Briala
@vampiress To be honest, there are not that many games where the term AI is actually a worthy description of NPC programming. (Alien: Isolation is one such, FWIW, and deservedly so.)
May 21, 2024 at 8:04amWeb
Show threadAlexander The 1stMay 21, 2024

@static @vampiress It's an issue because we're not necessarily talking about the fidelity of the A.I. Behavioral Trees in a game, but the general description of...building those.

Like,it occurs to me that even though I was specifically calling out UE4/5 Behavior Trees and AIControllers in an interview last year when discussing how I would try and do automated testing of a tool (Involving using the AIController to act as go-between between test and execution, like WebDriver.),maybe I hit a flag-

Show threadAlexander The 1stMay 21, 2024

@static @vampiress - even though I wasn't trying to be like "Let's use ChatGPT to do our automated tests!", I'm now worried that's why I ended up in second place in the interview candidates in the final round...but that's absolutely not what i was saying.

But I also can't really call it Command/Control setup without elaborating either, so...?

Show threadBrialaMay 21, 2024
@AT1ST @vampiress That's a good point. I was actually thinking how the agent code in Alien: Isolation learns (somewhat) from the player behaviour. That used to be called heuristics, I think? But I don't think many game AIs do learn; they just have very well constructed decision trees.
Show threadAlexander The 1stMay 21, 2024

@static @vampiress Two main reasons that come to mind why they don't "Learn":

1.) That's significantly harder to test that they don't learn some really odd behaviours;

2.) It's a lot of work that often won't be noticed unless you have the A.I. decide what it's going to do, and then communicate it with a vocal cue to the player (The Alien agent in Alien: Isolation gets around the vocal cue stuff by...not having to speak the language of the player.).

Show threadAlexander The 1stMay 21, 2024

@static @vampiress With 2.), I'm reminded that F.E.A.R.'s A.I. was notoriously good at stuff like flanking - but it required marking a lot of the map up with markers so that the enemy characters could say "Great, we're going to flank the player" - but if the enemy characters didn't *say* "Let's flank 'em!", it... probably wasn't going to be noticed that it *was* using advanced techniques.

Or put another way; a lot of work for something that looks very similar to significantly less work.

Show threadBrialaMay 21, 2024

@AT1ST @vampiress Which is basically pre-generated data for the decision trees.

Sometimes, I kinda wish I programmed these things for a living, TBH. It sounds like my kind of data-driven shenanigans. :) But I have to be happy reading about it and theorising.

Show threadAlexander The 1stMay 21, 2024

@static @vampiress It does sound amazing, but from a User Experience standpoint, it just... doesn't really pan out.

Like, it's lower than the Star Ocean PA systems, which tend to be more opaque than most relationship systems (And at least up until SO3 as I understand, much more complex than other relationship systems.); unless players run the game scenario a second time they aren't going to notice it doing different things.

Show threadAlexander The 1stMay 21, 2024

@static @vampiress (And they probably need to play it a lot to see the intricacies; I'm reminded of how #TwitchPlaysPokemon revealed how Lance's A.I. is designed to target weaknesses if the Pokemon can use them, and has infinite PP on his moves - because he spammed Reflect with a Dragonite against Butterfree that just spammed Poison Powder to win, and it took a lot of people by surprise that you could actually do that.

The game was more than 14 years old at that time, and yet...)

Show threadBrialaMay 21, 2024

@AT1ST @vampiress Yeo, the "learning" has to be fairly tightly controlled.

In Alien: Isolation, one thing I do know the alien "learns" is that if you use the same type of distraction (flare, noisemaker, etc...) too often it gets less likely to be fooled by it. Tightly controlled "learning". But effective.