Unsupervised Learning for Parametric Optimization

This letter proposes the unsupervised training of a feedforward neural network to solve parametric optimization problems involving large numbers parameters. Such training, which consists in repeatedly sampling parameter values and performing stochastic gradient descent, foregoes taxing precomputation labeled data that supervised learning necessitates. As an example application, we put this technique use on rather general constrained quadratic program. Follow-up letters subsequently apply it more specialized wireless communication problems, some them nonconvex nature. In all cases, performance proposed procedure is very satisfactory and, terms computational cost, its scalability with problem dimensionality superior convex solvers.