Developed a Manifold-based modified ADMM algorithm tailored for large-scale or distributed binary-constrained optimization problems. The method is evaluated on binary distributed optimization instances, large scale ML feature selection problems, and challenging MIPLIB NP-hard classified instances to assess convergence behavior and solution quality for various applications.
- Distributed Optimization: Across the binary distributed optimization problems, the method consistently reached the a solution within a <10% optimality gap.
- ML Feature Selection: The combination of the algorithm and gurobi accelerates exact optimization by reducing time to high-quality solutions by 4.5%–over 95% across instances while maintaining comparable or better solution quality.
- MIPLIB: The algorithm warm-starting for Gurobi reduces computational time substantially on challenging instances, with savings exceeding 70% , while offering limited benefits or increasing runtime on easier instances.