A dynamic multi-objective optimization evolutionary algorithm based on classification of decision variables and inter-layer collaborative prediction

Abstract

The Dynamic Multi-objective Optimization Evolutionary Algorithm (DMOEA) demonstrates outstanding performance in addressing complex Dynamic Multi-objective Optimization Problems (DMOPs). However, existing DMOEAs lack a mechanism for classifying decision variables, which makes it challenging for the population generated by the response mechanism to achieve a balance between convergence and diversity, thereby compromising the algorithm’s overall optimization performance. For this reason, this work presents a DMOEA based on decision variable classification and inter-layer collaborative prediction. First, the algorithm classifies decision variables into two types (elite decision variables and routine decision variables) by detecting their characteristics and designs corresponding response mechanisms for different variables. Second, an inter-layer collaborative prediction strategy based on the Gate Recurrent Unit (GRU) model is proposed to handle routine decision variables, while elite decision variables are optimized using a Latin Hypercube Sampling (LHS) strategy. Subsequently, the two types of optimized variables are combined to form the final predicted population. Finally, a population re-optimization strategy (including dominant solution filtering and adaptive mutation) is proposed to finely optimize the predicted population, thereby further improving prediction accuracy. Through experiments on 24 test functions with seven high-performance DMOEAs, it is demonstrated that the algorithm has significant advantages in both convergence and diversity.