Abstract 
Many technical issues encountered by coalition forces, including SDC resource sharing, network management, distributed analytics and machine learning, can be formulated as optimization problems. Gradientbased iterative algorithms have been widely utilized to solve these problems. The core difference among various algorithms is the way to update the solution based on the current gradient values in each iteration. Instead of designing a new method by exploiting the structures of the involved functions, employing machinelearning techniques to learn how to design an efficient solution framework for optimization problems is of interest.
We propose here a learning approach to solve nonconvex, constrained stochastic optimization problems with userdefined objective and constraint functions. Two coupled Long ShortTerm Memory (LSTM) networks are used to find the optimal solution. The new approach is validated by reproducing the solution to the nonconvex problem for innetwork data processing [1].
Efforts are in progress to extend the new technique for outstanding stochastic optimization problems and systems with temporal dynamics (e.g., SDC). Our latest results confirm the feasibility of the LSTMs to include temporal system evolution, while providing the optimal control decisions. Furthermore, it is also shown to solve some forms of stochastic optimization problems.
