Bézier Walk Evolution redefines global optimization with geometry-driven search
New algorithm uses Bézier curves to balance exploration and exploitation, beating CMA-ES.
A team of researchers from Shuimu AI and Beihang University have introduced Bézier Walk Evolution (BWE), a novel optimization framework that replaces traditional nature-inspired heuristics with a geometric approach. BWE models candidate solutions as control points on Bézier curves, with the curve order adaptively varying during the search. High-order curves, controlled by many population points, encourage broad exploration by generating complex trajectories, while low-order curves create near-linear paths for refined local exploitation. This built-in mechanism smoothly transitions the algorithm from global search to local convergence, offering a more interpretable alternative to black-box metaheuristics.
In extensive tests on 41 benchmark functions from CEC2017 and CEC2022 suites (dimensions 10–100), BWE demonstrated strong overall performance and favorable scalability. It outperformed 7 classical optimizers (e.g., differential evolution, PSO) and 6 state-of-the-art methods including L-SHADE and CMA-ES. Additional validation on five constrained engineering design problems (e.g., pressure vessel, welded beam) confirmed its practical applicability and robustness. The paper, published on arXiv, provides code and detailed results, positioning BWE as a promising new direction for global optimization tasks in AI, robotics, and engineering.
- BWE adapts Bézier curve order during evolution to smoothly transition from global exploration to local exploitation.
- Outperforms 7 classical and 6 state-of-the-art optimizers (including L-SHADE and CMA-ES) on 41 CEC2017/2022 benchmarks.
- Validated on 5 constrained engineering design problems, demonstrating real-world robustness and scalability up to 100 dimensions.
Why It Matters
Offers a more interpretable, geometry-driven alternative to nature-inspired optimizers, improving efficiency in engineering and AI hyperparameter tuning.