Unknown degradation from adding S31 shape features

Determine the cause of the observed degradation in generalization performance when augmenting the actor’s fully connected head with linear regression coefficients (slope and intercept) computed for four predefined regions of the S31 S-parameter, given that analogous S21-derived features improve performance in the same 1D ResNet-like actor for cavity duplexer tuning.

Background

The paper evaluates generalization of several 1D ResNet-like actors for cavity duplexer tuning across multiple datasets. In this paper, adding linear regression coefficients (slope and intercept) summarizing four regions of the S21 curve to the actor’s fully connected head improved generalization metrics.

However, the authors report that adding similar linear regression features derived from the S31 curve degraded the metrics. The reason for this degradation is explicitly stated to be unknown, posing an unresolved question about the interaction of these handcrafted S31 features with the learned representations and loss generalization.