I am trying to use bayesopt to optimize a number of parameters (from 2 to 16, depending on the problem setting). I see from this paper https://arxiv.org/pdf/0912.3995.pdf(table1) that the number of iterations to reach the optimum scales exponentially with then number of parameters using the Matern kernel. The relationship would have been polynomial if RBF kernel is used. Since evaluating my cost function is expensive (around 20 mins) and I can only afford to optimize 3 parameters within reasonable amount of time, I would like to try RBF in bayesopt. But the Matern kernel is default and it seems that RBF kernel is not even included in the fitrgp function. Should I write my own RBF kernel and supply it to fitrgp and modify the bayesopt source code? Is there any other methods to make high dimensional Bayesian optimization scalable? Thanks!