Table 5.
Improvement δ in accuracy for different iterative optimization settings in the sonar dataset.
| Linear [%] | KNN [%] | Bayesian [%] | SVC [%] | NN [%] | |
|---|---|---|---|---|---|
| Base accuracy | 75.195 | 81.343 | 67.700 | 84.052 | 84.024 |
| Diagonal | 0.076 | 0.290 | 1.829 | –0.395 | –0.957 |
| Spherical | 0.586 | 0.095 | 6.067 | –0.300 | –1.910 |
| MLE | –0.167 | 0.000 | 6.443 | 1.810 | –0.291 |
| Iterative grid search (δ) | 1.162 | 1.824 | 7.919 | 2.286 | –0.386 |
| Shift (δ) | 0.976 | 1.919 | 7.919 | 2.381 | –0.386 |
| Shuffle (δ) | 1.162 | 1.824 | 7.919 | 2.286 | –0.386 |
| Finer (δ) | 0.876 | 1.824 | 7.919 | 2.286 | –0.386 |
| Combined 1 (δ) | 0.600 | 1.838 | 7.157 | 2.190 | –0.386 |
| Combined 2 (δ) | 2.795 | 3.267 | 8.100 | 2.010 | –0.486 |
The proposed optimization achieved a consistent improvement except for the neural network. The different hyperparameter settings had a varying influence on the classifiers defined in Table 4. Combined 2 improved the linear, KNN, and Bayesian classifier, whereas Shift already delivered the best performance for SVC. Additionally, the proposed optimization achieved higher improvements than Diagonal, Spherical and MLE except for the neural network.