home | bio | papers | books | lab | helen


 

 Tim Menzies

Prof (full). Ph.D. Computer Science.
SE, AI, data mining, prog. languages.

  tim@menzies.us  

 

"Less, but better"

I find simple solutions to seemingly hard problems (see examples or read tutorial).

So what can I simplify for you?

For Students

I seek talented grad students for AI+SE. Is that you?

For Industry

Ask me how to innovate. On time. On budget. Case studies:

My Funding

Over $8M. From many sources, e.g.:

   

Meet the team

My current graduate students (at the RAISE lab- real-world AI for SE):

Wei Fu Rahul Krishna Vivek Nair George Mathew Zhe Yu Di (Jack) Chen Amritanshu Agrawal Jianfeng Chen Guilherme Ferreira Andrew Hill

My prior graduates:

Scott Chen David Owen Ashutosh Nandeshwar Ekrem Kocaguneli Abdel Salem Sayyad Fayola Peters Joe Krall Greg Gay

What we do

I challenge my students as follows:

  • Researchers usually seek solutions that are more intricate and complex;
  • Yet empirically & theoretically the world we can know is very simple;
  • So can you do "it" better, with less?

Here are a few examples of their "Less, but better" results:

Software defect prediction

2007

Surprisingly effective defect predictors can be built from simple static code attributes.

2010

Static code defect predictors have inherent limitations. But these limits can fixed via a new learner, very simple learner, that better understand the business goals.

2012

Very simple social metrics can generate near-optimal predictors for software quality.

2015

A simple feature and instance selectors let software projects share privatized data, without missing important patterns.

Search-based SE

2002

Search-based SE methods can easily and readily and critically assess long held SE truisms.

2013

Active learners can simplify and reduce the cost of search-based SE by orders of magnitude.

2016

Very simple optimizers can dramatically improve the performance of data miners learning software quality predictors.

Transfer learning

2009

A simple nearest neighbor relevancy filtering resulted in one of the first general results in software analytics: defect predictor learned from Turkish toasters could be successfully applied to NASA flight software (and vice versa).

2013

How to transfer lessons learned from past projects? Easy! Clustering tools enable transferring lessons learned between software projects.

2016

Ultra-simple transfer learning methods (called "bellwethers") enable effective transfer of lessons learned.


Requirements engineering

2002

Contrast set learners find simple controllers in requirements models.

2003

Contrast set learners can explain enormous decision trees (6000 node) learned from complex requirements models just 6 rules.

Effort estimation

2013

Active learners can easily estimate large software projects after just a few samples.

2016

The effort to build complex software can be estimated by very simple equations.

Other applications

1990

The lesson of decades of expert systems research is that, for specific domains, human expertise can be readily captured in just a few rules.

2010

Simple contrast-set learners out-perform state-of-the-art optimizers for spacecraft control;

2015

Data miners can greatly simplify and reduce the effort involved in data collection for community health studies.

2016

Text miners can succinctly summarize thousands of technical papers about SE.


BTW: for the origins of the "Less, but better" mantra, see Dieter Rams' 10 principles for good design.