Breeding Race Cars With Genetic Algorithms

smack-pot writes "Wired News has an article about how the Digital Biology Interest Group at University College, London is using genetic algorithms to breed superfast Formula-One race cars. 68 design parameters were configurable in the cars, and the generated designs were tested using the racing simulation software developed by the game developer Electronic Arts. According to the research it is possible to shave off 88/100th of a second per lap by using genetic algorithms to tune the cars. In an industry where a tiny fraction of a second matters, that's significant."

Comments

Difference between simulation and reality

[Minimind]

There is a large difference in evolved behaviour between physical things and models of those same things. GAs using physics simulators are very good at exploiting inaccuracies and subtle features of the simulation, making the transfer between the simulation to reality very difficult without the use of specialised techniques such as Minimal Simulations and Incremental Evolution.

This means you have to be skeptical with experiments performed just in simulation without testing the same model in reality.

what it shows ...

[curator_thew]

Is that genetic algorithms are nice for parametric optimisation, but not for breakthrough innovation.

Not very practical...

[Goonie]

I've seen this story floating round, and colour me unimpressed.

Genetic algorithms are terribly clever, and are useful for many purposes, but to make them work you need a "fitness function" - the ability to check how good a solution is. And, seeing you're going to need to apply it to every member of the population in each generation, it better be pretty bloody low-overhead, and be a pretty close approximation of the real-world fitness of a solution. In fact, in my admittedly limited experience with them I found that 99.9% of the difficulty in applying genetic algorithms to a problem is finding an appropriate fitness function.

The fitness function these guys have used is to use a racing simulation game and run the race electronically. That's good if you're trying to set up a car to win that game, but if you're actually trying to win a real car race with a real car, if the only fitness function you have is sending your driver out for a few million trial laps it's just not going to cut it.

If, on the other hand, they had built software that allowed them to specify the car settings and tell them what lap time the car would achieve, that would be really impressive, and then you could bolt on the GA optmizer to find the killer setup. But using GA's like they have done is just a party trick - cute, but not that impressive.