IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 26, NO. 5, OCTOBER 2052 1 Evolutionary Search With Multiview Prediction for Dynamic Multiobjective Optimization Zhang san , Li si Abstract—Dynamic multiobjective optimization problem (DMOP) denotes the multiobjective optimization problem which varies over time. As changes in DMOP may exist some patterns that are predictable, to solve DMOP, a number of research efforts have been made to develop evolutionary search with prediction approaches to estimate the changes of the problem. Index Terms—Dynamic multiobjective optimization, evolutionary search, prediction-based method. I. Introduction W ULTIOBJECTIVE optimization problem (MOP) considers the optimization of at least two conflicting objectives simultaneously [1][2].Unlike single objective optimization, no single solution can satisfy a given MOP, and the optima of a MOP are thus a set of solutions in the decision space, namely, Pareto-optimal solutions (POS)[3][4]. 1) A multiview prediction method is proposed to predict the changes of DMOPs from both the views of decision and objective spaces, which could provide more accurate guidance for evolutionary search in solving DMOP. 2) An efficient kernelized autoencoding model is derived to model the complex change patterns, which holds a closed-form solution and will not bring much computational burden to the prediction. 3) The proposed method is designed to regenerate an initial population for the evolutionary search when dynamic occurs, which can be easily integrated into existing MOEAs for solving DMOPs. II. Preliminaries In this section, we first give the brief introduction of DMOPs. Subsequently, the brief review of existing predictionbased evolutionary algorithms, and the motivation of our proposed evolutionary search with multiview prediction for solving DMOP are presented. A. Dynamic Multiobjective Optimization Problem In this article, we consider the DMOP as a timevariant MOP. In particular, minimization problems are investigated here. Formally, the mathematical definition of a DMOP is formulated as Manuscript received 17 March 2051; revised 20 July 2051 and 22 October 2051; accepted 4 December 2051. Date of publication 14 December 2051; date of current version 3 October 2052.. Algorithm 1 this is a algorithm Require: n ≥ 0 ∨ x ̸= 0 Ensure: y = xn y←1 if n < 0 then X ← 1/x N ← −n else X←x N ←n end if while N ̸= 0 do if N is even then X ←X ×X N ← N/2 else {N is odd} y ←y×X N ←N −1 end if end while TABLE I OVERVIEW OF REPRESENTATIVE PREDICTION-BASED DMOEAS. algorithm FPS PPS KF main idea using a forecasting techniques, predicting the whole popluation using a Kalman Filter-base model remarks 1 2 3 B. Prediction-Based Evolutionary Algorithms for Dynamic Multiobjective Optimization Generally, the workflow of existing prediction approaches can be summarized by Algorithm 1. In contrast to the flow of static MOEAs, change detection and population prediction are two additional components in the prediction-based DMOEAs. In particular, the common procedure of population prediction is as follows. First, a learning model is constructed to learn the change patterns based on past search experience. Next, the learned patterns are used to estimate subsequent change and to predict the initial individuals after the dynamic change. C. Motivation of Conducting Prediction From Multiple Views Real-world and benchmark DMOPs often possess dynamic changes with different patterns in POS and POF IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 26, NO. 5, OCTOBER 2052 2 TABLE II OVERVIEW OF REPRESENTATIVE PREDICTION-BASED DMOEAS. algorithm MDP PBDMO SVR main idea a mutidirectional prediction strategy a reaction mechanism that contains three different strategies constructing a predictor with SVR (i.e., dynamics in decision space and objective space). As mentioned in Section II-A, both POS and POF change over time in Type II DMOPs, and the change patterns may be significantly different. { min f1 (x) = g(x)(x1 + 0.1 sin(3πx1 ))α1 f2 (x) = g(x)(1 − x1 + 0.1 sin(3πx1 ))α2 (1) x = t + cos t + 1 (2) y = 2 sin t (3) III. PROPOSED METHOD In this section, the details of the proposed evolutionary search with multiview prediction for solving DMOPs are presented. In particular, as illustrated in Fig. 2, by employing a conventional MOEA as the basic optimization solver, the search process starts as routine, following the procedures as in the static environment. However, an additional environmental-change detection method is designed to trigger the population prediction, wherein a prediction model is constructed to predict new individuals for the subsequent search in the changed environments. The search process will be iteratively performed until predefined stop criteria are satisfied. To estimate dynamic changes that occur in different spaces of the DMOP, as shown in Fig. 3, we derive a kernelized autoencoding model to conduct the population prediction from the views of both decision and objective spaces. The input and output of this model are the nondominated solutions obtained along the search and the the newly predicted population, respectively. In what follows, the dynamic detection operator employed in the proposed method, the details of the proposed kernelized autoencoding model for multiview prediction, and the pseudocode of the proposed multiview prediction for solving DMOP are presented. A. Dynamic Detection Existing dynamic detection methods can be categorized into two groups, i.e., detector-based detection and behavior-based detection. In particular, the detector-based methods reevaluate some specific solutions (detectors) to detect changes in their function values or feasibility, while the behavior-based approaches consider to assess the behaviors of an algorithm for dynamic detection. Following the robust detection performance achieved in previous DMOEAs, to detect the dynamic changes, here we randomly select 5% individuals as detectors in the remarks 4 5 6 population. In particular, at the start of each generation, the detectors will be reevaluated. If the current objective values of these detectors vary from their stored ones in the last generation, it is recognized that an environmental change occurs. B. Proposed Kernelized Autoencoding Model for Multiview Prediction To learn diverse change patterns in DMOPs from multiple views, in this article, we propose a kernelized autoencoding model to track the moving of high-quality solutions in both decision and objective spaces. In particular, in the context of evolutionary dynamic multiobjective optimization, we denote the optimization problems of a DMOP before and after the dynamic occurs as OP1 and OP2, respectively. Suppose two sets of solutions P =p1,...,pN �Rd×N and Q =q1,...,qN �Rd×N, where N denotes the number of solutions in each set and d is the variable dimension are representative solutions of OP1 and OP2, respectively. By treating OP1 as the corrupted version of OP2, the learning of change pattern from OP1 to OP2 can then be modeled as the denoising autoencoding process by configuring P and Q as the input and output of the autoencoder, respectively, which is given by where M � Rd×d is the learned mapping from OP1 to OP2, tr(A) refers to the trace of the matrix A, and T is the transpose operation of a matrix. denotes the mapped solutions of P in an RKHS H that is transformed by a nonlinear mapping function. When a dynamic change of DMOP has been detected, multiview prediction will be performed using the kernelized autoencoder derived above with the nondominated solutions found along the evolutionary search process. In particular, suppose the dynamic occurs at time instance t+1, we propose to use the nondominated solutions obtained at time instance t and t−1, denoted by N DSt−1 and N DSt , respectively, to predict the new initial population for time instance t+1, from both decision and objective spaces. First, half of the solutions in N DSt−1 S and N DSt is randomly selected and stored in N DSt−1 S and N DSt , respectively. C. Pseudocode of the Proposed Multiview Prediction for DMOP Last but not the least, the combination of the solutions predicted from the views of both decision and objective spaces then composes the initial population IPt+1 as change reaction in DMOEA, thereby to guide the search toward the changed optima. It is straightforward to see IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 26, NO. 5, OCTOBER 2052 that the proposed evolutionary search with multiview prediction is optimizer independent, it thus can be easily integrated into any of population-based static MOEAs for tackling DMOPs. D. Complexity Analysis The computational cost of prediction component comes from learning the mappings via kernelized autoencoding model and solution prediction. The kernelized autoencoder in Algorithm. 2 holds a closed-form solution, which is obtained by solving the system of linear equations as shown in (8), because the computational cost of solving the system is dominated by the calculation of matrix inverse, and the learning complexity of MS k and MF k is O(d2 ∗ N) + O(m2 ∗ N) = O(d2 ∗ N), where N denotes the number of nondominated solutions, d is the variable dimension, and m is the number of objectives. In solution prediction, the interior point algorithm calls for the major time, which costs O(m3 ∗ N). Therefore, the overall complexity of the proposed multiview prediction is O((d2 + m3) ∗ N). IV. EMPIRICAL STUDY In this section, to evaluate the performance of the proposed multiview prediction for solving DMOPs, comprehensive empirical studies on commonly used DMOP benchmarks against four recently proposed DMOEAs are presented. A. Experimental Setup 1) Test Instances: The performance of the proposed prediction approach is investigated on the recently proposed DMOP benchmark for IEEE CEC2018 Competition on Dynamic Multiobjective Optimization. This test suite has 14 benchmark problems, including nine two-objective and five three-objective functions with diverse dynamic characteristics, such as variable linkages, time-dependent POS/POF geometries, irregular POF shapes, disconnectivity, knee, etc. 2) Performance Metrics: In the literature, there are several metrics that have been proposed to evaluate the performance of DMOEAs. In this article, we employ two groups of commonly used metrics, i.e., the inverted generational distance (IGD) metric with two dynamic variants, two dynamic variants of the hypervolume (HV) metric, to objectively measure the performan 3) Compared Algorithms: The proposed evolutionary search with multiview prediction is compared against four state-of-the-art prediction approaches and two baselines with the separate components of our proposed method. In particular, the algorithms for comparison in the experiments are as follows. 3 B. Results and Discussion In this section, comparison results of the proposed multiview prediction against three prediction-based approaches in terms of both optimization quality and efficiency are presented and discussed. Moreover, deeper insights of the interaction between multiple views of prediction are also investigated. 1) Comparison Against Existing Prediction-Based Approaches: The averaged MIGD and MHV values with the standard deviations obtained by the compared algorithms based on MOEA/D optimizer under five dynamic configurations (see Table II), over 20 independent runs on the IEEE CEC2018 DMOP benchmarks, are presented in Tables III and IV. The experimental results obtained by the algorithms based on MOPSO optimizer are presented in Tables I and II in the supplementary material. In these tables, superior performances are highlighted in bold and the Wilcoxon rank sum test at a 5% significance level is performed on the results to indicate the difference between the performances of baseline and the proposed algorithm. “+,” “−,” and “�” denote that the proposed MV-DMOEA is statistically significant better, worse, and similar to the compared algorithm, respectively. In addition, we summarize the results of the significance test with the obtained MIGD and MHV values on 14 DMOP benchmarks under five dynamic configurations (i.e., totally 70 DF instances) at the bottom row of each table. w/t/l indicates that the proposed MV-DMOEA wins on w instances, ties on t instances, and loses on l instances, when compared to the algorithm with the same optimizer, respectively. In the supplementary material, we also calculate the effect size (scientific significance) according to the Vargha–Delaney ˆ A12 statistic. 2) Discussion on Runtime of the Compared Algorithms: It is worth noting that the prediction approaches generally involve additional computational cost to train the prediction model, such as the TCA-based prediction model in and the SVR-based predictor. Therefore, here we further compare and discuss the runtime of all the compared algorithms under the same implementation environment. 3) Comparisons of Prediction From Two Views: To further investigate the deeper insights of the superior performance obtained by the proposed multiview prediction, we plot the initial solutions predicted from the views of both decision and objective space in Fig. 6 and in the supplementary material. Here, we modify the MV-MOEA/D to study prediction from each view separately. Specifically, the kernelized autoencoder model proposed in Section III-B is utilized to conduct prediction from two different views. Fig. 6 depicts the initial solutions predicted by the two independent kernelized predictions on two two-objective and a three-objective DMOPs at three time instances (i.e., IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 26, NO. 5, OCTOBER 2052 t = 15, 20, 25), respectively. As can be observed in these figures, solutions obtained by the prediction from the view of decision space (denoted by red circles in the figures) are more diverse but relatively far from the true POF, while the solutions predicted from the view of objective space (denoted by black circles in the figures) are lack of diversity but more close to the true POF. The comparisons between the prediction from the views of decision space and objective space show that the prediction with learning from multiple views could complement each other in terms of both diversity and convergence. Therefore, it is promising to combine the solutions estimated by the two-view prediction, and to generate an initial population as change reaction for accelerating the evolutionary search toward the moving optima in dynamic environments. 4) Sensitive Study of Parameter p in Multiview Prediction: In our proposed multiview prediction, the size of the training set for each view is equally set, whereas the proportion of either view may affect the performance of the prediction approach. In order to study how multiview prediction is sensitive to the proportion, we set a hyperparameter p to denote the proportion of predicted solutions from the view of decision space in the whole initial population. Specifically, we investigate five different p ranging from 0.1 to 0.9, on four representative benchmarks (i.e., DF1, DF9, DF10, and DF13, all of which possess drastically different change patterns in decision and objective space) with four different configurations (nt = 1, 2.5, 5, 10,�t = 10) are chosen to test MVMOEA/D in multiview prediction. Fig. 7 shows the averaged MIGD values obtained by MV-MOEA/D on these test instances. The results indicate that MV-MOEA/D performs relatively bad with too small or too large value of p (i.e., p = 0.1, 0.9), while the performance of the algorithm is obviously better on these benchmarks when p is close to 0.5. From these results, we can infer that both the prediction in decision space and objective space is important for predicting dynamic changes in solving DMOPs, which again confirmed the motivation of the proposed multiview prediction for solving DMOP. V. REAL-WORLD CASE STUDY ON DYNAMIC MULTIOBJECTIVE RECOMMENDER SYSTEM Recommender systems are essential to cope with information overload in online services and have been widely applied to various fields, such as e-commerce, movies, streaming music, and personalized news reading. Typical multiobjective recommendation aims to learn inherent and long-term user preferences on different properties of item,which are presumed to be static or slow varying throughout the recommendation process. However, the recommendation policy is generally modeled based on the record of user’s historical behaviors, and her/his taste of different items is actually changing over time. Thus, a 4 fixed policy may lead to inappropriate recommendations in the subsequent scenarios of the system. Bearing this in mind, to demonstrate the effectiveness of the proposed evolutionary search with multiview prediction in solving real-world DMOP, in this section, we present a case study with respect to the scenario of movie recommendation and take into account user’s varying preferences on two conflicting objectives of recommended movies. VI. CONCLUSION In this article, an evolutionary search with multiview prediction has been proposed to solve DMOP. In particular, a kernelized autoencoding model has been derived to perform the multiview prediction in an RKHS, which holds a closedform solution and easy-to-solve. In the proposed method, nondominated solutions found along the evolutionary search process have been used to build a prediction model from the views of both decision and objective spaces to generate an initial population, which serves as the dynamic response that can be integrated into any population-based multiobjective optimization algorithms for solving DMOPs. To evaluate the performance of the proposed multiview prediction, comprehensive empirical studies have been conducted on the IEEE CEC2018 DMOP benchmarks under five different dynamic configurations. The results obtained by comparing against four state-of-the-art prediction approaches and two baselines with the separate components of multiview prediction showed the efficacy of the proposed method for evolutionary dynamic multiobjective optimization. To further demonstrate the application value of the proposed method, a real-world DMOP of movie recommendation is formulated, and two traditional recommendation algorithms and two DMOEAs are compared with the proposed multiview prediction on the DMOP of movie recommendation. The results of prediction accuracy verified that the multiview prediction is able to solve real-world DMOPs. The proposed multiview prediction has shown promising performance on various benchmark problems and a practical problem; however, there are still several issues need to be addressed to further improve the multiview prediction for solving DMOP. First, the correlations between different views and how the dynamics in different spaces would affect the changes of a DMOP are keen to the design of multiview prediction for solving DMOP. Moreover, the definition of knowledge that is used for prediction when dynamic occurs may further affect the effectiveness and efficiency of multiview prediction in DMOP. Finally, constrained problems are commonly encountered in nature, and we did not make specific designs for constraints except the boundary check in the prediction. Therefore, for future work, we would like to consider the design of the dynamic detection method to adaptively determine the characteristics of dynamics, and suggest the appropriate interaction strategy for the subsequent multiview prediction. In addition, we would like to further explore dynamic multiobjective optimization with constraints considered in the prediction. IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION, VOL. 26, NO. 5, OCTOBER 2052 References [1] C. M. Fonseca and P. J. Fleming, “Genetic algorithms for multiobjective optimization: Formulationdiscussion and generalization,” in Proc. 5th Int. Conf. Genet. Algorithms, vol. 93, Jul. 1993, pp. 416–423. [2] R. T. Marler and J. S. Arora, “Survey of multi-objective optimization methods for engineering,” Struct. Multidiscipl. Optim., vol. 26, no. 6, pp. 369–395, 2004. [3] K. Deb, Multi-Objective Optimization Using Evolutionary Algorithms, vol. 16. Chichester, U.K.: Wiley, 2001. [4] C. A. Coello Coello, “Evolutionary multi-objective optimization: A historical view of the field,” IEEE Comput. Intell. Mag., vol. 1, no. 1, pp. 28–36, Feb. 2006. [5] AMS-StyleGuide-online.pdf, published by the American Mathematical Society [6] H. Sira-Ramirez. “On the sliding mode control of nonlinear systems,” Systems & Control Letters, vol. 19, pp. 303–312, 1992. [7] A. Levant. “Exact differentiation of signals with unbounded higher derivatives,” in Proceedings of the 45th IEEE Conference on Decision and Control, San Diego, California, USA, pp. 5585– 5590, 2006. 5 [8] M. Fliess, C. Join, and H. Sira-Ramirez. “Non-linear estimation is easy,” International Journal of Modelling, Identification and Control, vol. 4, no. 1, pp. 12–27, 2008. [9] R. Ortega, A. Astolfi, G. Bastin, and H. Rodriguez. “Stabilization of food-chain systems using a port-controlled Hamiltonian description,” in Proceedings of the American Control Conference, Chicago, Illinois, USA, pp. 2245–2249, 2000. Zhang san received the B.E. degree from the Department of Computer Science, Anhui University, Hefei, China, in 2048, where he is currently pursuing the Ph.D.degree.