A dynamic self‑adaptive optimization framework integrating the metaheuristic strategy for the non‑convex stochastic optimization problem

As a critical branch of optimization theory, adaptive optimization algorithms mitigate the limitations of classical gradient descent, such as slow convergence and vulnerability to local optima, particularly in high-dimensional, non-convex environments. This is achieved by dynamically adjusting learning rates and per-parameter update rules. By integrating metaheuristic strategies with adaptive gradients, a dynamic self-adaptive optimization framework (DSOF) is proposed in this paper. The DSOF introduces a history-informed second-order momentum mechanism, which leverages recent gradient statistics to improve robustness under non-stationary objectives. Additionally, it maintains population diversity through evolutionary crossover and mutation operations to prevent premature convergence, and adaptively contracts or expands parameter bounds to balance exploration and exploitation. Experimental results demonstrated that, under the evaluated settings, the DSOF improves optimization stability by reducing the gradient-variance fluctuation coefficient from 0.418 to 0.382 in the modified national institute of standards and technology classification task, achieves a 100% global optimum hit rate on the 50-dimensional Ackley function, and reduces NAS-Bench-201 search time to 25 hours with only 1.02 GPU-days. These results verify the empirical effectiveness of the DSOF in improving stability, global exploration capability, and computational efficiency.