Abstract:
Decision guidance (DG) analytics systems are a class of decision support systems that are designed to provide decision-makers with actionable recommendations---courses of action that, if taken, would lead to optimized outcomes given that all underlying assumptions are valid. These systems have a wide-range of real-world applications and support decision-making across diverse commercial and industrial domains, such as logistics, manufacturing and supply chain management. Despite significant technological advances, DG analytics systems are currently difficult to develop and often require a significant level of effort from teams of experts with a broad range of skills from subject matter expertise to data science and software engineering. A major reason for this difficulty is the inadequate level of abstraction for the development of different mathematical models needed for performing different analytical tasks, such as calibration, prediction and optimization, where the same underlying reality needs to be modeled multiple times using task- and tool-specific modeling languages and frameworks. To address this problem, in this dissertation I propose and develop the first prototype of a new class of systems that I call DG analytics management systems, which are analogous to database management systems for the development and operation of DG analytics systems. The novelty of DG analytics management systems is that they embody a model-driven approach for the rapid development of DG analytics systems from a repository of task- and tool-independent analytical models and related artifacts that can be reused for descriptive, predictive and prescriptive analytics. Specifically, the key contributions of this dissertation are: (1) a conceptual architecture for DG analytics management systems; (2) an analytics engine for processing analytics queries against task- and tool-independent analytical models that can be reused for different analytical tasks using commercial and open-source algorithms for calibration, prediction and optimization; (3) a grey-box optimization algorithm framework for the efficient processing of optimization queries against partially analytical models involving a mix of closed-form analytical expressions and evaluations of numerical black-box functions that can be non-differentiable and computationally expensive; and (4) a prototype implementation that demonstrates the use of DG analytics management systems for the composition and optimization of manufacturing service networks.
