Work

On solving computationally expensive multiobjective optimization problems with interactive methods

Year: 2014
Type: dissertation
Abstract: In this thesis, we address the challenges encountered when solving multiobjective optimization problems with interactive methods. We concentrate on practical issues, that is, we discuss structures that can be utilized when implementing an interactive method and practicalities of connecting an interactive method with a multiobjective optimization problem formulated e.g. in different modelling environments. As most real-world multiobjective optimization problems are computationally expensive, we also study how to facilitate the use of interactive methods for such problems. Even though interactive multiobjective optimization methods have been widely discussed in the literature, their implementations are rare. In this thesis, we define a core structure of interactive multiobjective methods, where we identify the main steps required for implementing an interactive method. As a concrete example of utilizing the core structure, we demonstrated the IND-NIMBUS software framework that includes implementations of several different interactive methods. When dealing with a real-world optimization problem, the problem is often modeled with some simulator that must be connected to the optimization method implementation. In this thesis, we study how such a connection should be constructed, in order to enable changing the interactive method or the problem being solved with as little effort as possible. In addition, as the interactive method involves a decision maker, we propose an algorithm, where the decision maker can be involved in verifying the model of the multiobjective optimization problem in addition to solving it. When solving a computationally expensive multiobjective optimization problem with an interactive method, the decision maker may experience unfeasible waiting times. We address this issue by introducing an agent assisted interactive algorithm, where we replace the computationally expensive multiobjective optimization problem with a computationally less-expensive surrogate problem. With the algorithm, the accuracy of the surrogate problem is maintained by identifying the areas of the Pareto frontier that the decision maker is interested in. The introduced algorithm is not specific to any particular method. As an example, we implement the algorithm for the interactive NIMBUS method and PAINT surrogate method.
Language: English
Open Access status: green
OpenAlex ID: W2200314815
Cites: 121
Cited by:
Related to: 10