Downloads

• Skynet 0.1.3.zip 99.4 KB
• Skynet 0.1.3 setup.zip 331.3 KB
• Download GNU tarball (latest version of the source project, straight from SVN)

You can also download the project code directly via SVN from the SourceForge source code repository, at https://csgalib-tsp.svn.sourceforge.net/svnroot/csgalib-tsp. From a command line, you can call the following: svn checkout https://csgalib-tsp.svn.sourceforge.net/svnroot/csgalib-tsp

Application features

• Graphical WPF interface
• Multiple selection algorithms
• Pause, resume and cancellation support
• Control over mutation ratio’s and elitism percentage
• Optional removal of identical twins or individuals
• “Take over the world” button

Background

The idea for creating this application came to me after reading the first part of Bio-Inspired Artificial Intelligence by Dario Floreano and Claudio Mattiussi. I figured I needed to do an implementation of what I’ve read to test myself. I split up the general GA code from the application itself and created GALib, a small C# Library that provides the scaffolding for Genetic Algorithm based functionality. All work was done in my free time.

See this page for more screenshots.

Using the application

Although the application interface should be pretty straight forward to use, I’ve written some documentation, which can be found here. If you have any questions, let me know

How it works

This section on my wiki explains how Skynet works as an implementation of GALib. If you are not familiar with how genetic algorithms work, you are advised to first have a good look at this Wikipedia article and related pages. This section will introduce you to how GA logic specific to the TSP works in a bottom-up fashion. For more information on the actual evolution, see GALib. I’m also planning to write a Code Project article about this implementation of GALib.

This is a full dependency digram generated in a Visual Studio solution containing both the GALib and Skynet projects.

Points of interest

Since my main motivation for creating this application was exercise, I learned a lot from building it. It’s my first decent C# application, as well as the first time I’ve created one using WPF and the first time I’ve done any GA programming (or AI in general). It also gave me the chance to familiarize myself with some of the new things of .Net 4.0, some profiling tools (see screenshot below), and have some fun with navigation based windows. The biggest challenge in the application itself (so not counting GALib), was definitely creating the crossover algorithm for the Route individual type. At first I simply took half of the connections of one parent, and then the other half from the other parent, but I rewrote this to take all common connections. Although the crossover algorithm works fine now, it’s pretty heavy on the cpu, and limiting the maximum speed of the application severely. If anyone finds a way to speed it up, be sure to let me know

I’ve been working on this application on an off for a month now, and have implemented everything I’ve planned and more. Although a lot of cool stuff could still be added, I’m quitting active development of both this application and GALib, so I can focus on new projects that allow me to further expand my understanding of AI.

1. I split the code into 2 projects: one holding the general GA code, which is now called GALib, and one holding the code specific to the TSP, now called Skynet (lol ). This allows for making any other GA implementation using GALib, even in .Net languages different from C#. The underneath diagram should give you a good idea of how the who thing works. The 4 most left classes are of Skynet, all remaining classes and interfaces are of GALib.

2. I rewrote the genotype of the Route individual type, the core of the TSP implementation. Instead of having a list of integers, referring to cities, I now have a list of Connection, each containing 2 instances of Location. I’ve done this to allow smarter crossover, decreasing the chance of not finding a better solution while it’s still in the converged search space significantly. As a result of this change, I also had to rewrite pretty much everything else of the Route class, including random initialization, mutation and fitness assessment. The biggest challenge here was preventing the creation of multiple loops in the crossover algorithm.

The interface of the TSP app (now called Skynet) also has seen quite some work. I figured out quite a lot of WPF stuff by further creating it. Some extra controls and progress indicators still need to be added though, and I also haven’t created the Cancel/Stop functionality.

Although the algorithm is efficient in the way that it takes relatively few generations to find a close to optimal solution, there remain a few mayor performance issues. If I compare the speed of evolution (measured in generations) with similar applications tackling the TSP, there are those performing up to 3 orders of magnitude faster. Although the fastest of those are written in C++ or C, I should be able to significantly speed up my application. Tracking down resource eating parts of code will be one of my next steps in developing this app. Another possible issue is diversity of the population. I have the suspicion that the diversity shrinks pretty fast, resulting into evolution driven only by mutation. I’m not sure of this though, and also have to investigate how I can prevent this from happening.

Here you have a few screenshots of the app in action