NoticeI have moved to Medium. You can find my new posts here. The below posts will stick around for posterity.
Metric K-Center [A]
So, my classes are over and the results are in. I’ve implemented a solution to the Metric k-center problem and, to be honest, the results are not all that interesting. But, I’ll go ahead and answer the questions that I proposed in my last post.
What will the solution look like?
The solution was composed of two parts, the runner and the viewer. Its important to solve the problem (obviously) it is also important to visualize you’re results. While it is easy to calculate metrics on how much your algorithm improves over a random solution, it’s hard to see where improvements can be made (or how they could be made).
The runner is targetted for JRuby and solves the application in parallel by solving for the various values of k. The output is then persisted to a MySQL database so that the logs can later be analyzed and rendered (by the viewer).
The viewer is just a simple Sinatra app that queries the DB for the latest runs and allows you to pick out a specific generation for a particular run and value of k and the results are rendered using the Raphael JS library.
What approaches will be best utilized for solving/estimating the problem?
By approaches I don’t mean frameworks or technologies, I am more referring to algorithms. For this, I chose to use a couple of different approaches and combine them together. The first is a genetic algorithmm (GA) and the second is the Wisdom of the Crowd (WoC).
The GA allows me to derive solutions rather quickly from an initial, random set of solutions (read up on the Wikipedia article if you need more information on GAs) but, like all solutions to NP-C problems it is merely an estimation to the solution. WoC allows us generate a consensus from the set of experts (top performing members) of a given population of solutions.
These two approaches will be the main driving factor for our runner.
How fast is it? (because all we care about is speed)
While I don’t have exact metrics, it takes about 32 seconds to solve a problem for 100 nodes for 80 value of k and 500 evolutions. That’s running on my home desktop.
Now, while I don’t have exact metrics I can talk a little bit about JRuby versus MRI Ruby. My choice to go with JRuby is two-fold. The first reason is that I get the benefit of real-threading. By this I mean that there is no GIL and when I run multiple threads, I get the benefit of JVM threads (running on multiple cores). This means that I can use libraries like peach. The second is faster overall runtime of the runner (regardless of whether or not threads are used).
Since these computations can be ran in parallel and since I have a quad-core machine, it only makes sense to run the tasks in parallel across multiple cores. However, if you’re a Rubyist, then you should be more than familiar with the limitations that come along with the GIL. From this, we can see why JRuby, in this circumstance, is a much better choice than the current MRI release.
What makes my solution better than other existing solutions?
Nothing really. There’s not much novel about my solution, but it was fun and I did suffice to learn quite a bit. I’ve posted the code (for anyone interested) on GitHub at JohnMurray/metric-k-center.git. I’ve also added a writeup of the project (PDF) that is included in the repository if you’re inclined to know a little more about it.