Can I use parallel computing when training a gaussian process with separate length scales for predictors with fitgrp?

Question

Lauri Nenonen 2018-8-29

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/416670-can-i-use-parallel-computing-when-training-a-gaussian-process-with-separate-length-scales-for-predic

编辑： Lauri Nenonen 2018-9-25

My code for training the GP looks like this:

gpMdlspeed1 = fitrgp(model1,Speed1,'Basis','constant','FitMethod','exact',... 'PredictMethod','exact','KernelFunction','ardsquaredexponential','KernelParameters',[sigmaM01;sigmaF01],... 'Sigma',sigma01,'Standardize',1,'HyperparameterOptimizationOptions',struct('UseParallel',true), 'Verbose',2 );

But this is not using the parallelpool since my understanding is that hyperparameteroptimizationoptions applies only for bayesianopt optimizer. Is there a way to train this Gaussian process with separate length scales for predictors using parallel computing?

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Gautam Pendse 2018-9-25

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/416670-can-i-use-parallel-computing-when-training-a-gaussian-process-with-separate-length-scales-for-predic#answer_338306

Hi Lauri,

Even when there are separate length scales for predictors, these are jointly optimized during training. This optimization proceeds serially by moving from one set of values of the length scales to another set of values such that the log likelihood increases.

To speed up the optimization, you can consider loosening the convergence criterion as in this example:

https://www.mathworks.com/help/stats/fitrgp.html#d120e330919

Hope this helps,

Gautam