4 4.5. Results 85 20 24 28 32 Conv. time (min) ddFT FT-MOEA-CM-All FT-MOEA-CM-Best FT-MOEA 16.5 18.0 19.5 21.0 MPPS 15 30 45 60 COVID-19 42 48 54 60 66 Conv. time (min) TS1 80 100 120 140 GPT12BE 240 300 360 420 GPT15BE Figure 4.5: Convergence time per case study and algorithm: FT-MOEA-CM-All (Blue box); FT-MOEA-CM-Best (Green box); FT-MOEA (Red box). convergence speeds were comparable. Nonetheless, in the GPT15BE case study, FT-MOEA converged faster but to a local optima. A di!erent perspective on the convergence process is illustrated in Figure 4.6, focusing on FT Size across generations and Figure 4.7, showing the computational time per generation. FT-MOEA-CM approaches larger FT sizes more rapidly (to attain global optima), then transitions to optimising the FT structure by reducing FT Size, a pattern also identified in Jimenez-Roa, Heskes, Tinga, et al., 2023. FT-MOEA 0 20 40 60 Generation 8 16 24 32 FTSize ddFT FT-MOEA-CM-All FT-MOEA-CM-Best 0 10 20 30 40 Generation 6 12 18 24 MPPS 0 25 50 75 100 Generation 0 20 40 60 80 COVID-19 0 20 40 60 80 Generation 15 30 45 60 FTSize TS1 0 25 50 75 100 Generation 0 25 50 75 100 GPT12BE 0 15 30 45 60 Generation 20 40 60 80 100 GPT15BE Figure 4.6: FT size across generations: Convergence for all case studies and algorithms based on FT size.
RkJQdWJsaXNoZXIy MjY0ODMw