– MPSO using the starting inertia weight, ending inertia weight, cognitive learning factor, and social learning factor being 0.5, 0.01, 0.5 and 0.5 respectively. The best solution for computing a velocity vector is randomly selected from a Pareto archive. The adaptive grid algorithm is used as an archiving technique.– NSGA-II using real codes with crossover mutation probabilities as 1.0 and 0.1 respectively.– RPBIL using real codes with NI = 40 where each probability tray produces five design solutions.– RPBIL-DE using real codes with NI = 40 where each probability tray produces five design solutions. Crossover probability, scaling factor and probability of choosing an element from an offspring in crossover for DE operators are set as 0.7, 0.8, and 0.5 respectively.– DMOEA using real codes with temperature and mutation rate of 1000 and 1 respectively.– DEMO using real codes with crossover probability, scaling factor and probability of choosing an element from an offspring in crossover for DE operators being 0.7, 0.8, and 0.5 respectively.– MOEA/D using real codes with number of neighbouring weight vectors, crossover and mutation probabilities being 6, 1.0, and 0.1 respectively.All MOEAs except UPS-EMOA use the population size of 100, and 250 generations for multiobjective unconstrained test problems. In addition, the population size and number of generations are 100 and 300 for multiobjective constrained test problems. UPS-EMOA uses different population size and iteration number as defined previously; however, their search procedures are terminated at the same total number of function evaluations. The Pareto archive of all MOEAs is set to be 500 except for NSGA-II and DEMO which contain Pareto solutions in a population. The optimisation parameter settings detailed above are obtained from using several settings for each optimiser and selecting the one that gives the best results.The ten multiobjective evolutionary optimisers are employed to solve each problem over 30 runs starting with the same initial population. The comparative results based on the hypervolume indicator are shown in Tables 3 and 4. The ZDT test problems are shown in Table 3 and the constrained test problems are shown in Table 4. Note that those hypervolume values are normalised for ease in comparison. The higher hypervolume value the better the Pareto front. With 30 runs, the mean value of the front hypervolumes is used to measure the algorithm’s convergence rate while the standard deviation represents the search consistency of the method. This means that a method with lower hypervolume standard deviation (STD) has better search consistency.From Table 3, the best performer (among compared MOEAs) according to the convergence rate (average hypervolume) for ZDT1, ZDT2, ZDT3, and ZDT6 is MOEA/D while NSGA-II is the best for ZDT4. The second best of ZDT1 and ZDT2 is BPBIL while RPBIL-DE is the second best of ZDT3 and ZDT6. For the ZDT4, the second best is MOEA/D. The worst method based on the convergence rate for ZDT1, ZDT2, and ZDT3 is MPSO while the worst of ZDT4 and ZDT6 is RPBIL. For a search consistency measure which is based on the hypervolume standard deviation, the most consistent method for ZDT1, ZDT2, ZDT3 and
đang được dịch, vui lòng đợi..
