4 ms·

For up to 100ish parameters, even mixed with floating point I recommend trying the midaco solver a friend of mine develops. MINLP, ant colony method (i.e. gradi

For up to 100ish parameters, even mixed with floating point I recommend trying the midaco solver a friend of mine develops. MINLP, ant colony method (i.e. gradient descent with many restarts). From my experience this runs circles around NNs for this class of problems (parameter optimization with relatively low complexity and/or limited amount of training data available).

wiz21c 6y ago
Yep there are well established and powerful tools for various problems : linear programming, boolean satisfiability, analytical solutions, etc. NeuralNetworks and co are for a very specific, yet large, class of problems.