Hacker News Reader

Comment by SavantIdiot | Hacker News Reader

Genetic algorithms work well in some circumstances, I recall that multi-band antenna design technology experienced a step-function boost in the 90's due to application of GA principles, which helped the cellphone market boom. (I could be mixing up two IEEE articles in memory.)

I think calling it a "hack" misses the nuance: mapping the physical metrics and telemetry into a genome is nontrivial.

Plus 40 generations is no where NEAR enough. A GA needs millions (or billions) of generations to avoid local minima/maxima. Especially given the size of the input vector. His vector is what, 2^100 states?

View on HN

GAs are great. Love 'em. Same with monte carlo -- too many people dismiss these approaches and get mired in gradient-based optimization (which is far harder, more expensive, and often doesn't work any better). There should just be a rule that you have to try linear regression before building a neural network, and GA/MCSA before doing gradient optimization.

> I think calling it a "hack" misses the nuance: mapping the physical metrics and telemetry into a genome is nontrivial.

Well, that's definitely what I was trying to say, so we're agreed.

I actually don't think GA a great fit for this problem as parking a car is a time- and circumstance-dependent process, and a fixed gene is...not. But that's a guess.

> I actually don't think GA a great fit for this problem as parking a car is a time- and circumstance-dependent process, and a fixed gene is...not. But that's a guess.

Oh definitely not! :)

Is Monte Carlo approach (versus a gradient based approach) basically trying a bunch of different things until something works, or is it more nuanced than that (perhaps with intelligent optimization of what to try next?)?

Monte Carlo is basically taking repeated random samples until you get a good approximation of the value you're interested in. Gradient descent is more like what you describe, except you don't try "a bunch of different things" but keep trying the same thing until you have a result you think is good.

To give a silly example, Monte Carlo is like trying to guess the surface area of a dartboard by throwing darts at it and counting how many hit it, while gradient descent is like throwing darts at the board until you hit the center (or something you think is the center).

Reply on news.ycombinator.com