AHPSO optimizer ranks first over CMA-ES with adaptive gradient control

Particle Swarm Optimization (PSO) is a widely used metaheuristic, but injecting gradient descent into it only helps after the swarm has found a smooth basin—not universally. A new paper by Aryan Gurudeo (arXiv:2608.11258) introduces Adaptive Hybrid PSO (AHPSO), which solves this by using a sigmoid function on swarm diversity to automatically adjust gradient influence. The mechanism keeps gradient near zero during exploration and near maximum during exploitation, eliminating the need for manual phase-switching. The paper reports extensive experiments: 40 configurations for budget-normalized comparisons and 29 functions with 42 configurations for iteration-matched comparisons, totaling over 14,700 independent optimization runs.

The results are nuanced but significant. Under budget-normalized comparison, standard PSO still wins 52.5% of configurations versus AHPSO's 20% (p = 7.0e-5, Friedman), showing that gradient injection is not universally beneficial. However, when both methods are given equal iteration counts, AHPSO-Adadelta ranks first among 9 methods—including CMA-ES—with p = 9.75e-4. The advantage is strongest on problems with smooth local basins (F8, F24-F27), where directed descent outperforms undirected sampling even at equal cost. The contribution is a principled characterization of when gradient injection provides value in swarm-based search, not a claim of universal superiority. This helps researchers choose the right optimizer for their specific problem landscape.