A simple multiple kernel learning framework is proposed in this paper,which is used to model the regression problem of complex data.The multi-scale Gaussian kernels are chosen as basic kernels,and the kernel scale parameters are set accord-ing to the random allocation method.We use the training sample to select the optimal model adaptively,combining the l1 regularization method,and construct a single task machine learning solution with sparse constraints.Then,the multiple kernel learning framework is extended to multi-task learning model,a multi-task model with group sparse is proposed to improve the traditional linear multi-task model by using the classical l2,1 regularization.On the base of these,the short-comings of random l2,1 regularization multi-task model in computational efficiency are analyzed.Based on the inherent mechanism of multi-task method and genera-tive network,the multi-task model of generating network with random multi-scale kernel is constructed,and the limit problem of residual is studied.The basic functions provided by the method of random set of kernel parame-ters have sufficient approximation ability,the parameters are obtained by random sampling in a preset probability distribution,and the probability distribution pa-rameters and kernel parameters are calculated by performing cross validation in the training samples.Data experiment is divided into two parts:single-task algorithm and multi-task algorithm,by modeling analysis of the simulation data and real data respectively.The results show that the stochastic multi-scale kernel method can adaptively select the parameters,and the generalization performance of the model is better than that of the original method. |