Interactive Evolutionary Multi-Objective Optimization Algorithms: Development, Improvements, Benchmarking and Analysis of Performance

  • Seyed Mahdi Shavarani

Student thesis: Phd

Abstract

Interactive multi-objective optimization algorithms exploit the preferences of a human Decision Maker (DM), elicited iteratively during the optimization process to direct the search and increase computational efficiency. Despite their undoubted promise, progress in the field has been hindered by challenges in experimenting with these methods, the main one being the fact that the results are affected by the decisions of the DM and the difficulty controlling or accounting for natural variation among DMs. One may propose using many DMs to eliminate biases in the experiments arising from DM variability. However, this would make the experiments expensive and non-replicable. The other approach is to replace the human DM with a Machine DM (MDM) in the experiments. Only a few studies have considered the simulation of decision-making behaviors in the context of interactive methods, most of which have used arbitrary utility functions to indicate the desirability of the solutions. However, it is not enough to replace the DM with a utility function assuming an ideal DM whose decisions are consistent throughout the experiments. This thesis builds on one of the only proposed MDMs that incorporate biases and proposes improvements by integrating a utility function based on psychological studies and widely used in the literature to simulate decision-making behaviors. We also expand on the types of non-idealities that are modeled, namely inconsistent decisions and preferentially dependent objectives. Moreover, we formally define and investigate the notion of irrelevant and hidden objectives and propose an approach to simulate them in the experiments. Irrelevant objectives are defined as those that are considered by the optimizer but not by the DM. Irrelevant objectives complement hidden objectives, defined as those that the DM considers but is not the optimizer. Noting that irrelevant objectives add to the complexity of the problem without improving the DM's satisfaction with the results, we propose a feature selection method that can be integrated into any interactive method to automatically filter out irrelevant objectives in an online mode and optimize only important ones. The research is further extended by proposing a new interactive method where the DM's preferences are estimated using decision trees, which naturally perform feature selection. The proposed preference learning technique targets limitations with other methods in the literature. We use our proposed MDM to compare the performance of this novel iEMOA with two state-of-the-art interactive methods. The results suggest that the superior performance of the proposed method is not degraded with the increase of the objective functions and is robust towards noise and non-idealities.
Date of Award1 Aug 2023
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorManuel Lopez-Ibanez (Main Supervisor) & Richard Allmendinger (Supervisor)

Keywords

  • Dimension Reduction
  • Operations Research
  • Support Vector Machines
  • Optimization Algorithms
  • Analysis of Performance
  • Evolutionary Computation
  • Hidden Objectives
  • Sigmoid Utility Functions
  • Design of Experiments
  • Irrelevant Objectives
  • Performance assessment
  • Interactive Evolutionary Multi-Objective Optimization Algorithm
  • Machine Learning
  • Decision Tree
  • Benchmarking
  • Feature Selection
  • Machine Decision-Maker
  • Preference Learning
  • Feature Elimination

Cite this

'