseline in my experiment.

And we have been through the philosophical fact that even though that is what the experiment shows a human would see the AI does not infact see.

NOW WE ARE GETTING TO THE MEAT!

Predator555 your acuteness of AI peripheral vision problem is accepted

We have through proper research refined what is the problem Predator555 has identified. The problem is AI peripheral vision is too acute. They spot 100% of the time in peripheral vision.

And Predator555 accepts he is a little lacadaisical with his TTPs.

KeyCat has scientifically tested my proposed AI Skill level solution. It does not work. We can now all repeat his experiment and confirm his results.

The solution is as KeyCat said really one for ArmA2 but BIS have been known to chuck in a fix to properly identified problems in patches in the past.

And of course ArmA has the best modding community of any game.

We now need to consider what is required for such a patch/mod:

1) How often does a human spot you in peripheral vision?

This requires either a military researcher to do the math with hundreds of people and an animatable dummy in various conditions and degrees of movement or some analogue of the same.

OR

That several subject matter experts come up with a formula to describe peripheral vision spotting.

OR AND THIS IS BEST!

That the research has already been done. We hunt around and find it on the web.

OR AND THIS IS DOABLE

A lot of ArmA players run an experiment designed as a game and have the system test how often we spot the AI.

NEXT WE NEED AN ALGORITHM

What kind of algorithm is needed?

For static subject.

Logarithmic curve to the edge?

http://en.wikipedia.org/wiki/Logarithm

Or some form of parametric curve as KeyCat suggests

http://en.wikipedia.org/wiki/Differential_geometry_of_curves

In the end the Algorithm needs to fit the research data and a parameterized solution allows for the most flexibility should research data improve.

What then are the parameters?

Sun should be an across vision solution so should we ignore it?

I would contend that in peripheral vision movement is a 100% spot but heck make it it a parameter too.

Camouflage again seems to be an across vision solution but do camouflage effects increase or decrease in peripheral vision?

Light levels

Background

I invite others to add in parameters.

There is a lot to cover.

This is how to do it in ArmA we deal in solutions.

Kind Regards walker