An improved memory prediction strategy for dynamic multiobjective optimization Jinhua Zheng School of Computer Science University of Xiangtan Xiangtan, China e-mail: jhzheng@xtu.edu.cn Tian Chen School of Computer Science University of Xiangtan Xiangtan, China e-mail:1174739397@qq.com Huipeng Xie Shengxiang Yang School of Computer Science School of ComputerScience and Informatic University of Xiangtan De Montfort University Xiangtan, China Leicester, U.K. e-mail:126469584@qq.com e-mail: syang@dmu.ac.uk Abstract—In evolutionary dynamic multiobjective optimization (EDMO), the memory strategy and prediction method are considered as effective and efficient methods. To handling dynamic multiobjective problems (DMOPs), this paper studies the behavior of environment change and tries to make use of the historical information appropriately. And then, this paper proposes an improved memory prediction model that uses the memory strategy to provide valuable information to the prediction model to predict the POS of the new environment more accurately. This memory prediction model is incorporated into a multiobjective evolutionary algorithm based on decomposition (MOEA/D). In particular, the resultant algorithm (MOEA/D-MP) adopts a sensor-based method to detect the environment change and find a similar one in history to reuse the information of it in the prediction process. The proposed algorithm is compared with several state-of-the-art dynamic multiobjective evolutionary algorithms (DMOEA) on six typical benchmark problems with different dynamic characteristics. Experimental results demonstrate that the proposed algorithm can effectively tackle DMOPs. Keywords- dynamic multiobjective optimization; memory strategy; prediction model; I. INTRODUCTION Many real-world problems are dynamic multiobjective optimization problems (DMOPs), with not only the conflict among multiple objectives but also the objective, constraint and related parameters may change over time [1], as well as the decision variables. As a consequence, the Pareto-Optimal Solutions (POS) and/or Pareto-Optimal Front (POF) may vary over time. A minimization problem is considered here without loss of generality. The dynamic multiobjective optimization problem [2] can be described as: 1 min (,) ( ( , ), ( , ),..., ( , )) .. (,) 0 1,2,... ; (,) 0 1, 2,..., T m i j Fxt f xt fxt f xt st g xt i ph xt j q subject to x  =   = = =     (1) where m is the number of objectives, t = 0,1,2... represents discrete time instants, x is the decision variable vector, and Ω is decision space. F(x,t) is the objective vector and consists of m time-varying objective functions that change intermittently. The function of gi ≤ 0 and hj = 0 present the set of inequality and equality constraints. DMOPs have increasingly caused the attention of the research community in recent years. Multiobjective optimization evolutionary algorithms (MOEAs) have been widely used to solve DMOPs [3]-[6]. However, the changes in the POF and/or POS in DMOPs still pose significant challenges to traditional MOEAs. In a dynamic environment, traditional evolutionary algorithm makes the population gradually lose ability to adapt to environmental changes, the reason for this is that the purpose of traditional evolutionary algorithm is to make the population gradually converge to get a satisfactory solution set, but this would make the population lose diversity, especially in the later stages of the evolution [7], which are the challenges of traditional evolutionary algorithm. The difficulty for a multi-objective evolutionary algorithm (MOEA) solving DMOPs is that the algorithm may not re-locate the varied POS and/or POF before the environment changing again[8].How to track the Pareto optimal solution set after the change has always been an important and challenging issue. Dynamic MOEAs(DMOEAs) were further proposed to track a moving POF/POS quickly and obtain PSs that are uniformly distributed over time. II. RELATED WORKS Most of the existing DMOEAs are composed of combining classical MOEAs and effective dynamic handling techniques, including prediction-based [9]-[11], memory-