I'm going to discuss the reasons and advantages of the learning algorithm I proposed and implemented for the diagnostic graph traversal, and why I think it is a good choice.
In the simplest case, we could've implemented a graph traversal algorithm that has no learning aspect. This would've been the easiest option as we all have implemented a DFS or BFS before, probably in a variety of settings, and will almost certainly do so again. But this would've been turning our backs on one of the best potential features of Mechanapp. Without any learning aspects to our application there would be nothing to set us apart from other similarly featured applications.
We also could have gone to the other extreme. Part of my research touches on the state of the art in machine learning. A variety of these algorithms would've worked for our purposes and maybe worked very well. There are problems with these though. Many of them require extensive training sets, training sets that we don't have. They may take a long time to become effective. The largest problem though, is that many of these algorithms are just too complicated. They are difficult to understand, difficult to implement, and difficult to debug. This just wouldn't work for us.
This is why I chose the simple learning algorithm that we've implemented. A greedy graph search with learned edge weights is simple. Simple enough that every member of the group has the background to understand it FULLY. I won't need to distribute machine learning papers to provide the other group members and myself with the background necessary to understand it. Additionally, in its simplicity it was easy to implement. I implemented two different versions in two different languages in one week. And it's exceptionally easy to debug due to its simplicity. These things are great, but the best part is, it will work. I believe it will work very well for our purposes. For these reasons, the algorithm is not too little, and it is not too much, I would say--It's just right.
No comments:
Post a Comment