New evolutionary algorithm beats SOTA on large-scale sparse optimization
Accurately identifies critical nonzero variables in high-dimensional problems with a novel score-based method
Large-scale sparse multiobjective optimization problems (LSSMOPs) involve many decision variables but only a few nonzero ones, making it hard to find optimal solutions. Researchers introduce a new evolutionary algorithm that tackles this by first generating scores that reflect each variable's importance and an initial mask vector template to pinpoint nonzero variables. This produces a high-quality initial population. The algorithm also calculates mutation probabilities per variable and uses a Pareto-guided normal distribution to optimize real variables, helping the population avoid local optima and converge quickly.
The method was validated on eight benchmark problems and three real-world applications, showing superior performance compared to current state-of-the-art algorithms. This approach is particularly valuable for fields like engineering design and data science where sparse solutions are common but scaling optimization remains a challenge.
- Novel initialization method produces importance scores and mask vectors to identify nonzero variables in high-dimensional spaces
- Pareto-guided normal distribution for real-variable optimization prevents local optima and accelerates convergence
- Outperforms state-of-the-art on all 8 benchmark problems and 3 real-world applications
Why It Matters
Enables efficient optimization of large-scale problems with sparse solutions, reducing computational cost in engineering and AI.