AI improvement suggestion: more human-like behavior

[Slappy]

I know some of you see the "AI improvement" subject and think "WTF the AI are already too strong!! Damn those RLF..." However let me explain.

We all know the AI has many advantages over a human (perfect aim, able to switch weapons or charge a melee weapon instantly, etc). These were put in to compensate for the AI's weakness - not being able to act as intelligently as a human guard.

What follows are my ideas to shore up the AI's weakness. This may result in the AI needing less perfect aim or other advantages/cheats.

(I think ElJoelio has done a great job with the AI, but as with any computer code, and ESPECIALLY AI, there's always room for improvement.)

1) Human guards almost never walk. One of the more common times a human guard might walk is when he heard or thinks he heard a thief, and wants to listen more closely, yet still move around. I propose that the AI run unless they heard something suspicious. Their hearing could be reduced while running, like a human guard's.

2) AI don't throw flares in situations where humans do. If a human sees a torch put out (which is different than seeing a doused torch - I'm talking about the situation where the human hears or sees the water arrow hit the torch), he won't run straight to the torch to re-light it. He'll throw some flares or flarescan the area before re-lighting the torch. Another situation that calls for throwing flares is when a thief was spotted, and then lost. The human will flarescan/throw flares into nearby shadows, while the AI just walks to the last known place of the thief and kind of bumbles around holding a flare.

Since we don't want the AI throwing unlimited flares, they would have to hit the supply chest when they run out. They'd probably have to get more than 3 flares per resupply though, because I doubt they could use them as effectively as a human.

3) I mentioned this one in an earlier thread: Respond to seeing a door/window open. At least get suspicious about it until a guard is seen through the door/window. If no guards are seen within a few seconds, and the door/window closes, and there is a shadow on the guard's side of the window, he should flarescan it.

I'm sure there are others. Add your own - lets add some things to ElJoelio's todo list!

[Machine]

Whilst these ideas seem simple in plain english, I would imagine them to be hard to code.

You hit on the AI's weaknesses a lot Slappy, without taking into merits their rediculous advantages :-

*Perfect Aim.
*Extended hearing.
*The ability to determine Theif and Guard noises at a distance.
*The ability to pick out a theif running around when all they can hear is the thundering trampling of 6 other guards in the area.
*The ability to "lock on" to a thief once spotted.

I do agree with much you've said, I don't like the idea of running around, i don't know why, it just destroys the atmosphere.

And I love the idea of AI having limited supplies like humans.

[The_Dan]

I'm guessing you haven't realised, but how well an AI plays a map is partially dependant upon the map itself.

I know for a fact that mappers can make AI respond to:
-stolen objectives
-deactivated electric lights
-custom triggers

You can also change AI properties so that they always run & hook up some triggers to make them change movement speeds after the appropriate stimuli.

I know joel is working to improve how the AI handle different situations, but mappers themselves can achieve a fair degree of realism if they experiment with TSupsiciousThing & TOrdersChanger.

[Machine]

Ahhhh, tis all clear now!

Then surely the next version of Aqua should allow the bots to turn the lights back on?

[Slappy]

What follows are my ideas to shore up the AI's weakness. This may result in the AI needing less perfect aim or other advantages/cheats.

Notice this part TM - what I meant is that if the AI is made better in certain ways, it wouldn't need to have all those advantages that you mentioned.

The_Dan - I know 2 things about mapping. Jack and shit.
Oh, 3 things actually - the third is, I can't do it.

I did read your comments in another thread about how the mapper of Aquatone could've made the AI turn the lights on.

However, I'm assuming this would make the AI run directly to a light and turn it on. A human would not do this. A human, when confronted with the darkened hallway, would toss some flares first to make sure a thief isn't right there by the lightswitch, waiting to KO. I'm assuming that this flare-tossing behavior is in ElJoelio's realm, and not the mappers.

It does sound like the mappers could greatly improve the quality of the AI, but since we can't even get proper locationID's everywhere, I don't see it happening

I'd especially like to see most maps re-done so that the AI can actually follow a human almost everywhere. Currently, the AI following is pathetic. I'm assuming this has to do with the pathnodes or whatever they're called. For example if you grab an AI in Nostalgia and tell him to follow you, then you run out the front door and make a loop around the house, the AI will NOT follow you the whole way. There are a couple places you need to "babysit" them or push them through doorways. This definitely needs fixing.

As for AI's running all the time - why not? Human guards run all the time. It's one of the major reasons that you can tell the AI is an AI just by glancing at him - its behavior is so obviously different than a human's, which is what I would like to see changed.

Perhaps, The_Dan, some map standards could be worked out regarding pathnoding and AI triggers (and locationIDs while we're at it)? Are there any standards the dev team places on maps before they can become an official TUT map?

Holy crap my capacity for rambling on endlessly seems to be greatly increased in the small hours of the morning.. enough for now.

[Machine]

Currently, the AI following is pathetic

Absolutely, and I can't believe that the Ai still can't get into the Greenhouse on Asylum!!

How hard is it to Pathnode? I've never done it before properly? From what I understand you place points in UnrealED that the bot can follow, each point having links to other nodes that can be directly reached.

It'd be cool being able to run around the level with an Ai following you and you wouldn't have to worry "I wonder if my ai is still there and not stuck in a doorway!"

I love the AI, it's often more helpful than a human because they can camp enlessly without getting board, and if your stood in the next room you can hear them when they are in trouble because they always should for help.

And i agree we should be making them less inhuman, but to a point I don't think they should be made more human. (If that makes sense)

[MortalMonkey]

Noise filter. It works by this rule:

Any noise that is louder (each noise should have a loudness factor that is multiplied by the sound's proximity to determine loudness) than other noises make the weaker noises even weaker in the AI's perception by an amout proportional to how much louder the loud voice is over the weaker.

[Machine]

Noise filter. It works by this rule:

Any noise that is louder (each noise should have a loudness factor that is multiplied by the sound's proximity to determine loudness) than other noises make the weaker noises even weaker in the AI's perception by an amout proportional to how much louder the loud voice is over the weaker.

WHA?!?

[LaughingRat]

Noise filter. It works by this rule:

Any noise that is louder (each noise should have a loudness factor that is multiplied by the sound's proximity to determine loudness) than other noises make the weaker noises even weaker in the AI's perception by an amout proportional to how much louder the loud voice is over the weaker.

WHA?!?

"TheMachine once again demonstrates the limits of AI programming...."

[The_Dan]

I did read your comments in another thread about how the mapper of Aquatone could've made the AI turn the lights on.

However, I'm assuming this would make the AI run directly to a light and turn it on. A human would not do this. A human, when confronted with the darkened hallway, would toss some flares first to make sure a thief isn't right there by the lightswitch, waiting to KO. I'm assuming that this flare-tossing behavior is in ElJoelio's realm, and not the mappers.

Right, but I tend to see things on a sliding scale of human-AI realism:

Current AI = Completely ignore deactivated lights = Bad

Proposed AI = Run towards lights & then switch them on = Improved

Humans = Check out the area, then switch lights on = Eventual target for AI


The proposed method isn't that easy to ambush the AI, as you'd think. When bots see a TSuspiciousThing, their overall awareness goes up & so they will be more sensitive to nearby thieves.

Eg. On most normal groud surfaces, you can easily run up behind an AI for a easy Blackjack KO, since his awareness is pretty low. However, if you let that AI see an unlit torch, & then run in for the BJ while he's relighting it, he'll quite likely turn around & look for you since his awareness of his surroundings has been raised.

The same thing applies to using TSuspiciousThing for electric lights. I've tried it out


Also, I think there is a trigger available that is activated by a bot gaining Line-Of-Sight to it. By placing this on a light switch, & then connecting that trigger to a TOrdersChanger (set to investigate), you can get a bot to investigate an area when he sees that the switch has been turned off.

Even if Joel does program AI to be more humanlike when dealing with darkened areas, I doubt he'll program them to check every switch in the map & see what switch the bot must locate to reactivate the light. Mappers need to give bots a helping hand in understanding what they must do in order to keep an area well-lit.

[The_Dan]

I was thinking of doing this anyway, but once I've got NeoPendragons scripting out of the way, I'll knock up a modified version of Aquatone with the suggested AI-improvement & then we can see how this compares.

[The_Dan]

Perhaps, The_Dan, some map standards could be worked out regarding pathnoding and AI triggers (and locationIDs while we're at it)? Are there any standards the dev team places on maps before they can become an official TUT map?

Currently, as long as a map has been properly pathnoded, it's playable.

To try & implement all these extra AI considerations on a new map isn unreasonable. However, I wouldn't mind see-ing them slowly phased into existing maps, I'd like to see the AI play more of a role in multi-player & provide a good human substitute in single-player.

The stuff I'm suggesting applies more to some maps than others. I keep mentioning aquatone, because it contains no torches (which AI can handle) but lots of electric lights (that are beyond the comprehension of current AI). This makes the single-player way too easy, & so the multi-player feels especially harder by comparison.

I'd like to be able to get as realistic as possible Thievery practise in single-player, so that I can stay sharp when I don't play online for a while.

[Bloody-Reaper]

And people are complaining the ai is to hard...
I wouldnt like to see a bot go flare scanning an area.

[CrouchingDork]

1. So far, people have said "Genius is ok, bla bla bla" but the basic thing they said was exploit their weaknesses, like how they leave their back open a lot, and walk, and don't get suspicious when doors open, etc. If we implemented this we'd up the %80 gaurd wins to %95.

(Please don't freak out on me, I'm entitled to my opinion, and my opinion just happens to be against yours)

2. I never walk because I don't even have a key bound. :grin:

3. However, I DO think some of the suggestions are good. It's just we're going to have to balance it out if they're going to be more human-like.

[Lurox]

Noise filter. It works by this rule:

Any noise that is louder (each noise should have a loudness factor that is multiplied by the sound's proximity to determine loudness) than other noises make the weaker noises even weaker in the AI's perception by an amout proportional to how much louder the loud voice is over the weaker.

WHA?!?

I must agree with you...

WHA?!?