Better reasoning about software engineering activities

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

• Review: Better reasoning about software engineering activities
• Problem Addressed
• Approach Proposed
• Details
• Conclusion
• Comments

Review: Better reasoning about software engineering activities

Problem Addressed

• Software management experience can be encoded into an oracle that can offer advise on how to structure a software project. Three components are required:

   a knowledge source (basis)
   a syntax (to encode)
   an interpreter (to excute)

• Stochastic simulation samples a subset of possible behaviors (by the form of datalog).
• Those possible behaviors can then be summarized to find the controllers.

Approach Proposed

• knowledge source: a detailed description of CMM2
• syntax: JANE language
• ``Chances'' weight used to handle degrees of belief, varies according to distributions during the simulation.
• TAR2 treatment learner is used to summarize the datalog and find controllers.

Details

• JANE supports 2 mechanisms for random search:

  When defining ``costs'' and ``chances'', a rane and a skew can be provided.

  JANE uses several novel operators including conjuctive operators, simple disjunctive operators and summing disjunctive operators.

• The CMM2 model is written in JANE.
• Learning from JANE is done by the TAR2 treatment learner.

Conclusion

TAR2 found three best treatments which are constraints that can improve the overall result.

Comments

• The approach discussed in this paper can be summarized as: lightweight modelling --> randomized simulation --> summarization via machine learning. The underlying idea is: sample possibilities through random search and find controllers in the search space.
• I think this is Dr. Menzies's research thesis and the answer to my initial question: what can artificial intelligence (or machine learning) do for software engineering. I'm so glad to have this whole picture after these months' study.
• For a more practical research project, I hope to have a real world model on which we can apply the approach and see if the learning derived will actually have the impact. Yeah, this is exciting.