Hacker News Reader

Comment by kvonhorn | Hacker News Reader

What field are you in?

Particle accelerator physics. They are used in the optimization of charged particle optics.

Purely out of curiosity - is there something about them that makes them better than other optimization algorithms, e.g. simulated annealing, random local search etc, or for continuous functions gradient descent and the like. Seems like GA's are always glossed over as a relic of past times, it's pretty interesting to hear about them in practice somewhere!

Haven't worked on them a lot recently, but AFAIK, the population based nature helps them escape local minima and do a better job at exploring the solution space.

We use them because they are particularly well suited for multi-objective optimization. They are also a global techniqe as opposed to gradient based methods and have great scaling properties. Important because each evaluation of the objective function takes ten minutes for us.

GAs have found a niche in a lot of "obscure" engineering situations. Anything where you need to optimize an expensive black box function is at least worth a try with a GA. I know they are used in analog IC design, civil engineer, and aerospace.

They can be very good when there are multiple objectives / tradeoffs. I wrote my thesis on that, which includes an introduction to the concept and how it works if you're interested. http://master.matsemann.com/

Reply on news.ycombinator.com