Research & Papers

DE-2LS boosts optimization by 11% with late-stage local search

New algorithm improves convergence speed and final quality by 34% over competitors

Deep Dive

DE-2LS, introduced by Dikshit Chauhan, enhances the RDEx evolutionary search engine with two conservative refinements for unconstrained single-objective numerical optimization (with variable bounds). The first is a smoothed exploitation-biased branch-rate update that shifts focus from global exploration to local exploitation in later stages. The second is a guarded coordinate-pattern local search that acts as a budget-aware refinement mechanism, ensuring function evaluations are used efficiently. Only objective values guide selection and acceptance decisions.

Experimental results demonstrate significant gains: DE-2LS increased RDEx's U-score from 33,602 to 37,448 (11.45% improvement). More impressively, compared to several competitive and IEEE CEC-winning algorithms, DE-2LS achieved the top overall U-score of 178,966.5—outperforming the next best by 34.43%. These results show that a carefully designed late-stage local-search strategy can simultaneously improve convergence speed and final solution quality. The source code is publicly available.

Key Points
  • DE-2LS adds smoothed exploitation-biased branch-rate update and guarded coordinate-pattern local search to RDEx
  • Achieved U-score improvement from 33,602 to 37,448 (+11.45%) over the baseline
  • Outperforms IEEE CEC-winning algorithms by 34.43% with best overall U-score of 178,966.5

Why It Matters

Better optimization algorithms mean faster, higher-quality solutions for engineering, ML, and scientific computing problems.

📬 Get the top 10 AI stories daily