Welcome to Henrik Smedberg's thesis proposal seminar

Date: Tuesday, 22 June 2021

Time: 13.00 – 15.00 CET

Place: https://his-se.zoom.us/my/sunithbandaru

External Reviewer: Mats Gustafsson, Professor at Department of Civil and Industrial Engineering, Uppsala University


Title: Interactive Knowledge Discovery for Knowledge-Driven Multi-Objective Optimization and Decision Support


Abstract

Multi-objective optimization involves the simultaneous optimization of several objective functions. Since these objective functions are often conflicting in real-world problems, the optimization process leads to several so called trade-off solutions, all of which are no better or worse than the others. Typically, decision makers have certain preferences that are used for selecting a final solution to be implemented in practice. Most multi-criteria decision analysis methods focus on the solutions’ performance in the objective space. However, it is in the decision space where practically relevant knowledge resides. Having access to this knowledge can help decision makers gain additional insights into the problem and the optimization process, leading to a more informed decision process. The present thesis employs and further develops methods for knowledge discovery in the context of multi-objective optimization, and investigates how the extracted knowledge may be processed and presented to the decision maker in an interactive manner for better decision support. Both implicit and explicit forms of knowledge are explored, however, an emphasis is laid on the latter due to their benefit of being interpretable, unambiguous and amenable for programmatic processing. Further, the thesis also investigates how knowledge extracted from preferred solutions can be incorporated into multi-objective optimization algorithms or the multi-objective problem itself, to improve the solution process computationally. Such knowledge-driven optimization (KDO) can be realized in two ways, offline and online. Offline KDO refers to incorporation of knowledge obtained from previous optimization runs into future problem scenarios of a similar nature, so that the search process can be limited to preferred regions of the objective space. The key focus here is on storage and retrieval of relevant past knowledge, and modifications to the optimization problem formulation through an expert system. In contrast, online KDO involves interleaving knowledge discovery methods with optimization algorithms, and utilizing the knowledge obtained during runtime to enhance the search process, thus driving faster convergence towards preferred regions of the objective space. This requires developing new search operators suited for algorithmically manipulating different forms of knowledge. In both offline and online KDO, the veracity of the knowledge and the degree to which it is applied play an important role. The thesis uses several benchmark optimization problems, real-world engineering test problems, and case studies from the manufacturing industry to validate and demonstrate the efficacy of developed methods and algorithms.


Aim and Research Questions:

The aim of this thesis is to employ and further develop appropriate knowledge discovery methods for use in multi-objective optimization and decision making, and thus contribute to improving the state-of-the-art in preference-based optimization algorithms and multi-criteria decision support. Three overarching research questions have been identified to realize this aim:

RQ1. Knowledge Discovery:

How can different forms of knowledge be extracted from the solutions of multi-objective objective optimization problems, while considering different variable types involved in multi-objective optimization?

RQ2. Decision Support:

How can different forms of knowledge be processed and presented to the decision maker in an interactive manner for better decision support?

RQ3. Knowledge-Driven Multi-Objective Optimization:

a. How can knowledge extracted from completed optimization runs be utilized offline in future optimizations of similar problem scenarios to focus the search towards preferred regions?

b. How can knowledge extracted during an optimization run be utilized online to improve the convergence behavior of preference-based optimization algorithms?


Contributions to Date:

RQ1. Knowledge Discovery:

● Trend Mining 2.0 ● Influence Score

RQ2. Decision Support:

● Interactive Decision Support System

RQ3. Knowledge-Driven Multi-Objective Optimization:

● FPM-NSGA-II ● Expanding Hypervolume Metric

14 views