An Inverse Modeling Multi-Objective Optimization Technique Based on Incremental Learning and Fuzzy Clustering

The use of inverse modeling-based crossover operators in multi-objective evolutionary algorithms (MOEAs) has recently received much attention. Sampling in the objective space is advantageous over sampling in the decision space as it allows selecting promising areas worthy to explore. This paper aims to develop an inverse modeling MOEA based on decomposition that employs an incremental learning-based support vector regression (SVR) model, as an alternative to the Gaussian process model, in order to improve the quality of obtained solutions and speed up convergence of the algorithm. Several inverse SVR models are constructed and the samples in the objective space are partitioned among them based on fuzzy clustering instead of hard clustering to enrich the training process. Extensive simulations on various benchmark problems show that the proposed algorithm drastically reduces the number of function evaluations required to reach an optimal solution compared to existing methods. The algorithm is also tested on the pathfinding problem, the community detection problem, the sparse portfolio problem, and other real-world problems, all of which confirmed the scalability and superiority of the proposed algorithm.