How description length tames bloated symbolic regression

Genetic programming for symbolic regression tends to produce bloated, overfitted equations—especially with noisy data. This work applies description length (DL), an information-theoretic criterion, as a principled way to select compact models that generalize better. Across noisy synthetic and real-world datasets, DL post-selection outperformed standard heuristics like AIC and BIC. However, using DL directly as a fitness function caused premature convergence; the sweet spot was multi-objective search for accuracy and compactness, then DL-based selection afterward.