I'm not sure what that means. Science is empiricism, and genetic algorithms are a wide field of research with a lot of experimentation behind them. There are many hypothesis about how they work, why they work and what their limitations are. These hypothesis are tested through experimentation, just like in other scientific fields.
> Smart heuristics beat these hybrid-genetic-memetic-algorithms all the time.
Sure, this is almost guaranteed by the "no free-lunch theorem": https://en.wikipedia.org/wiki/No_free_lunch_theorem
There might be some exotic cases where genetic algorithms are actually the best solution, but nobody knows that.
But that's not the point of genetic algorithms. The point is having a generic solution that can be applied with some success when you don't have the time or the resources to find those "smart heuristics". And there might be cases where we are just not smart enough to find them. Some of these cases might be related to the effort towards developing generic AI. You can also claim that it is always possible to write a more efficient computer program directly in machine code, if you ignore time and resources.
Another interesting property of evolutionary computation is enabling some form of artificial creativity. Check out this antenna, created by NASA using evolutionary computation:
https://en.wikipedia.org/wiki/Evolved_antenna
Also, if you will forgive me a bit of self promotion, check out my own work on the discovery of complex network generators:
Scientific contributions are mostly in operators, I guess. I can't think what else can someone working with genetic algorithms contribute because adding elitism, or 3-way crossover isn't really the solution to a problem, it's just metaheuristics jibber-jabber.
So a lot of the research goes in the direction of finding the "missing ingredients". On the evolutionary computation front, one of the topics I find the most interesting is the genotype-phenotype mapping. Naive genetic algorithms have genotypes that map directly to phenotypes (e.g you are evolving some vector which is a direct solution to a problem). More complex mapping (like what we see in nature) encode a generator for a solution.
On this note, exciting stuff happens on the genetic programming front (where you are evolving computer programs, commonly Lisp-like trees).
It really is a vast field, ranging from direct engineering applications to artificial life simulations aimed at theoretical biology. Like everything else, you have to be familiar with the literature.
I do not like the research. It is bad, it's rarely reproducible, source code is almost never public thus time/convergence measurements are meaningless.
Indeed, the same could be said of any "smart heuristics" too:
> Exact solutions beat these smart heuristics all the time.
;)
Same goes for heuristics being faster than exact solutions for harder problems.
Actually, No Free Lunch guarantees that such cases exist.
"Dr. Thompson peered inside his perfect offspring to gain insight into its methods, but what he found inside was baffling. The plucky chip was utilizing only thirty-seven of its one hundred logic gates, and most of them were arranged in a curious collection of feedback loops. Five individual logic cells were functionally disconnected from the rest— with no pathways that would allow them to influence the output— yet when the researcher disabled any one of them the chip lost its ability to discriminate the tones. Furthermore, the final program did not work reliably when it was loaded onto other FPGAs of the same type."
Now, I agree with you that, even for a problem as simple as that one (or specially because it's a simple problem, maybe) a "normal" algorithm would perform better. But the fact that the genetic algorithm was able to find the effects of magnetic flux (those five disconnected logic cells) in that specific FPGA is nothing short of brilliant.
The problem I've always seen (not an expert, by any means, but I've played around with them a bit) with genetic algorithms is that they require a weird balance. For simple problems, a normal algorithm will probably work better. As the complexity of the problem increases, the scoring and evaluating rules' complexity increases as well (I've got no hard data, of course, but my feeling is that the complexity of the scoring process increases much faster than the complexity of the problem), making a well thought algorithm a better/easier solution. This, and that any mistake in the scoring process might throw away thousands and thousands of generations.
[0]http://www.damninteresting.com/on-the-origin-of-circuits/
http://people.mpi-inf.mpg.de/~doerr/papers/gecco2008_1.pdf
There is also a book on the topic that summarizes some of the current research
This is true. It's been formalized as the No Free Lunch theorem by Wolpert and Macready.
> Genetic algorithms aren't really that brilliant. They are simple, and that's what makes them popular among researchers. There's no formal science behind them.
There is a formal science behind them -- conferences, journals, the whole nine yards. The problem is, there's a dominant notion of genetic algorithms that's 25 years out of date, and nobody's been able (or tried too hard) to break through the noise.
(Disclaimer: I work in the field.)
Which is why GA/GP matters: It's part of a growing toolbox to allow us to find interesting candidate algorithms for certain classes of problems without spending as much human time to figure out cleaner solutions.
If you wish you can then have humans figure out why the solution works and use that knowledge to come up with something better.
Genetic algorithm thrives when you have a lot of variables that can be optimized to gain a favorable behavior and those optimizations are unknown.
Saying it's not brilliant because it's simple is like saying that binary search or merge sort is not brilliant because they are simple.
On the other hand you have TSP which murdered all those GA algorithms with a proper Lin-Kernighan heuristic implementation.
I'm not sure if we are talking about same problems. > Saying it's not brilliant
I said they were simple and that's why people like them.
http://github.com/verdverm/pypge
http://www.math.binghamton.edu/dept/ComboSem/worm-chiu.pge_g...
I bet that thinking about the problem more would allow you to find an algorithm that works better than genetic algorithms all the time.
Just look at travelling salesman problem, or vehicle routing problem. Smart heuristics beat these hybrid-genetic-memetic-algorithms all the time.