Sunday, November 18, 2007

Making NPC Behaviors More Plausible +

Originally Posted by Ereiid:
I have to admit, I'm a little confused now. AI is still predictable. There seems to come a point in the learning curve of most games where a player will learn to recognize and anticipate even diverse NPC enemies. With the GTAs, you figure out at some point that the Colombians are fond of their semi-automatics, whereas the Cosa Nostra have a penchant for particularly nasty shotguns. Announced intentions certainly assuage early-learning curve problems adapting to game mechanics -- but I'm unclear how they address the recurring problem of the predictability of stereotyped AI coding.
My point (such as it was) was that it's not necessary to have any stereotyped NPC behaviors (unless one wants them, of course). Again, this goes back to my original suggestion that "real AI" isn't necessary for a game; all that's needed is the perception of plausibility. The beauty of signaling intentions is that once you've done that, the natural human reaction is to try to find some way to relate the subsequent behaviors to the stated intentions. As long as the two things -- the statement and the action -- are roughly consonant, people will perceive intelligence on the part of the actor.

Sadly, this means there's not much call for the development of behaviorist AI methods. As a software guy with a longtime interest in AI, that's a little disappointing to me -- who wouldn't love to get the assignment to build real AI into NPCs? :)

On the other hand, given the complexity of a MMORPG and the limited development resources, a "fake" (i.e., perception of plausibility) solution that generated something like 50% plausibility that was relatively simple to code and maintain would have an enormous practical advantage over a 90% solution that can't be understood by anyone but the PhD who wrote it. A 50%-plausible NPC algorithm would be such an improvement over what we have now (NPCs as quest-dispensers and loot-bags) that it would be worth having even if it "cheated" by being more like ELIZA than a truly intelligent system.

Does this come anywhere close to answering your question? Basically I'm saying that it's not necessary to have predictable behavior for classes of NPCs; I think players would be OK as long as when an NPC signals some intention, the actions that the NPC takes make sense relative to the prior communication. As long as word and deed are in alignment, there shouldn't be any need for classes of NPCs to share the same predictable behaviors.

Again, though, even if my suggestion were followed there's nothing stopping a developer from giving groups of NPCs similar behaviors if it's felt there's some value in doing so. All I'm suggesting is that it may not be necessary to blindly copy that approach to NPC behavioral design.

No comments:

Post a Comment