fig6
Figure 6. Convergence at different batch sizes with the same total evaluation budget of 24 × 8. (A) ZDT1; (B) ZDT2; (C) ZDT3; (D) MW7. We omitted qNEHVI for a batch of 16 due to the prohibitively high computation cost when scaling up. Plots are taken with mean and 95% confidence interval of log10(HVmax - HVcurrent), with HVmax being computed from known PF in pymoo. We follow the same details as for Figure 5. Results suggest that qNEHVI works better with low batching on disconnected PF. HV: hypervolume; PF: Pareto Front; qNEHVI: q-Noisy Expected Hypervolume Improvement; U-NSGA-III: Unified Non-dominated Sorting Genetic Algorithm III.
