The Future of Optical Fibre

An anonymous reader writes "An Australian researcher has come up with a novel way of developing optical fibres. Steven Manos, a researcher at the Optical Fibre Technology Centre in Sydney, Australia has developed a method of using genetic algorithims for discovering optimal designs of optical fibres. An article on his work had this to say "The problem with designing optical fibres is starting with a specific set of criteria and then coming up with a design to fit this. The computer program developed by Manos, which is run on supercomputers, does this by mimicking the process of evolution. The computer program combines two patterns to create a third fibre 'offspring', which Manos described as "similar but a bit different". This process is repeated thousands of times with the 10 designs best suited for the particular application chosen to 'breed' again." Another case of "When in doubt, use brute force"?"

Comments

I thought the future of fiber optics was...

[Karpe]

...wireless!! ;)

Re:Brute force? Not exactly

[russellh]

Ok, its not like a design from first principles, but its still way to search the parameter space without having to test all coimbinations of parameters

So-called first principles are explanation, not design tools. In other words - guess what? - nature is still surprising even if it can be "explained" by what we already know. We can explain stuff. It's the construction that we don't understand so well.

massive fiber overcapacity already?

[vijayiyer]

What is the purpose of this? Specialized interconnects? As far as I know, there is already massive overcapacity for telecom fiber networks.

Genetic Algorithm + Hill climbing

[G4from128k]

For exploring real-valued phase spaces, one solution is to combine a GA with a classical hill-climber. A hill-climber evaluates the local gradient (the partial derivatives of fitness with respect to the independent variables) and then makes a directed adjustment of the solution in the direction of better performance. Hillclimbers can reach optima in floating-point spaces very quickly, but tend to get stuck on local solutions.

GAs are great for jumping out of local optima to find new realms of the solution space, but don't converge as quickly on the neighborhood optima. So the combination of a GA with more classical optimzation can work well.

Re:Genetic Algorithm + Hill climbing

[namidim]

I would argue that though GAs are better than simple hill climbing they are far from being "Great" at getting out of local optima. In fact, a lot of the theory surrounding GAs has to do with how to avoid exactly that since the basic GA of some mutation combined with splicing for reproduction tends to get stuck extremely easily being limited almost entirely to the values represented in the original population. Evolutionary algorithms might be a better choice to pair with hill climbers. Evolutionary algorithms use vector cross products and gausian probabalistic mutation as the main operations for reproduction and my uderstanding is that they tend to get much better coverage of the search space than GAs. For instance a big problem in standargd GAs is that you get stuck with various combinations of whatever values were in the original population (only mutation can put in new values and mutated "genes" tend not to stick around for very long in the later stages of the evolution) . In EAs where most of the reproduction is based on probabalistic calculations however, you get a whole gradiant of values that change all the time.

Re:Brute force? No way?

[saigon_from_europe]

But only problem, just as parent said, is that

b) Some alternative minima are missed

A friend of mine got a job to work on genetic algorithms, in an academic institute. Being an engineer, he asked to be moved to another position 3 months after.

His explanation was very short: in GA you look for problems and you try to prove that they could be solved by this method.

Actually, all methods used in engineering were invented to solve some problem; not vice versa. Ok, maybe they not all of them, i canno say for sure since we learned about methods, not about their history. Most succesful methods moved from one area to another one, but basicaly it was problem before solution, not solution before problem.

The interesting point is that this is second story about GA in only couple of days; there must be some conspiracy theory to explain this.

Genetic Algorithms are so cool

[shaka999]

I've written my open GA for doing circuit optimizations and it works very well. The thing I love most is that they are so simple to write. There are things you can do to speed up convergence but the basic algorithm is very straight forward.

The difficult thing is how to score individual trials. I don't know how many times I've checked things after a overnight run and found that my results aren't what I expected. Pretty much everytime this comes down to how I've scored a trial. Just remember you get what you ask for.

For a circuit example, suppose I ask for a certain power comsumption and speed but I overstate the speed goal. Because I'm so far off the speed goal the power will largely be ignored. There are easy ways to tweak this but the point is...again...you get what you ask for.