Sean's pick this week is also Wing Designer by John Rogers.

Contents

Background

A few weeks ago when Will picked Wing Designer I immediately had to go download it and try to beat his high score of 29252. After a bit of twiddling I was able to! By hand, I got a score of 30706.9.

But then I wondered what kind of score I could get with an optimizer from the Optimization Toolbox or the Global Optimization Toolbox?

Fortunately with the way John wrote this, I could tie into the scoring engine by emulating the output of user selections from the user interface.

Picking an Optimizer

Looking at the tunable parameters there are quite a few and many of them are integer constrained. Since the actual engine for this app is a black box to me, I am assuming that the score function is highly nonlinear with lots of local minima. Thus I need to take a global approach to this. The first function that jumps out is ga since this allows integer constraints. However, genetic algorithms don't always hold up well in as many dimensions as we have.

So instead I turned to patternsearch. Patternsearch is a direct search method that is often recommended over ga by Alan, who's a genius. But since patternsearch does not support integer constraints, I need to evaluate it on every combination of integers and then pick the best from these results.

Implementation

Wing Designer's user interface returns a geometry structure that I needed to emulate. Here's an example of part of the structure:

I wrote a function that takes in some of the various parameters that would be selectable in the user interface to create this structure.

Next, I needed to make some design decisions regarding my wing since trying every integer combination is just not feasible in the amount of time allotted. I spoke with one of our aerospace experts and he provided a few suggestions:

I made an additional few:

Making this Feasible

Since patternsearch has to run many times, once for each of the airfoils, and the function is called many times inside of patternsearch, I have to parallelize it. I could set the 'UseParallel' flag in psoptimset, but this would only parallelize individual patternsearch polls. Ideally, the calls to patternsearch should be parallelized, so instead we'll use a parallel for-loop (parfor) over the airfoils.

My laptop only has two cores, not enough to make this happen in the 12 hours before the blog post is due. So instead, I used a cluster running MATLAB Distributed Computing Server on Amazon EC2. I fired up a cluster with 36 workers which makes it much more feasible. I chose 36 workers: one for each airfoil, one to run the job itself, and an extra because it likes even numbers.

This is the body of OptimizeWing.m, the function that calls patternsearch.

% Lower and upper bounds
lb = [50 10 5 0 0 0 0 300 35000 0 0 5000 0];
ub = [200 40 10 5 5 5 5 660 50000 7 15 80000 100];

% x0, half way between bounds
x0 = lb + ((ub-lb)./2);

% Options
options = psoptimset('TolMesh',0.1,'MaxIter',100);

nfoils = 34; % 34 airfoils
fval = zeros(nfoils,1); % fval at xopt
xopt = zeros(nfoils,numel(x0)); % optimal x

% Run the parfor loop
parfor ii = 1:nfoils
    % Wrap the call in a try block in case some air foils cause failure
    try
        [xopt(ii,:), fval(ii)] = patternsearch(@(x)wingdesignerScore(x,ii),x0,[],[],[],[],lb,ub,[],options)
    catch
        % iith airfoil failed
        xopt(ii,:) = nan;
        fval(ii) = nan;
    end
end

First, I ran patternsearch once on one airfoil with very loose tolerances just to make sure it worked as expected. It did! Then it was time to submit the main function to the cluster, using batch.

batch('OptimizeWing.m',2,'Pool',34,'AttachedFiles',files)

And wait... Well actually, shut off my computer and go home for the evening. Since the job is running on EC2 somewhere in Virginia, there is no dependency on my machine.

Results

After retrieving the completed job in the morning, the best score is found by taking the minimum of the scores vector, fval and putting the corresponding xopt parameters into the App.

That's about 153000x better than we did by hand! The optimized wing is large, symmetric, and mostly filled with fuel. I'd be curious to hear thoughts from someone in the industry on the feasibility of a wing like this?

Comments

Do you have a challenging optimization problem or some really computationally expensive work where cluster or cloud computing could help?

Let us know about it in the comments!

Get the MATLAB code

Published with MATLAB® R2014a