Trend analysis-based prediction strategies for dynamic multi-objective evolutionary optimization

Abstract

Dynamic multiobjective optimization problems (DMOPs) change over time, which require Evolutionary algorithms (EA) to track Pareto-optimal solutions (PS) and/or Pareto-optimal front (PF) in a dynamic environment. Most prediction-based algorithms solely use a single model to learn the changing pattern for solution prediction. In the face of complex DMOPs, they may achieve an unsatisfactory performance. To address this issue, a novel trend analysis-based prediction strategy (TAP) is proposed in this paper. Based on previous population information, a simple trend analysis is designed to extract the changing pattern of each solution, and classify them into different types: irregular, translational, and stationary. For irregular changing solutions, a neural network nonlinear model is presented to predict the new location. For translational changing solutions, a simple linear model is built to estimate their new positions. For stationary solutions, they are preserved. As a result, TAP is more responsive to different dynamic environments. TAP is incorporated into the dynamic multiobjective evolutionary algorithm (DMOEA) based on decomposition (MOEA/D) to construct a novel algorithm denoted as MOEA/D-TAP. To verify the performance of the proposed method, comparison experiments are carried out on 26 test instances of four different benchmarks compared with six state-of-the-art methods. The test results indicate that TAP is highly competitive.