摘要
针对超参数梯度目标过大,引起的搜索调优效率低问题,提出基于大数据的复杂超参数优化组合方法,来实现有效解决。使用汤普森采样法函数收集关键超参数数据,通过高斯现象计算超参数优化组合的观测向量值及期望值,求取观测数据集最优求解,建立初步优化组合框架,计算每一层次框架中对应超参数先、后验概率以及最大似然函数,对比推导出优化函数的状态分布规律,寻找符合条件的导数因子,作为最大优化矩阵的初始代入数值,实现超参数的最优解组合。仿真结果证明,所提组合下参数平均求解误差较低、整体优化性能表现较为优异。
Aiming at the low efficiency of search and optimization caused by too large hyperparameter gradient target,a complex hyperparametric optimization combination method based on big data is proposed to effectively solve the problem.The Thompson sampling function is used to collect the key hyperparameter data,calculate the observation vector value and expected value of the hyperparameter optimization combination through the Gaussian phenomenon,obtain the optimal solution of the observation data set,establish the preliminary optimization combination framework,calculate the first and a posteriori probabilities and maximum likelihood functions of the corresponding hyperparameters in each level framework,and compare and deduce the state distribution law of the optimization function,and find the qualified derivative factor as the initial value of the maximum optimization matrix to realize the optimal solution combination of super parameters.Simulation results show that the average solution error of the parameters under the proposed combination is low and the overall optimization performance is excellent.
作者
李发陵
彭娟
LI Fa-ling;PENG Juan(Chongqing Institute of Engineering,Chongqing 400056,China)
出处
《计算机仿真》
北大核心
2022年第7期332-336,共5页
Computer Simulation
基金
重庆市教育委员会科学技术研究项目(KJQN201901908)
重庆工程学院科研基金资助项目(2020xzky02)。
关键词
搜索调优
复杂超参数
汤普森采样法
梯度分布层次
最大似然函数
Search tuning
Complex hyperparameters
Thompson sampling method
Gradient distribution hierarchy
Maximum likelihood function
