Machine Learning for Requirements Engineering

(File Last Modified Wed, May 29, 2002.)

• Graduate Studies: The First Impression
• Summer Project
• The Result

Graduate Studies: The First Impression

This is the first paper professor Menzies gave me at the very beginning of May 2001 for my summer project. It effictively destroied my confidence and expectations for the coming school life of graduate studies in UBC. More than 3 times had I abandoned it in desperate and doubted my ability to read in English. At that time, I absolutely couldn't imagine it was this paper that became the foundation of my research afterwards.

Things sometimes just change quite dramatically if you do not give up easily( maybe this is the way life should be?). Summer has passed and new term began. Instead of writing a review, I'd like to summarize the three months' research experience for it was the actual consequence of reading this paper.

Summer Project

The project called ``Treatment Learning'' aims at designing and implementing a new machine learning technique, or more specifically, a data mining/summarization tool. Before I started the project, Dr. Menzies had built a prototype with prolog based on his previous research. Due to the limitation of prolog itself and the unfinished algorithm, its performance was not quite satisfactory at that time. My goal was to fully understand the underlying principle and implement it in C with optimized algorithm.

I started from learning the basic background knowledge in AI. The first 2 weeks were classical machine learning literature review with special focus on a benchmark decision tree learner C4.5. This was followed by another 2 weeks concentrating on our specific algorithm. Dr. Menzies met me once a week to discuss the progress and technical details. By the middle of June, I had finished the C implementation, which is called ``the TAR2 TreatmentLearner''. When I came to him with this working TAR2, he was too surprised to believe. :-)

After some modification, we planed and conducted large-scale data experiments, which included the development of a N-way cross validation facility. TAR2 was run on 11 datasets from the UC-Irvine machine learning data repository as well as corresponding 10-way cross validation trials. According to the satisfactory results on those datasets, we performed 2 case studies: one was related to his reachability theory and the other was an attempt of using TAR2 to find out control variables from software risk estimation data generated by the COCOMOII model. These applications finally led to 2 research papers submitted to MBRE01(The first workshop of model-based requirement engineering).

The Result

The TAR2 treatment learner is an intuition of new mining algorithm dedicated to summarize data log generated from early life cycle activities such as requirement engineering, and by doing so, try to find controllers in the large search space. It is the tool for studying possiblilities and the endeavor to master them.

As a new graduate student, I gained necessary background knowledge as well as some insights of my future research from this project, which, in my perspective, are both a good start and strong motivation.