In videogames, difficulty is a difficult thing to get right. It’s one of the reasons multiplayer is so popular after all; to date we’ve not come up with an AI that comes close to a human opponent, outside of chess at least. Now, it’s not hard to just make an enemy hit harder, have more health, or shoot with greater accuracy. Those things aren’t difficulty in a meaningful sense, but they do make the game harder.

Still, there’s not a lot original to say about this tendency to take the easy route and bump up the enemy’s pure abilities rather than their intelligence. What I want to talk about is a different aspect of AIs, which is something I’ve not seen often addressed, but which will ultimately be core in creating convincing enemies who are challenging, but can be defeated.

That aspect is making mistakes. Making believable mistakes, based on oversight, or failure to account for something by accident, and so on and so forth, rather than the result of glitches or the programmer’s failure to account for something. This may not seem like a huge concern while we’ve still got to figure out a way to be outwitted by the AI, but as we do get better at that this sort of thing is going to be crucial to correct for it in order to keep the game both fun and engaging.

A lot of victories in real conflicts are borne from taking advantage of mistakes the enemy makes. Sometimes this is a tactical error, sometimes strategic, and sometimes it’s more deeply rooted and occurs in the years before the war breaks out, when someone’s guess about the important factors of the next war prove to be incorrect. Oftentimes these things will be corrected over the course of the conflict, but sometimes not. In any event the point is that for the player to remain engaged and interested there can’t be an optimum strategy in all situations, which a ‘good’ AI would seemingly be prone towards, and which would thus force the same degree of efficiency from the player.

Of course in the real world there are all kinds of factors that are very hard to emulate. The Confederacy’s best option was probably a Fabian strategy – ceding land for time, and winning by attrition. But the political nature of the CSA meant that border states could not be sacrificed in such a fashion, and they had to be fought for (Well, except when McClellan was in charge of the Army of the Potomac, then not much of anything needed to be done by the Confederates). You can, to some extent, work with this in a game through mechanics like supply lines, dissent, and partisans, but it really has trouble with the nuances of the situation.

Now, getting games to that stage would be a tall order of course. Nevertheless I think we could stand to start thinking about how AIs might make believable, varied mistakes. Things that an astute player can see and exploit, but which the AI might realize and fix very quickly as well. This isn’t a completely untried concept of course, Galactic Civilizations 2 is the obvious example of an AI being designed to do this sort of thing, and it’s a commendable attempt, especially because the AI is actually pretty darned smart without cheating. Halo likewise had some clever foes, for its day, and their dynamic nature meant mistakes on their part could emerge pretty naturally and an observant, smart player could exploit those very well.

What do you all think about this idea? Am I getting too far ahead of our current, braindead AIs, or is this something we should look towards?