Review of reviews

So I got two reviews on my proposal, which was posted last Monday (2/3).  I intend to make revisions based on the comments I received, but also from my own thoughts.  I think reading a couple other proposals in the last week and hearing what some other people had to say about what makes one a winner or a loser has given me a better idea of where to go with it.

Some notes from the review by Ronald Shaw:

1.) Ronald said in a couple different ways that I should pick a platform and some of the key technologies (such as programming language) in the proposal.  I think he's right and I intend to incorporate that into the next release.  I think it will go along with what we talked about in class last week, "making nothing into something".
2.) He's also right that I did not fully identify my stakeholders.  I mainly only described my customer base.  I need to also think about and include the development team, automotive dealers, mechanic shops, and potentially negative stakeholders such as AllDATA (maker of some existing automotive repair software).

Some thoughts on the review by Kevin Dilts:

1.) Kevin mentioned this project proposal reminded him of some discussions from the CS 427 (Intro to AI) class last semester.  That is about the time this idea started brewing in my mind, actually.  We did a couple (rudimentary) examples of a theoretical logic system that could diagnose a car.  Much later in the class, we discussed different types of AI expert systems, including case-based expert systems.  I think this is how the idea got into my head that a hybrid expert system (case-based and rule-based) for car repair could be something new and potent.  I was also pleasantly surprised to learn that such a system apparently does not exist at present.
2.) Kevin mentioned a potential challenge of the system not addressed by the proposal.  He said there seems to be a risk that a system applying rules for diagnosis in a probabilistic manner (based on case histories, as this project proposes to do) has some risk of giving bad diagnoses.  I do not believe this risk exists, but it illustrates that perhaps I didn't explain this part of the system very well.  Imagine the system is leading the user through a diagnostic chart as if it were a tree, using some kind of graph search algorithm like Breadth-First Search (BFS).  However, in this system, the search is directed towards the branches of the tree which are believed to be the most promising (i.e. are the most common causes of a particular problem on that vehicle or class of vehicles in the past).  So, the search is traversing the most promising branches of the search space first (we hope).  Even with the worst possible information guiding this process, the search algorithm is still complete despite its probabilistic nature (after all, it's built on top of BFS).  I intend to simplify and detail this description and add it to the proposal so it may be more easily understood.