Why JMP is very slow on OSX (El Capitan) for the Gaussian models (kriging)

Question

Hi,

I have bought and installed JMP 12 on a Macbook air with Intel Corei7 processor and 8Gb of RAM.

JMP is nice, but I have bought it mainly for estimating metamodels using the kriging approach ( Models, Gaussian).

Unfortunately, the estimation process is hugely slow. I am using a subset of data with 2400 points of observation, with 8 independent variables, and it takes forever (more than 12 hours until now), and the progress bar barely moves.

Is this normal? I have read that this platform is not able to exploit multi-cores (I have two cores on which I can run 4 threads). How could I accelerate this process?

I have also a MacBook Pro which has 4 cores. Can I also install my JMP license on it? Well, if it does not use multi-cores, that would not solve my problem, I imagine...

I will have many Kriging estimations to do for my research, so any idea would be very welcome.

Very cordially,

Prof. Murat Yildizoglu

aab · Answer

Dear Ryan,

I'm currently testing the JMP 12 demo and one of our major issues was also the comparably very slow speed of the gaussian process. Dataset of 10 Variables and 1000 rows. Python 4s, JMP >5hours.

I guess you can't say anthing about a possible release period of JMP 13? Will this faster engine for gaussian processes only be available for the Pro version?

kind regards

Andreas

Bill_Worley · Answer

Professor Yildzoglu,

I agree, Neural Nets do offer their own set of challenges with parsimonious models being one of them.  Are you using JMP Pro?  There are several options in JMP Pro that will allow you build more robust predictive models with the NN platform.  These models will very likely not be as parsimonious as any gaussian models you build for the same data, but their predictive capability will likely be just as good or better.

The reason I like NN better (in most cases) than gaussian is that although gaussian fits every point exactly, the fit between two consecutive data points is difficult to determine and is extrapolation at best.  One way to check this is to look at your Actual by Jackknife Predicted fit for you gaussian fit. Draw a 45 degree line from the upper right hand corner to the lower left.  If your data is tight on this line then there is little bias in the data and you can have more trust that you are fit is good even between the points.  You have one advantage here in that you have a lot of data and you may not have to worry so much about what is going on between points.

The Prediction Profiler is where you will also see how the gaussian fit fluctuates between consecutive points.  In the attached data table there is a profiler comparison for a gaussian fit, an ordinary least squares (OLS) fit, and a NN fit of the same data set.  The NN fit was with 3 hidden nodes in the Tan H activation function and 5 layer K-fold cross-validation.  The OLS fit was built using RSM with more or less stepwise variable reduction.  The one thing I would like to point out is how "wiggly" the gaussian profiler is compared to the other two.  As you move the value in any one of the independent variables you will see the predicted values shift, but for the gaussian profiler you will also see that the individual prediction lines for the independent variables can move rather dramatically compared to the other two fit models.

I also ran a comparison of the three models using a Fit Y by X.  The gaussian fit shows a perfect fit for the model, but that does not line up with what you see in the profiler.  If there were no bias the profiler for the guassian fit would be as flat as the other two when you move from point to point.

The data I used here is stochastic in nature, but it is not unusual to see the same type of results with deterministic systems such as computer simulations.

Best,

Bill

Production BBM 12315 .jmp

yildi1 · Answer

Thank you very much Bill and Ryan!

I have already followed your suggestion Bill. Indeed ANN is much quicker. Unfortunately, the estimation seems quite unstable when I change the number of hidden nodes: the relationship between the dependent variable and the independent ones can just get inversed when I change this parameter. Which is not very assuring...

To give you an idea about what I am trying to do: I am trying to understand the relationship between the parameters and the output of an economic  simulation model. Using a DoE and a kriging estimation gives nice and parsimonious results in our context, as we have shown it in a Computational Economics paper.

I. Salle, M. Yildizoglu, Efficient Sampling and Metamodeling for Computational Economic Models, forthcoming in Computational Economics, published on-line, November 2013.