Prediction Strategy Assisted by LSTM Neural Network for Dynamic Multi-Objective Optimization

Dynamic multi-objective optimization problems (DMOPs) exist widely in real life, and the algorithm needs to track changing Pareto optima accurately and quickly after the environment changes when dealing with these DMOPs. In response to the above issues, this paper proposes a prediction strategy assisted by long short-term memory recurrent neural network (PS-LSTM) for DMOPs. In the proposed method, when environmental changes occur, the difference of the center point of Pareto solutions is predicted by long short-term memory recurrent neural network and then the evolutionary direction of Pareto solutions can be calculated to better track changing Pareto optima by selecting excellent individuals in a new environment. Moreover, the next initial population is predicted based on the difference of the center point and the evolutionary direction of Pareto solutions. Finally, the PS-LSTM strategy is compared with four representative strategies on 11 benchmark test problems. Experimental results indicate that the strategy has better performance.