Dynamic Models under Uncertainty for Large-Scale Enterprise Systems

Thesis (Ph.D.)--University of Washington, 2016-06

Many enterprise applications such as assortment planning, inventory control and supply chain management rely on forecasting, optimization and machine learning methodologies. While many methods have been developed for static deterministic systems there is a need for methodologies that are dynamic, involve uncertainty and are computationally practical. Additionally, new methods are needed for online applications that have a constant stream of new data arriving. Current methods are often unable to handle these online applications as resolving the entire problem every time new data arrives is intractable. This dissertation presents new models and algorithms for large-scale dynamic systems. The new methodology utilizes a framework for dynamic models applicable to online applications where there is not complete information and streams of new data arrive. The methodology for dynamic models consists of a probabilistic forecaster and an algorithm for approximating the solution to nonlinear least squares problems for updating parameters. The methods are developed to scale well with a large amount of data and to be used in near real-time applications. The models and algorithms developed can be applied to an enterprise system. We analyze the methods developed in this dissertation for the example of demand forecasting. Like many problems in enterprise systems, demand forecasting problems are dynamic and change over time and have a great deal of uncertainty. Furthermore demand forecasting problems are often large, containing either a lot of data or a complex model with many parameters. Cloud computing is becoming a major computational resource for streaming large scale data. Cloud computing can be used to increase the effectiveness of shared resources and solve problems that require a massive amount of data. However, it is not always clear how to maximize the effectiveness of shared resources on the cloud. A meta-control framework is proposed to achieve good performance for cloud applications. The measures of performance for the meta-control are latency, delay, restart and cost. This framework will design rules to characterize cost, quality of service and other criteria for implementing the meta-control feedback schema.