Research framework for geometric reformulation of black-box optimization. SGOLab proposes a domain-codomain decoupling principle, navigating the search space through an independent geometric reference system rather than explicit landscape reconstruction.

A memory-efficient, gradient-free zeroth-order (derivative-free) optimizer designed to solve the "Curse of Dimensionality" in Black-Box optimization and memory-constrained Machine Learning. It provides an O(log D) gradient estimation approach that can successfully train Neural Networks without ever calculating analytical derivatives or Backprop

High-dimensional sparse regression in R: penalized Huber, SVM, quantile, and Wilcoxon rank regression via the finite smoothing algorithm. Fast C++ (Rcpp) coordinate-descent kernels with elastic-net penalties, cross-validation, and SCAD/MCP non-convex penalties.