Evolutionary dynamic multiobjective optimization using a Bayesian vector autoregression prediction model

Dynamic multiobjective optimization problems (DMOPs), inherently involving the simultaneous optimization of conflicting objectives under time-varying environments, exhibit ubiquitous presence in real-world applications spanning diverse engineering disciplines. A prevalent limitation in existing dynamic multiobjective optimization evolutionary algorithms (DMOEAs) lies in inadequate utilization of historical information and neglect of interdependencies among decision variables, which frequently induces suboptimal initial population predictions deviating from true Pareto optimal sets. To mitigate these limitations, we propose MOEA/D-BVAR, a novel DMOEA framework incorporating a Bayesian vector autoregressive (BVAR) model that conceptualizes solution dynamics through holistic vector forecasting rather than isolated variable-specific prediction. The algorithm initially clusters variables via mutual information correlation analysis, subsequently constructing BVAR models for each cluster to project their evolutionary trajectories. Independently varying variables are rapidly predicted through differential forecasting models. A multivariate interaction optimization mechanism enhances search efficiency. Comprehensive empirical evaluations on 14 benchmark suites compare MOEA/D-BVAR against six state-of-the-art DMOEAs developed over the past five years. Statistical analysis of experimental outcomes demonstrates the proposed algorithm’s superior competitiveness in handling complex DMOPs.