Integrating hybrid structures and reinforcement learning into multi-objective evolutionary algorithms: enhancing efficiency and adaptability

Abstract

This thesis addresses critical challenges in multi-objective optimization by integrating hybrid structures and reinforcement learning (RL) with multi-objective evolutionary algorithms (MOEAs).A key innovation is the use of RL to automate the selection of evolutionary algorithm (EA) operators, significantly enhancing the adaptability and efficiency of MOEAs. Furthermore, the hybridization of EA operators with MOEA structures markedly improves optimization efficacy.

The research introduces three novel algorithms: the Hybrid Multi-objective Teaching Learning based Optimization Algorithm based on R2 Indicator (R2-HMTLBO), the Hybrid Multi-objective Equilibrium Optimizer and Whale Optimization Algorithm based on R2 Indicator (R2-HMEWO),and the Reinforcement Learning-based Multi-objective Evolutionary Algorithm based on R2 Indicator (R2-RLMOEA). These methodologies are assessed using two metrics—Inverted Generational Distance (IGD) and Spacing (SP)—to evaluate the convergence and distribution of the solutions. The algorithms are also compared with existing methods to highlight their advancements.

Specifically, R2-HMTLBO integrates the Nondominated Sorting Genetic Algorithm-III (NSGA-III)with Teaching Learning-based Optimization (TLBO) and Genetic Algorithm (GA) operators, incorporating the R2 indicator to enhance diversity and convergence. R2-HMEWO combines the Whale Optimization Algorithm (WOA) and Equilibrium Optimizer (EO) with elite archive strategies, demonstrating improved bi-objective optimization performance. R2-RLMOEA employs reinforcement learning to adaptively select the most effective strategies, enhancing performance across various benchmarks.

The algorithms address existing limitations of MOEAs and provide insights into the integration of diverse optimization mechanisms. Experimental analysis validates the superiority of these hybridizations over traditional approaches, showcasing significant performance improvements inmost multi-objective benchmarks. The thesis also explores theoretical implications and future research avenues, particularly in balancing exploration and exploitation in dynamic environments. It highlights strategies to enhance convergence speed and solution quality diversity

Date of Award	Sept 2024
Original language	English
Supervisor	Debbie Rankin (Supervisor) & Damien Coyle (Supervisor)