210 Appendix B. Appendixes: FT-MOEA (a) (c) (e) (b) (d) (f) Figure B.4: Visualization of metrics (εs, εd, εc) over the generations considering the percentiles 25%, 50%, 75% and 100% for the case study MPPS (ps =400, ng =100, uc =20). In (a), (c) and (e) using the m.o.f. sdc, and in (b), (d), and (f) using the m.o.f. d. Figures B.4(c) and B.4(d) illustrate the convergence of ωc. A similar pattern is observed, where our approach maintains more “variety” of FT structures between generations, with each structure being Pareto e"cient in at least one metric. In Figure B.4(d), we observe that variance decreases with generations, indicating that FTs tend to become more similar in their MCS matrices in later generations. Figures B.4(e) and B.4(f) illustrate the convergence of ωs. Here, the most significant di!erences between both approaches can be seen. In Figure B.4(e), we observe that the FTs tend to be small. It is important to note that just before finding the global optimum, the FTs within a generation tend to increase in size. Once the global optimum is found (i.e., ωd =ωc =0), the remaining part of the process naturally focuses on minimising the FT size, resulting in a compressed version of the found global optimum. On the other hand, Figure B.4(f) shows that not controlling the size of the FT results in a structural explosion, yielding massive structures that are not beneficial for the inference process.
RkJQdWJsaXNoZXIy MjY0ODMw