Games and AI
Jan. 26th, 2006 01:34 pm![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
luckykaaBy all accounts, Sony are pushing AI for the PS3, and want developers to really invest in this.
There's a widely held belief amongst gamers that decent AI is just a matter of CPU power, and so the PS3 with all its parallel cells should be remarkably good in this respect. This isn't the case. This is a fairly pointless ramble that #I've been meaning to ramble about for a while.
The problem is, AI takes a lot of developer time, and isn't really something gamers notice. Good character AI would have characters behaving fairly normally. Normal doesn't stand out. Tactical AI can also be a problem. There are a lot of situations where the optimal tactics can be proved mathematically, making the AI a totally unbeatable opponent, given enugh CPU time. At a simple level, Nobody can defeat a competent Naughts and Crosses game. This can happen with other games as well. A lot of games have a very limited set of possible moves, and given perfect timing which a computer has), so tend to play too perfect a game. The game often has to make suboptimal decisions, without looking like it's doing something stupid to help the player to win.
Most of the AI type systems in games have been quite simple, and often rather gimicky. The Thing had the marketting gimmick of a Trust/Fear mechanism. Trust was just a variable that went up if you gave the character a gun, and down if you shot him. Fear worked in a similar way, except required adrenaline shots, but generally this was all pretty scripted. The game works better that way. We don't want people going off and doing unexpected things. If people did sensible things then the game wouldn't be completable.
At the other end of the spectrum is killer tactics. There are often things you can do that will mislead the AI. Go in one direction. The AI predicts you will do X, and does Y. Player does something compeltely different which would normally be illogical but since the AI has left the previously defended location vulnerable in doing Y, it works. This is actually fair enough. Where it falls down is that these tactics can often work again. In a lot of cases we need to generalise what the player does. This is not an easy task. Machine Learning is a tricky subject.
What could work well is better army logic. Lord of the Rings apparently had some AI in the CGI animation model. Soldiers would pick a target, and attack it (or run away at times).
There's a widely held belief amongst gamers that decent AI is just a matter of CPU power, and so the PS3 with all its parallel cells should be remarkably good in this respect. This isn't the case. This is a fairly pointless ramble that #I've been meaning to ramble about for a while.
The problem is, AI takes a lot of developer time, and isn't really something gamers notice. Good character AI would have characters behaving fairly normally. Normal doesn't stand out. Tactical AI can also be a problem. There are a lot of situations where the optimal tactics can be proved mathematically, making the AI a totally unbeatable opponent, given enugh CPU time. At a simple level, Nobody can defeat a competent Naughts and Crosses game. This can happen with other games as well. A lot of games have a very limited set of possible moves, and given perfect timing which a computer has), so tend to play too perfect a game. The game often has to make suboptimal decisions, without looking like it's doing something stupid to help the player to win.
Most of the AI type systems in games have been quite simple, and often rather gimicky. The Thing had the marketting gimmick of a Trust/Fear mechanism. Trust was just a variable that went up if you gave the character a gun, and down if you shot him. Fear worked in a similar way, except required adrenaline shots, but generally this was all pretty scripted. The game works better that way. We don't want people going off and doing unexpected things. If people did sensible things then the game wouldn't be completable.
At the other end of the spectrum is killer tactics. There are often things you can do that will mislead the AI. Go in one direction. The AI predicts you will do X, and does Y. Player does something compeltely different which would normally be illogical but since the AI has left the previously defended location vulnerable in doing Y, it works. This is actually fair enough. Where it falls down is that these tactics can often work again. In a lot of cases we need to generalise what the player does. This is not an easy task. Machine Learning is a tricky subject.
What could work well is better army logic. Lord of the Rings apparently had some AI in the CGI animation model. Soldiers would pick a target, and attack it (or run away at times).
