Firemint discusses 'human like' Real Racing AI
Genetic algorithms used to teach the iPhone to drive
Over on the TouchArcade forums, Firemint has been explaining how the CPU has been taught to drive in its highly anticipated racing game, Real Racing.
Rather than using a pre-programmed difficulty system, in which the computer drivers simply get better and better, Real Racing employs 'genetic algorithms' to provide a far more life-like opponent.
"What is great about genetic algorithms (GA) is that they produce AI that is much smarter, but also more human like," explains Firemint CEO Rob Murray. "What isn’t so good (as we have discovered) is that just like good human players, they misbehave and find exploits."
This new level of organic intelligence in the computer drivers has proven quite extraordinary, though it brings its own set of problems. As the AI learned to drive, it began to exploit certain areas of the game's design.
For example, the genetic algorithms allowed the CPU to figure out that hitting a corner at full thrape, then bouncing off the walls to face the right direction again was the quickest way to corner, and was an acceptable trade off against the amount of damage the car took in the process.
Once this was addressed, it began cutting the corners and driving on the grass - using the small loss in speed as a braking system.
"We have been addressing these issues as well as other more mundane ones, but it has taken time. The interesting thing is that these are the same exploits that human players could use, so in a way the AI training system has been acting like a massive beta testing team," Rob concludes.
It's a fascinating problem that promises a superb racing experience even in single player mode, though it's also the reason for Real Racing's launch delay. It should all help toward making for a superior product when it does hit the App Store, of course.
