We are currently just choosing a bunch of random values and whichever values give the lowest "error" is returned.

We can improve in a number of ways:

• We can do some sort of gradient descent to find better values. We still need to do random sampling as well in case there are local minima, but from there we can do gradient descent.
• We can do a multi-threaded optimization (completed as part of bug: fix algorithm overflow issues #2173)

We should also redefine "error". Currently it's defined as as the sum of abs(profit) for each block. It's possible we are running in to overflows for the larger sets.