Characterize hyperparameter sensitivity and performance degradation

Characterize how RADAR's performance degrades when its reward-function hyperparameters—\(\alpha\), \(\beta\), \(C\), \(k\), and the penalty midpoint—depart from the manually selected configuration.

Background

RADAR balances trace entropy and collection volume through a reward function controlled by the hyperparameters $\alpha$, $\beta$, $C$, $k$, and the sigmoid penalty midpoint. The reported experiments used values selected through manual experimentation and showed stable behavior for that configuration.

The paper explicitly states that it does not characterize performance outside the selected configuration. A systematic sensitivity analysis is therefore needed to determine robustness and quantify how changes in the reward-function parameters affect convergence, resource consumption, entropy, and rare-trace preservation.