AI Discussion (dev branch)

[gammadust 12]

i'm less and less qualified to share specific opinions, since there is a long time i don't dive into arma's AI, but 2 notes:

 

- in another thread cosmic10r shared an interesting video, where AI would scan for a threat recently made aware. This is interesting because it hints at the expected human-like behaviour when one does not exacly know where a communicated threat is. One of my earlier issues with AI target sharing was its 1 to 1 precision, not merely the no-delay and range aspects. In the video i can't tell if the lack of precision is due to communication imposed message degradation or the low precision information available to the originator AI. The scaning behaviour, while not new, looks very believable in this context.

 

- i share the opinion that inter-group information sharing should be considered another beast, i'm sceptical on attempts to address it before basic inner-group communication can be found acceptable. Also this is relatively easy to address via mods.

 

Things that i started to look out for regarding target acquisition via communication:

 

- Rate, Accuracy, and Immediacy of Target sharing (ingroup)
- Accuracy of prediction routines
- Likelyhood of shooting under low expectations of success

[Gudsawn 93]

Regarding disableRemoteSensors - is this disabling an old feature, or a new feature (i.e. the new AI-can't-see-through-trees update)?

 

Also, would I be right in saying that you'd want to run this command on the server (where all AI are local to the server, and no AI are local to clients), to improve performance?

[seba1976 98]

Regarding disableRemoteSensors - is this disabling an old feature, or a new feature (i.e. the new AI-can't-see-through-trees update)?

 

Also, would I be right in saying that you'd want to run this command on the server (where all AI are local to the server, and no AI are local to clients), to improve performance?

As far as I understand it, no, because players are there and they are not local to the server.

 

 

 

These raycasts are used to determine what other entities an entity can see, and they take a lot of CPU time. This is of course a bit of a trick, because rather than a true optimization, it disables part of the simulation. However, there are certainly types of scenarios where these raycasts are not needed. An example is a fully Player-versus-Player scenario, where the visibility between every combination of player entity is not needed. So why not disable this by default? There are cases where you do require these raycasts, for example in stealth scenarios. Without them, commands like knowsAbout, nearTargets and targetKnowledge will only function for local units! The commands themselves are local, can be used on servers and clients, and the state is reset when the scenario ends. So, carefully consider whether your scenario can benefit from this method.

[Gudsawn 93]

As far as I understand it, no, because players are there and they are not local to the server.

 

 

Commands disableRemoteSensors and getRemoteSensorsDisabled let you disable raycasts for remote entities in a group that doesn't have any local entities.

So we'd want to disableRemoteSensors on the server, right (since we, the remote clients, don't have any local entities)?

[seba1976 98]

 

So we'd want to disableRemoteSensors on the server, right (since we, the remote clients, don't have any local entities)?

AFAIK, no!  :) You can't use those commands in coop scenarios. Whatever side with the remoteSensors disabled, won't see units that aren't local. If you disable the remoteSensors in the server, the AI created on the server, won't be able to detect any player units. That's at least what I can make out of the descriptions given.

[Gudsawn 93]

AFAIK, no!  :) You can't use those commands in coop scenarios. Whatever side with the remoteSensors disabled, won't see units that aren't local. If you disable the remoteSensors in the server, the AI created on the server, won't be able to detect any player units. That's at least what I can make out of the descriptions given.

 

 

 

Hmmm, are you sure? Doesn't disableRemoteSensors disable sensors for remote users (hence the name). In other words, the remote users (players) can't sense the local users (AI). It seems that AI will still be able to detect players as normal, but the server won't be able to detect if we (the remote players) have detected the AI, hence this note:

 Without them, commands like knowsAbout, nearTargets and targetKnowledge will only function for local units!

 

Which means we can still use these commands on the server (since AI are local to the server) but not for players (who are remote). As long as all AI are local to the server (and disableRemoteSensors is ran only on the server), there should be no issue. The only issue I can see is that the server's AI won't be able to detect if we (the players) have detected them.

[SilentSpike 84]

You guys are actually both wrong (mistergoodson is mostly right) :P

 

Basically, wherever you run the command, ray casting calculations are disabled for any group which doesn't contain a unit local to that machine.

By default:

 

1. Players are local to their own client.

2. AI placed in the SQM file are local to the server.

3. AI spawned via script are local to whichever machine they were created on (this is the basis behind headless clients).

 

I'm also not sure if:

1. AI in a player driven vehicle become local to that player.

2. AI in a group with a player become local to that player (or what happens if there are multiple players in said group).

 

Maybe oukej can clarify those two.

[inlesco 233]

Fundamentally speaking, shouldn't any soldier go down instinctively when shot by any caliber higher than 9mm?

 

The impact immediately makes you stop (basic human instinct), not continue to run forward as it is now when AI, for instance, hit with a silenced 7.62 in the leg. 

 

So, multiple custom animations should be triggered on 'Hit' EH to override the current movement direction and speed.

 

Perhaps someone already implemented this realistic behavior via a mod?

[seba1976 98]

Basically, wherever you run the command, ray casting calculations are disabled for any group which doesn't contain a unit local to that machine.

 

So, AI local to the server will detect the players because raycasting will be enabled for them to use them to detect what ever unit's around, regardless of the target locality, I'm I right? But for example a trigger like EAST detected by WEST, will not trigger because the server will not be raycasting from remote BLUFOR player units, detecting local to the server EAST units. Is that what you're saying?

[SilentSpike 84]

Yes that sounds correct.

[seba1976 98]

Well then, in a nutshell, disabling remote sensors in a client, or server, means: this client, or server, will not know if non local units have detected local units. Is that right?

[SilentSpike 84]

Well then, in a nutshell, disabling remote sensors in a client, or server, means: this client, or server, will not know if non local units have detected local any units. Is that right?

 

Almost :)

[x3kj 1247]

Fundamentally speaking, shouldn't any soldier go down instinctively when shot by any caliber higher than 9mm?

 

The impact immediately makes you stop (basic human instinct), not continue to run forward as it is now when AI, for instance, hit with a silenced 7.62 in the leg. 

 

So, multiple custom animations should be triggered on 'Hit' EH to override the current movement direction and speed.

 

Perhaps someone already implemented this realistic behavior via a mod?

When you are sprinting from one cover to another the basic human instinct is not to stop right there in the open

[ruPal 143]

Since 1.52 GL5 is useless... I don't know what did you change in AI behaviour but AIs are too timid since stable patch.

[Variable 322]

When you are sprinting from one cover to another the basic human instinct is not to stop right there in the open

I think that when under suppressive fire, AI should sprint to cover if cover is close (say, 20 meters), if not, change stance to prone and hold position. That's only in terms of movement, of course.

[oukej 2910]

Since 1.52 GL5 is useless... I don't know what did you change in AI behaviour but AIs are too timid since stable patch.

Do you have any more details on that please?

[KeyCat 131]

 

Added: userType action can define a maximum distance for AI

 

Wonder what  kind of distance are affected, hearing, vision, movement...?

 

/KC

 

[Jona33 51]

I've had the opposite experience to Kukus, on dev at least AI seem to have got really aggressive. Was testing on Panthera with me vs 5 AI, and they'd actively move towards me without any waypoints and try to kill me. Even running the same tiny piss-about test session was enjoyable to see if I could win, I couldn't. :D Might be placebo but they seemed really good.

[en3x 209]

Kukus does run mod called GL5 so. ( Group link 5 )

[ruPal 143]

Yes, I mentioned that. GL5 did a good job. It made AI much more real but since 1.52 it makes AI too timid. They don't react like in 1.50. In most situations they go prone and move to player in such manner. When they are close to player they just wait him to come.
I downgraded to 1.50 and it works great. They move fast to covers, try to flank you and 1-2 of them try to  come close to player as fast as possible. It was a great mod till 1.52. I don't know what was changed and don't know how to fix it.
There are several paragraphs about AI in change log and here is the one that I worry about :

Removed: Class-based differences in AI spotting abilities (it is covered by other systems instead)

I think that changes in AI caused problems with other AI mods. I know that there is no any reported but this one is the first. GL5 System is broken and works worse on 1.52 than 1.50.
I can make comparative video if you are interested in it, just ask.

[andersson 285]

Maybe BI have improved AI so the "fixes" in AI mods are no longer valid or need some tweaking. Lets hope that is the case, and what jona33 wrote is really promising :)

[en3x 209]

Since 1.52 GL5 is useless... I don't know what did you change in AI behaviour but AIs are too timid since stable patch.

That AI addon probably explains the different results... for tests everyone should run stock A3 without any addons and preferable use the same AI settings in the <name>.Arma3Profile so we compare apples to apples ;)

/KC

[ruPal 143]

 

Since 1.52 GL5 is useless... I don't know what did you change in AI behaviour but AIs are too timid since stable patch.

That AI addon probably explains the different results... for tests everyone should run stock A3 without any addons and preferable use the same AI settings in the <name>.Arma3Profile so we compare apples to apples ;)

/KC

 

Don't take in your account and thank you for your opinion, but my post was addressed to developers. In this situation you cannot help me somehow.

[andersson 285]

The rule is to not mix in mods into dev feedback. Is BI responsible of how the AI behaves in GL5? No, the developer of GL5 is who you should adress.

 

For sure, the AI could have some problems and you see it manifest in GL5 mod, but you/we can't know until you try without that mod. So if you try dev AI without any mods and find the same timid behavior, or something else which is not to your liking, then it's a good thing to report. But right now it's quite wrong for BI to change their AI so a mod that you uses will work better.

[chortles 263]

Maybe BI have improved AI so the "fixes" in AI mods are no longer valid or need some tweaking. Lets hope that is the case, and what jona33 wrote is really promising :)

Not surprising, a workaround for animating grenade launchers was broken when the devs did an official version.