This paper proposes a novel population initialization algorithm, termed Memory and Prediction-based Population Initialization (MPPI), for addressing dynamic multi-objective optimization problems (DMOPs). MPPI monitors environmental vectors to detect changes in the environment. Upon detecting a change, the algorithm employs three mechanisms to initialize the population: First, it establishes a memory bank based on center points, knee points, and boundary points associated with the environmental vector, and generates a portion of the population based on the memory bank. Second, it predicts a subset of solutions based on feedforward center points. Finally, it generates a number of random solutions to maintain population diversity. Since the memory-based initialization mechanism does not rely on recent solutions, the proposed algorithm is capable of solving dynamic multi-objective optimization problems with dramatically and irregularly changing Pareto optimal sets (POS). The MPPI algorithm is validated on a suite of benchmark functions and a random DTLZ switching experiment. We compare the experimental results against those of other state-of-the-art population initialization methods. The comparison demonstrates that MPPI can quickly track the POS in response to environmental changes, exhibiting strong performance on both periodic and non-periodic problems.
Liu et al. (Thu,) studied this question.