University Events

Description: Topic: Improve Decision Support System Operations Through Evidence Based Knowledge Evolution

Location: Lester W. Cory Conference Room, Science & Engineering Building (Seng), Room 213a

Abstract:
This dissertation describes a new and novel approach to machine learning design maximizing the balance between accuracy, efficiency, solution justification, and rule evolution. This dissertation improves four factors of machine learning within a single design. Different components of the decision support systems (DSS) design shall aim to improve one of the four factors compared to traditional designs. During this dissertation, traditional machine learning methodologies were altered to improve different aspects of machine learning, tested against a single application.

The researched examples depict the tradeoff between system accuracy, time efficiency, and storage space efficiency. Typically, to increase the accuracy of a machine, the system requires more time for computation and more historical data points for comparison. Similarly, to improve run-time efficiency or storage efficiency, the user must trade off the solution accuracy.

A key area investigated was how to build a DSS system capable of providing solution justification to the user. A DSS's purpose is to aid the user in making decisions. However, if a user does not understand why the DSS is providing a given recommended solution, the user is not able to make an informed decision. An outcome of this dissertation is an investigation of how a DSS can provide the probabilistic rationale behind the decisions the DSS presents to the operator.

The second area of focus is the DSS operational efficiency, including both the time and storage requirements. As part of this dissertation, an investigation into methods to reduce the quantity of historical data required to provide consistent rule validation was conducted. New rule evolution and frequent pattern tree algorithms were tested in an attempt to reduce the time necessary to refactor the decision tree and provide results based on a new case. New algorithms from Amazon and Google were tested to see if advancements in the last ten years provide accurate results within the bounds of space and time requirements.

Accompanying the operational efficiency is improving the accuracy of the knowledge base. Beyond testing new algorithms, data requirements for providing accurate results were investigated. When a knowledge base is given bad data to train, the knowledge base learns incorrect rules. For a knowledge base to learn and validate rules, the knowledge base needs to be trained using useful, valid data. During this dissertation, new and novel ways to both minimize the data required for a knowledge base, and ways to provide the right data were explored.

The final area of investigation was an improvement in rule evolution within a knowledge base. A crucial piece of a knowledge base is the ruleset used for evaluating new cases. Rules are generated based on expert knowledge or by allowing a machine to learn autonomously through accessed data. An outcome for this research area is determining how a knowledge base can do rule validation after rules are added, removed, or altered. To validate new rule sets some traditional methods require rerunning all previous cases through the new rule set. Part of this dissertation was exploring ways to validate rule changes against a smaller historical set, without requiring a massive amount of computational resources necessary to process a full historical data set.

At the end of this dissertation, a single design was delivered addressing each of the four factors in a single design. The final design depicts a single working decision support system that improves accuracy, efficiency, justification, and rule evolution. The implementation compares the improved system against a traditionally designed system. Each improvement is system-agnostic, allowing use in other systems without the other alterations.

NOTE: All ECE Graduate Students are ENCOURAGED to attend.
All interested parties are invited to attend. Open to the public.

Advisor: Paul J. Fortier
Committee Members: Dr. Honggang Wang and Dr. Liudong Xing, Department of Electrical & Computer Engineering, UMass Dartmouth; Dr. John T. Hays, Senior Fellow, General Dynamics Mission Systems

*For further information, please contact Dr. Paul J. Fortier at 508.999.8544, or via email at pfortier@umassd.edu.