A Pre-Sorted Autoencoder-based Prediction for Dynamic Multiobjective Optimization

In this paper, a pre-sorted autoencoder-based prediction algorithm, named PSAE, was proposed to solve dynamic multiobjective optimization problems (DMOPs). Since transfer learning enables knowledge acquisition from related tasks, its application to predict the Pareto-optimal solutions has gained increasing attention. Especially, the autoencoder (AE) was introduced to solve DMOPs and has gained promising results. However, the current AE-based prediction lacks an effective mechanism to match input and output knowledge, leading to inconsistencies between the input and output solutions. For solving this problem, for the first time, the thin-plate spline robust point matching (TPS-RPM) method is introduced to achieve knowledge matching between inputs and outputs, thus enhancing prediction accuracy. To verify the effectiveness of proposed algorithm, some recently proposed dynamic multiobjective evolutionary algorithms are selected in the experiments and the results show the superiority of PSAE.