Feedback Directed Optimization

Chun Wai Liew

HPCD-TR-13, CAP-TR-23

Optimization is a very important part of the design process.  There
are few design problems where concerns for either cost, quality,
design time, etc., are not important. A great deal of time and design
effort is spent on determining how to generate a solution that is
optimized for a particular set of criteria, e.g., cost or time.
However, optimization remains an ill-understood part of the process in
many design problems.

This thesis describes an innovative approach towards finding good 
solutions, i.e., optimized solutions, by using information about interactions 
between components (local interactions) gleaned from earlier solutions.
The approach is called Feedback Directed Optimization (FDO).

FDO is an iterative design approach based on the assumption that
information about local interactions between solution components are
essential towards being able to converge on good solutions to resource
optimization problems. The approach includes techniques for (1) 
credit-blame assignment to determine where local interactions occur
that might have been overlooked by the problem solver (2) 
controlling the problem solver on subsequent iterations to generate
better solutions.  These techniques have been developed and tested on
several problem solvers in multiple domains. In addition, we have
analysed the approach and have developed a prescriptive framework
whereby new and existing problem solvers can use our techniques.