Artificial Stupidity

Back when I was working on Combat Flight Simulator we had a problem. The game had varying levels of difficulty, but what did that mean? What’s the difference between Easy, Medium, and Hard? There were lots of things we did before we gave anyone a chance to try it out. Some were simple, like unlimited ammo and fuel, or invulnerability. Others were a little more involved, like changing how accurate your aim needed to be and the damage radius of things. Once we had that working we let others try it.

And it still wasn’t “fun”. This led to long discussions about what “fun” is, but one of the things we settled on was not just the ability to return to base successfully, but to complete the mission, and for air superiority missions, that meant stopping enemy aircraft. Unlimited ammo/fuel helped because you didn’t need to worry about having the perfect shot. Invulnerability or increased “hit points” meant you survived. But neither of those added up to stopping the other side.

What we found was that our AI was too good. Or at least never bad. We gave them a set of rules and models and they followed them. Exactly. Every time. Which meant that even if you knew the rules as well, if you made the smallest mistake they’d end up on your “six”. And that just wasn’t fun.

So what we ended up doing was implementing artificial stupidity. Depending on the difficulty setting, some number of the enemy AI and your wingmen will make suboptimal decisions. They might roll too far out of plane, or not turn at the right radius. And sometimes they’ll just do something random (simulating panic).

And of course, what’s old is new again. Seems like some folks working on chess AI took that approach and went a step farther. Not only does this new version make mistakes, it can make mistakes like a specific person. Of course, being able to predict a person’s moves, their tell, isn’t new. Apparently what’s old is new again. 

In retrospect, having your opponent be perfect isn’t much fun. We should have known that.