稳健性(进化)
算法
趋同(经济学)
理论(学习稳定性)
数学
变量(数学)
数学优化
计算机科学
标识
DOI:10.1007/s00034-022-02015-5
摘要
The convergence, robustness and steady-state performance are three key points to consider when evaluating variable step-size algorithms. The theoretical depth of the algorithm, the stability of step-size and the performance analysis all affect these key points. Based on these points, a novel variable step-size algorithm was derived from a steady-state mixed error cost function. The algorithm was developed according to basic theoretical formulas and provides a certain level of theoretical depth. In addition, to balance the relation between convergence and robustness, an improved restricted method that uses different restricted parameters at different stages was combined with the step-size algorithm. This method maintains the convergence speed while reducing fluctuations. Furthermore, the proposed algorithm eliminates numerous parameters and approaches the optimal step-size without increasing the computational burden. Weight misalignment, stability and complexity performance analyses are also provided. According to the experimental results and analyses, the proposed algorithm has a good performance.
科研通智能强力驱动
Strongly Powered by AbleSci AI