This paper is solving (basically) high-school-level problems by training neural networks on the "obvious" cost function. All of those problems can be solved much cheaper by standard numerical solvers for ordinary differential equations. They don't even compare to standard methods.<p>So what's the point? Riding the neural network hype?