Parameter estimation fits a model; model selection chooses among plausible models to represent the underlying process. Since added parameters often improve fit to observed data, selection balances goodness of fit against complexity. section 3.7
A disciplined comparison keeps model choice separate from final evaluation:
- Compare candidate models using criteria such as likelihood ratio tests, AIC, or BIC. section 3.7
- For kernel configurations, fit each candidate on the same training folds and compare validation performance using a selected classification measure. section 8.6.3
- Refit the selected configuration on all training data; keep the test set untouched until selection is complete. section 8.6.3
Choosing an SVM kernel configuration
For an RBF kernel, γ controls boundary responsiveness and interacts with the soft-margin penalty C. Large values of both can produce a highly responsive boundary; cross-validation compares candidate settings for likely generalization rather than choosing by training accuracy. section 8.6.3











