Experimental survey of L-SHADE and SHADE-based adaptive differential evolution algorithms

Adaptive Differential Evolution (DE) methods are currently among the most efficient Evolutionary Algorithms. In the recent years different Success-History-Based Adaptive Differential Evolution algorithms (SHADE), often with linear population size reduction (commonly known as L-SHADE), have won numerous Competitions in Evolutionary Computation. Since 2014, the number and the variety of SHADE or L-SHADE-based algorithms flourished, encompassing novel operators and procedures. However, it is unclear which new SHADE/L-SHADE operators and procedures are the most successful, or efficient, for specific kinds of problems. After more than a decade of rapid development, some large-scale empirical tests are needed to select the best SHADE/L-SHADE-based algorithms for different purposes. This paper aims at a wide-scale inter-comparison between 32 SHADE/L-SHADE-based variants on large sets of various-dimensional benchmarks and on numerous real-world problems. We point at SHADE/L-SHADE-based algorithms that perform best for low-, or for high-dimensional problems. We determine variants that outperform others on simple problems, and those that perform best for more difficult tasks. Finally, we analyze which variants are best-suited for real-world applications, considering different computational budgets. Results indicate that much different SHADE/L-SHADE-based algorithms perform best for real-world problems than for numerical benchmark functions. Also, different algorithms may be recommended for higher, than for lower-dimensional problems, and other methods perform better for difficult problems than for unimodal ones. This discrepancy poses a challenge for choosing the appropriate algorithm for the specific application, and casts doubts on the classical way of justifying the introduction of novel variants.