Over at next-gen, they have an article by Edge on the difficulties of creating artificial intelligence for NPCs/opponents in games.

On the one hand, you don’t want the computer to be too smart. How many times have you played a strategy game, for instance, and had the feeling the game was cheating on you? It probably was, because the AI had been too well-constructed.

The same is true for shooters. Being computer-guided, enemies have a much easier time of targetting you, because the program is always faster than you are, and far more precise.

On the other hand, you don’t want the opponents to be too stupid, either, or the game has no challenge. If all the enemies are graduates of ISSOM (the Imperial Stormtrooper School of Marksmanship), there’s little tension and no sense of achievment in getting through a level.

So the goal, essentially, is to create an AI and then “dumb it down” to where it approximates a real human being, neither too smart nor too stupid. And, of course, capable of errors in judgment, just like us.

What would be interesting is if the AI could learn as we do. If we see a pattern in enemy behavior, we can adapt to that and come up with a method to defeat it.

We have patterns, too. What if the AI could detect that, and adapt itself accordingly? This would probably not work well in a shooter or RPG, where typically enemies die by the cartload, leaving no survivors. In a strategy game, though, it might be possible.

However, would that be making the AI “too smart” again? Would players think the game was cheating? Or would they catch on, and realize they’re becoming predictable? And do you think it’s possible sometime in the future to have AI that’s indistinguishable from human behavior?

Artificial Idiocy on next-gen