Page 32 - 2025S

UEC Int’l Mini-Conference No.54 25

Improving Adaptation in Evolutionary Dynamic Multi-Objective

Optimization via Sparse Solution Selection

Tasin Mobin MOHTADI ∗1 and Hiroyuki SATO 2

1 UEC Exchange Study Program (JUSST Program)
2 Department of Computer and Network Engineering, The University of
Electro-Communications, Tokyo, Japan

Abstract

This paper presents a comprehensive enhancement of the Dynamic Non-dominated Sorting Genetic
Algorithm II (DNSGA-II) for improved Hypervolume (HV) performance in dynamic multi-objective
optimization. We introduce a novel framework featuring adaptive mutation control and diversity-
sensitive reinitialization mechanisms specifically designed to address HV degradation during environ-
mental transitions. Five distinct adaptive variants were developed and rigorously evaluated on FDA
1-FDA 5 benchmark problems across multiple change frequency scenarios (τ t = {10, 25, 50} gener-
ations). Experimental results demonstrate consistent HV improvements up to 18.7% over standard
DNSGA-II, with the hybrid adaptive mutation-reinitialization approach achieving superior perfor-
mance across all test cases. The proposed method features three key innovations: 1) Real-time HV
monitoring for adaptive parameter control, 2) Diversity-guided reinitialization intensity modulation,
and 3) A counter-based change response mechanism. Statistical validation confirms significant im-
provements in both HV quality (p < 0.01) and stability metrics, particularly during rapid environmen-
tal changes. The approach effectively balances exploration-exploitation trade-offs while maintaining
solution diversity across evolving Pareto fronts.

Keywords: dynamic optimization, hypervolume, DNSGA-II, adaptive mutation, reinitialization
strategy, evolutionary computation

1 Introduction tion quality in the new environment.

Traditional multi-objective algorithms
1.1 Problem Context and Challenges
like NSGA-II [1] and its dynamic variants
Conventional methods for dynamic multi- (DNSGA-II) encounter significant difficulties
objective optimization retain solutions after en- in maintaining solution quality during environ-
vironmental changes by preserving a randomly mental transitions. Three primary limitations
persist: First, substantial hypervolume degra-
selected subset comprising 1 − ζ of the popula-
tion. This stochastic approach risks introduc- dation occurs during change periods, reflecting
ing selection bias, as the retained solutions may compromised Pareto front quality. Second,
inadequately represent critical regions of the these algorithms exhibit delayed response
Pareto front or lack necessary diversity. Con- characteristics, resulting in slow convergence
following environmental shifts. Third, diversity
sequently, the optimization efficiency for the
changed problem is compromised due to subop- collapse during stable periods leads to pre-
timal transfer of historical information, poten- mature convergence. These limitations stem
tially delaying convergence and degrading solu- primarily from static re-initialization strate-
gies that fail to adapt to changing problem
∗ Supported by JASSO Scholarship. landscapes.