Training-Free Diffusion-Driven Modeling of Pareto Set Evolution for Dynamic Multiobjective Optimization

Jian Guan; Huolong Wu; Zhenzhong Wang; Gary G. Yen; Min Jiang

cs.NEMar 23, 2026

Training-Free Diffusion-Driven Modeling of Pareto Set Evolution for Dynamic Multiobjective Optimization

Authors: Jian Guan, Huolong Wu, Zhenzhong Wang, Gary G. Yen, Min Jiang

Organizations: Institute of Artificial Intelligence, Xiamen University, Xiamen, 361005, China · Department of Artificial Intelligence, School of Informatics, Xiamen University; the Key Laboratory of Multimedia Trusted Perception and Efficient Computing, Ministry of Education; and the Key Laboratory of Digital Protection and Intelligent Processing of Intangible Cultural Heritage of Fujian and Taiwan, Ministry of Culture and Tourism, Xiamen, 361005, China · Department of Artificial Intelligence, Sichuan University, Chengdu, 610065, China

Abstract

Dynamic multiobjective optimization problems (DMOPs) feature time-varying objectives, which cause the Pareto optimal solution (POS) set to drift over time and make it difficult to maintain both convergence and diversity under limited response time. Many existing prediction-based dynamic multiobjective evolutionary algorithms (DMOEAs) either depend on learned models with nontrivial training cost or employ one-step population mapping, which may overlook the gradual nature of POS evolution. This paper proposes DD-DMOEA, a training-free diffusion-based dynamic response mechanism for DMOPs. The key idea is to treat the POS obtained in the previous environment as a "noisy" sample set and to guide its evolution toward the current POS through an analytically constructed multi-step denoising process. A knee-point-based auxiliary strategy is used to specify the target region in the new environment, and an explicit probability-density formulation is derived to compute the denoising update without neural training. To reduce the risk of misleading guidance caused by knee-point prediction errors, an uncertainty-aware scheme adaptively adjusts the guidance strength according to the historical prediction deviation. Experiments on the CEC2018 dynamic multiobjective benchmarks show that DD-DMOEA achieves competitive or better convergence-diversity performance and provides faster dynamic response than several state-of-the-art DMOEAs.

Figures & tables

Fig. 1: Two evolutionary modes of the POS in dynamic environments . (a) The single-step, direct mapping employed by existing prediction methods. (b) The gradual and smooth evolutionary process of the POF in the real world. (c) Correspondingly, in the decision space, the evolution of the POS exhibits a continuous manifold pattern. The red dashed box ( $\Delta t$ ) indicates the key intermediate transitional states often ignored by the methods in (a).

Fig. 2: Overview of the proposed DD-DMOEA. (a) The AKP strategy partitions the objective space and predicts the positions of the knee points ( $k^{t}_{i}$ ) at the moment $t$ . This prediction assists in generating an approximate target POS distribution based on $POS_{t-1}$ . (b) The DDM treats the $POS_{t-1}$ as a noisy distribution, and a training-free diffusion model executes an efficient and progressive denoising process. (c) As the environment changes ( $t$ -1 $\to$ $t$ ), this denoising process gradually transforms the “noisy” old population into a refined one adapted to the new environment, enabling rapid tracking of the POF’s evolution ( $POF_{t-1}\to POF_{t}$ ).

Fig. 3: Illustration of dividing objection space by the knee point. The number of subspaces here is 3.

Prob.	( $n_{t},\tau_{t}$ )	KT-DMOEA	IGP-DMOEA	DIP-DMOEA	KTM-DMOEA	PSGN	VARE	DM-DMOEA	DD-DMOEA
DF1	(10, 10)	0.0402±9.53e-03(+)	0.0248±3.59e-03(+)	0.0264±2.27e-03(+)	0.0242±2.03e-03(+)	0.0497±5.53e-03(+)	0.0360±5.01e-03(+)	0.0304±3.55e-03(+)	0.0227±6.70e-04
	(5, 10)	0.0549±1.18e-02(+)	0.0281±4.23e-03(-)	0.0400±3.18e-03(+)	0.0363±3.15e-03(+)	0.0496±5.64e-03(+)	0.0395±5.31e-03(+)	0.0343±3.79e-03(=)	0.0348±5.90e-04
	(10, 5)	0.1241±2.96e-02(+)	0.0704±8.84e-03(+)	0.0665±7.10e-03(+)	0.0635±7.65e-03(+)	0.0687±6.94e-03(+)	0.1285±1.16e-02(+)	0.1155±6.37e-03(+)	0.0544±2.65e-03
	(5, 5)	0.1609±3.40e-02(+)	0.0849±1.19e-02(+)	0.1086±8.80e-03(+)	0.0817±5.78e-03(=)	0.0840±9.32e-03(+)	0.1361±1.45e-02(+)	0.1246±1.13e-02(+)	0.0820±3.06e-03
DF2	(10, 10)	0.0383±7.69e-03(+)	0.0275±4.63e-03(+)	0.0291±3.28e-03(+)	0.0275±2.69e-03(+)	0.0455±4.93e-03(+)	0.0385±3.45e-03(+)	0.0338±5.20e-03(+)	0.0227±8.83e-04
	(5, 10)	0.0465±1.12e-02(+)	0.0320±5.41e-03(+)	0.0413±4.82e-03(+)	0.0365±3.89e-03(+)	0.0470±5.54e-03(+)	0.0402±4.12e-03(+)	0.0368±6.13e-03(+)	0.0318±1.32e-03

TABLE I: Mean and standard deviation of MIGD obtained by the comparison algorithms in four dynamic environments.

Fig. 4: IGD curves for all algorithms on 14 test functions on dynamic environments ( $n_{t}=10$ and $\tau_{t}=10$ ).

Fig. 5: Average running time of all algorithms on some test functions on dynamic environments ( $n_{t}=10$ and $\tau_{t}=10$ ).

Fig. 6: MIGD values generated by different noise schedules $\alpha$ for all test functions in four dynamic environments. For example, DDM/L( $\alpha$ )-1 and DDM/C( $\alpha$ )-1 denote the linear and cosine schedules on the dynamic environment ( $n_{t}=10$ and $\tau_{t}=10$ ).

Fig. 7: Average running time and MIGD value for six test functions at different time steps on dynamic environments ( $n_{t}=10$ and $\tau_{t}=5$ ).

Algs.	$\text{ENV}_{1}$	$\text{ENV}_{2}$	$\text{ENV}_{3}$	$\text{ENV}_{4}$	$\text{ENV}_{5}$
KT-DMOEA	2.29e-01±(2.11e-02)	2.14e-01±(9.86e-02)	1.83e-01±(9.29e-02)	1.79e-01±(6.59e-02)	2.14e-01±(5.41e-02)
IGP-DMOEA	2.02e-01±(9.95e-02)	2.20e-01±(5.50e-02)	1.81e-01±(7.00e-02)	2.03e-01±(4.13e-02)	1.93e-01±(5.41e-02)
DIP-DMOEA	2.05e-01±(4.32e-02)	2.02e-01±(3.76e-02)	1.98e-01±(2.47e-02)	1.85e-01±(2.07e-02)	2.01e-01±(2.54e-02)
KTM-DMOEA	1.99e-01±(4.22e-02)	2.00e-01±(3.25e-02)	1.88e-01±(3.31e-02)	2.02e-01±(3.46e-02)	2.03e-01±(2.76e-02)
PSGN	2.57e-01±(1.07e-01)	2.39e-01±(1.05e-01)	1.89e-01±(1.05e-01)	2.31e-01±(6.41e-02)	2.39e-01±(6.58e-02)
VARE	1.88e-01±(5.47e-02)	2.06e-01±(7.38e-02)	1.86e-01±(7.19e-02)	2.16e-01±(5.80e-02)	1.88e-01±(3.27e-02)

TABLE II: Mean and standard deviation of HV obtained by the comparison algorithm for five environmental changes in the ROA problem.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Problem	objectives	Dynamics	Remarks
DF1	2	mixed convexity-concavity,location of optima	dynamic PF and PS
DF2	2	switch of position-related variable, location of optima	static convex PF, dynamic PS, severe diversity loss
DF3	2	mixed convexity-concavity,variable-linkage, location of optima	dynamic PF and PS
DF4	2	variable-linkage, PF range, bounds of PS	dynamic PF and PS
DF5	2	number of knee regions, local of optima	dynamic PF and PS
DF6	2	mixed convexity-concavity, multimodality, location of optima	dynamic PF and PS

Appendix

TABLE S-I : Dynamic characteristics of 14 test functions in the CEC2018 dynamic multi-objective optimization competition.

Prob.	( $n_{t},\tau_{t}$ )	KT-DMOEA	IGP-DMOEA	DIP-DMOEA	KTM-DMOEA	PSGN	VARE	DM-DMOEA	DD-DMOEA
DF1	(10, 10)	0.5854±1.19e-02(+)	0.6079±6.27e-03(+)	0.6038±4.04e-03(+)	0.6089±3.67e-03(+)	0.4666±5.34e-03(+)	0.5783±7.94e-03(+)	0.6053±6.33e-03(+)	0.6096±1.29e-03
	(5, 10)	0.5605±1.52e-02(+)	0.5988±7.57e-03(=)	0.5854±6.22e-03(+)	0.5956±6.00e-03(+)	0.4689±5.97e-03(+)	0.5767±8.36e-03(+)	0.5973±7.54e-03(=)	0.5971±1.46e-03
	(10, 5)	0.4706±2.47e-02(+)	0.5366±1.07e-02(+)	0.5390±1.20e-02(+)	0.5457±1.10e-02(+)	0.4359±7.94e-03(+)	0.4469±1.42e-02(+)	0.5454±1.04e-02(+)	0.5547±3.89e-03
	(5, 5)	0.4358±2.95e-02(+)	0.5126±1.46e-02(=)	0.4731±1.33e-02(+)	0.5149±8.83e-03(=)	0.4220±1.13e-02(+)	0.4418±1.52e-02(+)	0.5204±1.55e-02(=)	0.5103±3.14e-03
DF2	(10, 10)	0.8125±1.10e-02(+)	0.8291±8.00e-03(+)	0.8245±5.74e-03(+)	0.8312±4.56e-03(+)	0.6634±6.50e-03(+)	0.8182±3.99e-03(+)	0.8368±8.88e-03(=)	0.8357±1.14e-03
	(5, 10)	0.8016±1.64e-02(+)	0.8214±1.01e-02(=)	0.7992±8.82e-03(+)	0.8127±6.28e-03(+)	0.6638±6.10e-03(+)	0.8134±5.73e-03(+)	0.8222±1.10e-02(=)	0.8184±2.17e-03

Appendix

TABLE S-II : Mean and standard deviation of MHV obtained by the comparison algorithms in four dynamic environments.

Fig. S-1 : IGD curves for all algorithms on 14 test functions on dynamic environments ( $n_{t}=5$ and $\tau_{t}=10$ ).

Fig. S-2 : IGD curves for all algorithms on 14 test functions on dynamic environments ( $n_{t}=10$ and $\tau_{t}=5$ ).

Fig. S-3 : IGD curves for all algorithms on 14 test functions on dynamic environments ( $n_{t}=5$ and $\tau_{t}=5$ ).

Fig. S-4 : Average running time of all algorithms on 14 test functions on dynamic environments ( $n_{t}=10$ and $\tau_{t}=10$ ).

Fig. S-5 : Average running time of all algorithms on 14 test functions on dynamic environments ( $n_{t}=10$ and $\tau_{t}=5$ ).

Fig. S-6 : Average running time of all algorithms on 14 test functions on dynamic environments ( $n_{t}=5$ and $\tau_{t}=10$ ).

Fig. S-7 : Average running time of all algorithms on 14 test functions on dynamic environments ( $n_{t}=5$ and $\tau_{t}=5$ ).

Prob.	$(n_{t},\tau_{t})$	DD-DMOEA/ $\text{v}_{1}$	DD-DMOEA $/\text{v}_{2}$	DD-DMOEA $/\text{v}_{3}$	DD-DMOEA
DF1	(10, 10)	0.0251±2.92e-03(+)	0.0394±4.01e-03(+)	0.0312±2.17e-03(+)	0.0235±2.09e-03
	(5, 10)	0.0374±3.86e-03(+)	0.0472±5.75e-03(+)	0.0404±3.76e-03(+)	0.0356±3.91e-03
	(10, 5)	0.0571±5.50e-03(=)	0.1338±1.39e-02(+)	0.0910±4.30e-03(+)	0.0576±5.87e-03
	(5, 5)	0.0934±7.35e-03(+)	0.1372±1.30e-02(+)	0.1130±1.14e-02(+)	0.0872±9.67e-03
DF2	(10, 10)	0.0264±4.45e-03(+)	0.0388±6.20e-03(+)	0.0253±4.03e-03(+)	0.0252±4.58e-03
	(5, 10)	0.0346±4.99e-03(+)	0.0392±2.89e-03(+)	0.0298±2.91e-03(-)	0.0327±5.70e-03

Appendix

TABLE S-III : Mean and standard deviation of MIGD values of all test functions for different strategies on four dynamic environments.

Prob.	$(n_{t},\tau_{t})$	DD-DMOEA/ $\text{v}_{1}$	DD-DMOEA $/\text{v}_{2}$	DD-DMOEA $/\text{v}_{3}$	DD-DMOEA
DF1	(10, 10)	0.6057±5.12e-03(+)	0.5800±6.93e-03(+)	0.5945±3.93e-03(+)	0.6081±3.66e-03
	(5, 10)	0.5904±7.15e-03(+)	0.5722±9.67e-03(+)	0.5829±8.10e-03(+)	0.5936±8.81e-03
	(10, 5)	0.5501±8.02e-03(=)	0.4359±1.69e-02(+)	0.4934±5.33e-03(+)	0.5486±8.66e-03
	(5, 5)	0.4953±1.25e-02(+)	0.4285±1.52e-02(+)	0.4607±1.75e-02(+)	0.5040±1.43e-02
DF2	(10, 10)	0.8302±6.77e-03(+)	0.8067±1.08e-02(+)	0.8343±4.35e-03(=)	0.8316±7.90e-03
	(5, 10)	0.8128±1.02e-02(+)	0.8070±5.20e-03(+)	0.8233±5.19e-03(=)	0.8159±1.08e-02

Appendix

TABLE S-IV : Mean and standard deviation of MHV values of all test functions for different strategies on four dynamic environments.

Prob.	$(n_{t},\tau_{t})$	DDM/L( $\alpha$ )-1	DDM/ $K_{1}$	DDM/ $K_{2}$	DDM/ $K_{3}$	DDM/ $K_{4}$	DDM/ $K_{5}$	DD-DMOEA
DF1	(10, 10)	0.0257±2.33e-03(+)	0.0251±2.01e-03(+)	0.0245±2.65e-03(+)	0.0249±2.28e-03(+)	0.0241±2.02e-03(+)	0.0254±2.75e-03(+)	0.0235±2.09e-03
	(5, 10)	0.0395±4.38e-03(+)	0.0419±4.53e-03(+)	0.0362±3.26e-03(+)	0.0363±3.55e-03(+)	0.0362±2.91e-03(+)	0.0369±5.08e-03(+)	0.0356±3.91e-03
	(10, 5)	0.0662±4.51e-03(+)	0.0619±6.57e-03(+)	0.0576±4.24e-03(=)	0.0598±6.94e-03(+)	0.0580±5.04e-03(+)	0.0603±3.99e-03(+)	0.0576±5.87e-03
	(5, 5)	0.0872±7.04e-03(=)	0.0944±6.28e-03(+)	0.0868±7.26e-03(=)	0.0888±8.88e-03(+)	0.0846±7.99e-03(=)	0.0812±6.30e-03(-)	0.0872±9.67e-03
DF2	(10, 10)	0.0267±3.26e-03(+)	0.0263±3.41e-03(+)	0.0256±2.87e-03(+)	0.0252±3.09e-03(=)	0.0249±4.55e-03(=)	0.0259±4.54e-03(+)	0.0252±4.58e-03
	(5, 10)	0.0375±4.59e-03(+)	0.0355±4.04e-03(+)	0.0351±4.97e-03(+)	0.0341±4.30e-03(+)	0.0351±4.92e-03(+)	0.0360±4.87e-03(+)	0.0327±5.70e-03

Appendix

TABLE S-V : Mean and standard deviations of MIGD values obtained from all test functions in four dynamic environments by different time steps

K

and noise scheduling

\alpha

Prob.	$(n_{t},\tau_{t})$	DDM/L( $\alpha$ )-1	DDM/ $K_{1}$	DDM/ $K_{2}$	DDM/ $K_{3}$	DDM/ $K_{4}$	DDM/ $K_{5}$	DD-DMOEA
DF1	(10, 10)	0.6035±4.29e-03(+)	0.6055±3.76e-03(+)	0.6069±4.15e-03(+)	0.6057±3.84e-03(+)	0.6069±3.89e-03(+)	0.6044±4.67e-03(+)	0.6081±3.61e-03
	(5, 10)	0.5849±7.85e-03(+)	0.5817±8.11e-03(+)	0.5921±6.73e-03(+)	0.5926±6.42e-03(+)	0.5925±5.09e-03(+)	0.5919±8.78e-03(+)	0.5936±8.21e-03
	(10, 5)	0.5345±5.43e-03(+)	0.5413±1.09e-02(+)	0.5512±5.77e-03(=)	0.5462±9.73e-03(+)	0.5488±7.76e-03(=)	0.5449±6.02e-03(+)	0.5484±8.81e-03
	(5, 5)	0.5022±1.16e-02(+)	0.4921±9.55e-03(+)	0.5046±1.07e-02(=)	0.5046±1.17e-02(=)	0.5104±1.23e-02(=)	0.5160±1.00e-02(=)	0.5040±8.76e-03
DF2	(10, 10)	0.8292±6.01e-03(+)	0.8306±5.78e-03(+)	0.8317±5.83e-03(=)	0.8313±5.19e-03(+)	0.8309±7.81e-03(+)	0.8307±7.20e-03(+)	0.8316±7.79e-03
	(5, 10)	0.8080±9.36e-03(+)	0.8115±8.34e-03(+)	0.8109±9.33e-03(+)	0.8124±9.09e-03(+)	0.8118±9.85e-03(+)	0.8102±9.12e-03(+)	0.8159±1.40e-02

Appendix

TABLE S-VI : Mean and standard deviations of MHV values obtained from all test functions in four dynamic environments by different time steps

K

and noise scheduling

\alpha

Prob.	$(n_{t},\tau_{t})$	DDM/ $\psi_{1}$	DDM/ $\psi_{2}$	DDM/ $\psi_{3}$	DDM/ $\psi_{4}$	DDM/ $\psi_{5}$	DDM/ $\psi_{6}$	DD-DMOEA
DF1	(10, 10)	0.0267±4.13e-03(+)	0.0247±3.89e-03(+)	0.0227±4.13e-03(=)	0.0286±4.26e-03(+)	0.0241±3.11e-03(+)	0.0247±3.42e-03(+)	0.0230±2.36e-03
	(5, 10)	0.0364±4.53e-03(+)	0.0359±5.33e-03(+)	0.0321±4.24e-03(-)	0.0403±6.02e-03(+)	0.0389±4.87e-03(+)	0.0399±5.97e-03(+)	0.0349±3.34e-03
	(10, 5)	0.0786±1.26e-02(+)	0.0727±1.07e-02(+)	0.0644±9.43e-03(+)	0.0826±6.54e-03(+)	0.0669±7.74e-03(+)	0.0646±5.91e-03(+)	0.0566±6.05e-03
	(5, 5)	0.1047±1.21e-02(+)	0.1035±1.02e-02(+)	0.0913±1.15e-02(+)	0.1131±1.57e-02(+)	0.1041±8.69e-03(+)	0.1042±1.18e-02(+)	0.0842±6.18e-03
DF2	(10, 10)	0.0380±6.97e-03(+)	0.0340±6.48e-03(+)	0.0291±5.23e-03(+)	0.0359±4.59e-03(+)	0.0336±4.80e-03(+)	0.0333±3.44e-03(+)	0.0243±4.27e-03
	(5, 10)	0.0391±4.75e-03(+)	0.0385±5.81e-03(+)	0.0363±6.82e-03(+)	0.0452±6.32e-03(+)	0.0457±5.42e-03(+)	0.0470±5.96e-03(+)	0.0317±7.79e-03

Appendix

TABLE S-VII : Mean and standard deviations of MIGD values obtained from all test functions in four dynamic environments by the standard deviation

\psi

of different Gaussian distributions.

Prob.	$(n_{t},\tau_{t})$	DDM/ $\psi_{1}$	DDM/ $\psi_{2}$	DDM/ $\psi_{3}$	DDM/ $\psi_{4}$	DDM/ $\psi_{5}$	DDM/ $\psi_{5}$	DD-DMOEA
DF1	(10, 10)	0.5985±7.46e-03(+)	0.5999±5.74e-03(+)	0.6041±6.85e-03(+)	0.6021±6.81e-03(+)	0.6057±5.21e-03(+)	0.6047±5.41e-03(+)	0.6088±3.61e-03
	(5, 10)	0.5841±1.04e-02(+)	0.5848±8.61e-03(+)	0.5917±7.71e-03(+)	0.5889±7.37e-03(+)	0.5862±8.81e-03(+)	0.5864±1.02e-02(+)	0.5963±8.21e-03
	(10, 5)	0.5149±1.01e-02(+)	0.5271±1.65e-02(+)	0.5407±1.35e-02(+)	0.5332±1.84e-02(+)	0.5428±1.17e-02(+)	0.5436±7.73e-03(+)	0.5511±8.81e-03
	(5, 5)	0.4886±1.89e-02(+)	0.4857±1.36e-02(+)	0.5016±1.56e-02(+)	0.4940±1.46e-02(+)	0.4907±1.29e-02(+)	0.4945±1.67e-02(+)	0.5075±8.76e-03
DF2	(10, 10)	0.8263±7.41e-03(+)	0.8251±8.75e-03(+)	0.8301±8.09e-03(+)	0.8284±9.65e-03(+)	0.8273±8.74e-03(+)	0.8271±5.77e-03(+)	0.8329±7.79e-03
	(5, 10)	0.8094±1.15e-02(+)	0.8109±1.14e-02(+)	0.8145±1.11e-02(+)	0.8145±9.36e-03(+)	0.8065±1.01e-02(+)	0.8035±1.25e-02(+)	0.8178±1.40e-02

Appendix

TABLE S-VIII : Mean and standard deviations of MHV values obtained from all test functions in four dynamic environments by the standard deviation

\psi

of different Gaussian distributions.

Multi-Objective Optimization

Multi-objective optimization studies how to find and choose solutions that balance competing goals, constraints, and preferences.

See all 265 papers

Apr 1, 2026

A Decoupled Basis-Vector-Driven Generative Framework for Dynamic Multi-Objective Optimization

Dynamic multi-objective optimization requires continuous tracking of moving Pareto fronts. Existing methods struggle with irregular mutations and data sparsity, primarily facing three challenges: the non-linear coupling of dynamic modes, negative transfer from outdated historical data, and the cold-start problem during environmental switches. To address these issues, this paper proposes a decoupled basis-vector-driven generative framework (DB-GEN). First, to resolve non-linear coupling, the framework employs the discrete wavelet transform to separate evolutionary trajectories into low-frequency trends and high-frequency details. Second, to mitigate negative transfer, it learns transferable basis vectors via sparse dictionary learning rather than directly memorizing historical instances. Recomposing these bases under a topology-aware contrastive constraint constructs a structured latent manifold. Finally, to overcome the cold-start problem, a surrogate-assisted search paradigm samples initial populations from this manifold. Pre-trained on 120 million solutions, DB-GEN performs direct online inference without retraining or fine-tuning. This zero-shot generation process executes in milliseconds, requiring approximately 0.2 seconds per environmental change. Experimental results demonstrate that DB-GEN improves tracking accuracy across various dynamic benchmarks compared to existing algorithms.

Yaoming Yang, Shuai Wang, Bingdong Li +2

Multi-Party Multi-Objective Optimization Multi-Objective Optimization Multi-Objective Evolutionary Optimization

Ask PDF

Aug 1, 2026

SDDMO-Bench: A Benchmark Suite for Streaming Data-Driven Dynamic Multi-Objective Optimization

Streaming data-driven dynamic multi-objective optimization requires algorithms to track time-varying Pareto fronts using only sequential observations under concept drift. However, systematic evaluation remains difficult because real-world problems usually lack ground-truth optima, drift annotations, and controllable conditions, while existing benchmarks provide limited support for standardized comparison. This paper proposes SDDMO-Bench, a benchmark suite that transforms classical dynamic multi-objective test problems into streaming environments by combining intrinsic objective-mapping evolution, controllable distributional drift, and sequential data revelation. By combining five representative time-dependent base functions with six distributional drift patterns, SDDMO-Bench constructs 30 scenarios with diverse levels of non-stationarity, problem complexity, sample-distribution variation, and Pareto-front evolution. Experiments with representative evolutionary algorithms demonstrate that SDDMO-Bench provides challenging and discriminative test scenarios, offering a standardized, controllable, and reproducible benchmark for evaluating adaptability, robustness, and Pareto-front tracking in streaming data-driven dynamic multi-objective optimization.

Wenjie Xiao, Hui Bai, Junhao Chen

School of Xiangtan University Xiangtan, China

Multi-Party Multi-Objective Optimization Benchmark Design Multi-Objective Evolutionary Optimization Synthetic Benchmark Generation Multi-Objective Optimization

Ask PDF

Oct 26, 2026

A Dynamic Decomposition Strategy for the MOEA/D With Proven Performance Guarantees

The MOEA/D is one of the most successful multi-objective evolutionary algorithms. It operates by decomposing the optimization problem into single-objective subproblems, which are then solved in a co-evolutionary manner. The performance of the MOEA/D depends critically on this decomposition. In this work, we propose a novel way to adjust this decomposition on the fly, for the case of discrete bi-objective optimization. Different from all previous works in this direction, we support our new algorithm with a proven performance guarantee. This is the first work that rigorously proves the runtime of a dynamic multi-objective evolutionary algorithm. For monotonic distortions of the simple LOTZ benchmark, it is known that classic decompositions leave large holes in the Pareto front, making it very hard for the MOEA/D to cover the entire Pareto front. In contrast, we prove that our self-adjusting MOEA/D solves all these LOTZ problems in an expected number of

O(n^4)

function evaluations. This relatively low runtime as well as the details of our mathematical proof show that the dynamic decomposition strategy indeed evolves a decomposition adjusted to each of these problem instances.

Benjamin Doerr, Martin S. Krejca, Noé Weeks

Laboratoire d’Informatique (LIX), CNRS, École Polytechnique, Institut Polytechnique de Paris, Palaiseau, France · École Normale Supérieure, Paris, France

Multi-Objective Evolutionary Optimization Multi-Objective Optimization Runtime Analysis of Evolutionary Algorithms

Ask PDF

Abstract

Multi-Objective Optimization

Explore similar work

A Decoupled Basis-Vector-Driven Generative Framework for Dynamic Multi-Objective Optimization

SDDMO-Bench: A Benchmark Suite for Streaming Data-Driven Dynamic Multi-Objective Optimization

A Dynamic Decomposition Strategy for the MOEA/D With Proven Performance Guarantees