Killer Robots - U N Seeks Debate

[Riedquat]

It's a challenging task..

Imagine for a FPS shooter. In order for the AI to be 'realistic' (and not totally unfair) it needs to work in a similar way to the human players. Which means taking a rendered image and working out what objects are present - which is a hard task, but largely done. From this it has to learn a sequence of actions likely to 'win'. To some extent these can be learnt separately.

If it's really clever it would be learning from it's mistakes as you played.

However, I honestly think that the main problem in this area is that a computer AI that is better than human players would be boring. Like playing chess against the best computers now, you are just not going to win.

I don't think that FPS AIs try to find the players by rendering an image; that would be difficult (computer vision still has quite a way to go) as well as very slow, although they will calculate things like whether the player is in their line of sight or not. They'll probably cheat occasionally if nothing much is happening by going to the player that they couldn't really know is there.

I've found it quite interesting playing Skyrim, sometimes if you're standing somewhere where the enemy can't get to you but you can keep raining arrows on them they'll scurry off and actively search for a place where you can't hit them. Conversely they're incredibly stupid a lot of the rest of the time (creep up so they can't see you, stick an arrow in them and they'll wander around saying "Could've sworn I heard something. Oh well," with an arrow sticking out of their head).

I think that we've a long way to go where an FPS AI could do better one-on-one tactics-wise than a sensible player but if it was programmed to do anything to win it would still win due to being able to out-shoot and out-react.