3 3.4. Experimental evaluation 71 All Espresso FT-MOEA No Sym. No rec. Sympy SC SS TS1 TS2 TS3 (b) 0.00 0.06 0.12 0.17 0.23 SC SS TS1 TS2 TS3 (c) 0 17 35 52 69 2 4 6 8 Time (hours) SC SS TS1 TS2 TS3 (d) 0 1 2 3 Time (min) SC SS TS1 TS2 TS3 (a) 0.00 0.06 0.12 0.18 0.24 c d s Figure 3.5: Results for the case studies and di!erent metrics: (a) error εc based on the MCSs, (b) error εd based on dataset, (c) FT size εs, and (d) runtime. is the original implementation (Jimenez-Roa, Heskes, Tinga, et al., 2023) without modules and symmetries. • Espresso translates a set of MCSs CDinto a Boolean formula &C→CD%b→C b and simplifies it via the ESPRESSOalgorithm (Brayton, Hachtel, McMullen, et al., 1984) available in pyeda∗. The resulting formula is then translated into an FT. • Sympy is similar to Espresso but uses the sympy library† for simplification. We ran all case studies three times on a CPU with 2.3 GHz and 8 GB of RAM. Results. We compare the FTs for case TS1 inferred viaFT-MOEA(Figure 3.4(c)) and via SymLearn in configuration All (Figure 3.4(d)). Colours depict the connections of the BEs to the components in Figure 3.4(a). SymLearn identified the symmetry (between yellow and blue BE) and was able to infer the left sub-tree using FT-MOEA while the right sub-tree was obtained by simple mirroring. The box charts in Figure 3.5 compare the di!erent configurations in all five cases w.r.t. the three metrics in Section 3.2: the size |F| of the FT, the error ωd based on the failure dataset, and the error ωc based on the MCSs. From Figure 3.5(a) and 3.5(b), we see that the SymLearn configurations based on Boolean functions as a back-end (i.e., Espresso and Sympy) always yield an FT that exactly matches the input, i.e., ωc =ωd =0. This is expected since the Boolean logic formula perfectly ∗https://pyeda.readthedocs.io/en/latest/2llm.html †https://docs.sympy.org/latest/modules/logic.html
RkJQdWJsaXNoZXIy MjY0ODMw