Improving the applicability of genetic algorithms to real problems

The current state-of-the-art of genetic algorithms is dominated by high-performing specialistsolvers with fast convergence. These algorithms require prior knowledge about the characteristics of the optimised problem to operate effectively, although such information is not available in most real cases. Most of these algorithms are only tested on a narrow range of similar benchmarking problems with lower complexity than the real cases. This leads to the promotion of high-performing strategies for these cases, but which might prove to be ineffective on practical applications. This hypothesis is supported by a low uptake of the current specialist/convergence algorithms on real-world cases; NSGA-II remains the most popular algorithm despite being developed in 2002. It is suggested that this is due to its uniformly good performance across a wide range of problems with distinct characteristics, indicating high generality, and a high diversity retention across iterations. To assess if increasing the generality and diversity of the search improves the performance on real problems, the Multi-Level Selection Genetic Algorithm (MLSGA) is extended to develop a “diversity-first” general-solver genetic algorithm. It is selected as it shows high promise for the diversity-oriented methodology. Firstly, the reasons behind why it exhibits high diversity are investigated, as it is shown that the collective-level mechanisms create additional evolutionary pressure, while the fitness separation approach leads to collectives targeting different regions of the search space. This creates unique region-based search which leads to retention of higher diversity of solutions between generations. Secondly, as the MLSGA is exhibiting a poor convergence, the algorithm is combined with the current state-of-the-art algorithms in the hybrid approach (MLSGA-hybrid) to offset this problem. MLSGA-hybrid focuses on increasing the convergence of the search, over the original MLSGA algorithm, while retaining its emphasis on the diversity. The results demonstrate that this improvement leads to top performance on a range of problems. This is particularly the case on constrained problems indicating that the diversity has been retained. Thirdly, the co-evolutionary variant is introduced and tested (cMLSGA), which combines multiple evolutionary algorithms to improve the generality of the method. To validate the performance of the “diversity-first” general-solver approach, the algorithm is tested on 100 benchmarking problems and compared with top algorithms from the current state-of-the-art. It is shown that cMLSGA is the best iv general-solver, due to the most robust performance across the evaluated cases, while maintaining a higher focus on problems where elevated diversity of the search is preferred, such as discontinuous, constrained and biased cases. Finally, the cMLSGA approach is benchmarked on 3 engineering cases with a wide range of diverse characteristics and compared with other leading genetic algorithms. It is shown that the convergence-oriented solvers are ineffective for real-world applications due to higher complexity of practical problems, whereas performance of specialist-solvers is low due differences between real-world cases and benchmarking functions they are adjusted to. According to that, the “diversity-first” genetic algorithms with a high generality are preferred and there should be more focus on algorithms with these characteristics in the future