Decision support for complex planning challenges - Combining expert systems, engineering-oriented modeling, machine learning, information theory, and optimization technology
Doctoral thesis, 2014
This thesis develops an approach for addressing complex industrial planning challenges. The approach provides advice to select and blend modeling techniques that produce implementable optimal solutions. Industrial applications demonstrate its effectiveness. Industries have a need for advanced analytic techniques that encompass and reconcile the full range of information available regarding a planning problem. The goal is to craft the best possible decision in the time allotted. The pertinent information can include subject matter expertise, physical processes simulated in models, and observational data. The approach described in this paper assesses the decision challenge in two ways: first according to the available knowledge profile which includes the type, amount, and quality of information available of the problem; and second, according to the analysis and decision-support techniques most appropriate to each profile. We use model-mixing techniques such as machine learning and Kalman Filtering to combine analysis methods from various disciplines that include expert systems, engineering-oriented numerical and symbolic modeling, and machine learning in a graded, principled manner. A suite of global and local optimization methods handle the range of optimization tasks arising in the demonstrated engineering projects. The methods used include the global and local nonlinear optimization algorithms. The thesis consists of four appended papers. Paper I uses subject matter expertise modelling to provide decision analysis regarding the environmental issue of mercury retirement. Paper II provides the framework for developing optimal remediation designs for subsurface groundwater monitoring and contamination mitigation using numerical models based on physical understanding. Paper III provides the results of a machine learning study using the Compiling Genetic Programming System (CGPS) on multiple industrial data sets. This study resulted in a breakthrough for identifying underground unexploded ordnance (UXO) and munitions and explosives of concern (MEC) from inert buried objects. Paper IV develops and uses the model mixing and optimization approach to expound on understanding the MEC identification technique. It uses the methods in the first three papers along with additional technology. Each thesis paper includes complimentary citations and web links to selected publications that further demonstrate the value of this approach; either via industrial application or inclusion in US government guidance documents.
analytic hierarchy processes
data modelling engineering-oriented modelling