Neural Network for Change Direction Prediction in Dynamic Optimization

Abstract

In dynamic optimization problems (DOPs), the environment changes with time, often accompanied by the movement of the optima. Once the environment changes, how to find the new optimum as soon as possible is a challenging issue. An important question to ask is, “Are the past solutions useful for the optimization in new environments?” Since the successive environments are usually correlative and the change of the optimum may obey a stable law, if we can capture the change law from past information, then we can predict the new optimum of new environments. Furthermore, in complex problems (e.g., multimodal problems), solutions in different subareas often obey different change laws, and the local optimum of a subarea, which is poor in past environments, may become the global optimum in new environments. Thus, this paper proposes a neural network (NN)-based change prediction method, named NN-based change prediction method (NNCP), to discover the change law of the optima in different subareas and predict new optima. In the proposed NNCP, the search space is decomposed into multiple subareas and the change law of the local optimum in each subarea is extracted from the solutions found in past environments by NNs. Then, the NNs reuse the past solutions to predict new solutions. Based on the proposed NNCP, most of existing evolutionary algorithms (EAs) for DOPs can reuse the past information to predict new optima. To verify the effect of NNCP, we incorporate the proposed method into the five typical state-of-the-art EAs for comprehensive study. They are evaluated on the widely used moving peaks benchmark with two famous performance measures. Experimental results show that the proposed NNCP significantly enhances EAs’ performance in solving DOPs.