Latin Hypercube of discrete and continuous variables

I have a set of 6 variables, 2 continuous, 2 discrete, 1 variable whose values to be picked from a list, 1 constant. I have the min value, max value and a starting value for the discrete and continuous types. For Ex:

['var1':{'type':'continuous', 'lb':-1.2, 'ub':4.80, 'start':0.0}, 'var2':{'type':'discrete', 'lb':0, 'ub':100, 'start':10}, 'var3':{'type':'list', 'start':[1,2,10,20]}, 'var4':{'type':'discrete', 'lb':-12, 'ub':80, 'start':0}, 'var5':{'type':'continuous', 'lb':0.0, 'ub':0.8, 'start':0.0}, 'var6':{'type':'constant', 'start':3.14}]

These variables may be correlated. I want to use Latin Hypercube to sample these variables together preserving correlation between them. Please can you suggest how to do it.

Is there any MATLAB method to do this?

I have only found all continuous type sampling so far.

Thanks in advance.

[Yuhui Jia]

I am in a similar situation. Have you solved this problem? Any suggestion? Thank you very much.

Answer 1

[Marc]

Does the MixedLinearModel class come close to what you are trying to do?

Maybe you can start with that and then try and adapt a Latin Hypercube type code to handle your DOE design?

The optimization toolbox has mixed integer linear programming solvers but I know that this probably lacks the problem setup that you are going for.

I think in general the problem leads to multiple "best cases" so to speak. Ideally you don't want to do 100 experiments but there is probably "20" similar 16 experiment designs that give you the best chance at explaining how the 100 would react. Only problem is, which one should you pick?

When I set up a fractional factorial from mulitple levels, think 5 variables with 5 levels, I end up with a bunch of "best case experimental design". So, I try and weight the variables to create another objective function to week out the "best" from the bunch.

This is really all I can say about this.