Multi-loop Adaptive Differential Evolution for Large-Scale Expensive Optimization

Data-driven evolutionary algorithms (DDEAs) have shown promising performance in solving small- and medium-scale expensive optimization problems (EOPs) with low and medium dimensions. However, the performance of existing DDEAsis still not good enough for large-scale EOPs. To efficiently solve the large-scale EOPs, this paper proposes a new offline DDEA based on adaptive differential evolution (DE), called ADE-DDEA. To obtain a sufficiently accurate surrogate, ADE-DDEA introduces a Latin hypercube sampling strategy with standardization for collecting data and a grid searching based approach for training the surrogate. Moreover, ADE-DDEA employs a state-of-the-art adaptive DE (i.e., the jSO) as the optimizer and a multi-loop strategy to avoid trapping into local optima. In the experiment, this paper compares ADE-DDEA with the traditional algorithms on five commonly used test functions with different dimension scales. ADE-DDEA shows significant advantages on large-scale EOPs with only 10% computational budgets of traditional methods. Even if compared with the state-of-the-art DDEAs for large-scale EOPs, ADE-DDEAshows sufficient competitiveness. Furthermore, the experimental results also show the advantage of the algorithm in terms of running time consumption.