Coefficient Calibration as Selection Pressure in Symbolic Regression
The paper introduces Dirichlet-Sinkhorn Constant Averaging (DSCA), a new calibration strategy for symbolic regression that uses controlled, heterogeneous data partitions to improve candidate evaluation. Instead of a single centralized fit, DSCA calibrates each candidate independently on different subsets and averages the results. This method can introduce selection pressure during the evolutionary search, potentially leading to better functional recovery and accuracy-complexity trade-offs.