A Large-Scale Sparse Multiobjective Optimization Algorithm Based on Optimal Performance Scores

What the math gives to ML

The paper's transferable contribution is a two-layer representation that separates a variable's real value from an explicit binary activity mask, combined with importance scores derived from non-dominated performance across value intervals. This turns sparse-variable discovery into a score-guided combinatorial search rather than relying on uniformly random masks. In neural networks, the same machinery can drive dynamic structured sparsity: maintain continuous weights separately from binary gates, estimate each gate's usefulness over several perturbation magnitudes, and use score-weighted mask mutations. The strongest initial test is sparse training or pruning of input features, neurons, heads, or blocks under a jointly measured accuracy–cost objective.