3 ms·

I hate to be "that guy", but I'm somewhat confused why everyone online seems to be so obsessed with genetic algorithms. What they are is a kind of optimization

by apu 15y ago

I hate to be "that guy", but I'm somewhat confused why everyone online seems to be so obsessed with genetic algorithms. What they are is a kind of optimization technique, and one of the worst ones at that. (Genetic algorithms are what you resort to in the rare cases when no other optimization technique will work.) As 'lincolnq said, standard hill-climbing would probably work much better for this kind of problem. Still, cool results, I guess.

Strilanc 15y ago
It has a wider appeal because of its similarity to biological evolution.
yid 15y ago
In my experience (I worked on convergence analysis of some early GAs), genetic algorithms suffer from a combination of extremely easy implementation, a "cool" factor that plagued neural-networks-as-classifiers research in the early 90s, and a belief that biological "analogies" would perform better than decades of research in stochastic optimization. Still, they have pedagogical value because they are easy to implement, and for many students, are their first real soup-to-nuts "AI" algorithm implementation. I've seen them serve as the gateway for students to get into CS research, but IMO aspects such as the last resort principle, no free lunch theorem, and principles of stochastic optimization are generally under-stressed, which leads to some abhorrent research papers along the lines of "Problem X using GAs".
brianobush 15y ago
I use GAs initially to find out how hard the problem is that I am attempting to solve. Why? They are simple and work well.
mdda 15y ago
As a counterpoint to "that guy", what I like about GAs is the localization aspect : While modern mathematical methods can do sampling (and so optimization) more efficiently, they seem to be in the "Intelligent Design" camp. GAs can work using extremely local information - and let the force of numbers pull the process through. <grin> Besides, evolution is the optimization method that God chose. </grin>
evolvingstuff 15y ago
Standard hill-climbing would indeed work better under the assumption that the fitness landscape is fairly unimodal, in terms of local optima. In this problem that may or may not be the case. I think that this is the motivation for a somewhat more "global" search strategy (something that can escape local optima) like genetic algorithms. That being said, amongst the possible candidates for search strategies, genetic algorithms are fairly lousy. Differential Evolution or CMA-ES would likely work far better.

by apu 15y ago

Strilanc 15y ago

It has a wider appeal because of its similarity to biological evolution.

yid 15y ago

In my experience (I worked on convergence analysis of some early GAs), genetic algorithms suffer from a combination of extremely easy implementation, a "cool" factor that plagued neural-networks-as-classifiers research in the early 90s, and a belief that biological "analogies" would perform better than decades of research in stochastic optimization. Still, they have pedagogical value because they are easy to implement, and for many students, are their first real soup-to-nuts "AI" algorithm implementation. I've seen them serve as the gateway for students to get into CS research, but IMO aspects such as the last resort principle, no free lunch theorem, and principles of stochastic optimization are generally under-stressed, which leads to some abhorrent research papers along the lines of "Problem X using GAs".

brianobush 15y ago

I use GAs initially to find out how hard the problem is that I am attempting to solve. Why? They are simple and work well.

mdda 15y ago

As a counterpoint to "that guy", what I like about GAs is the localization aspect : While modern mathematical methods can do sampling (and so optimization) more efficiently, they seem to be in the "Intelligent Design" camp. GAs can work using extremely local information - and let the force of numbers pull the process through. <grin> Besides, evolution is the optimization method that God chose. </grin>

evolvingstuff 15y ago

Standard hill-climbing would indeed work better under the assumption that the fitness landscape is fairly unimodal, in terms of local optima. In this problem that may or may not be the case. I think that this is the motivation for a somewhat more "global" search strategy (something that can escape local optima) like genetic algorithms. That being said, amongst the possible candidates for search strategies, genetic algorithms are fairly lousy. Differential Evolution or CMA-ES would likely work far better.