: . ural networks Deep learning put has two major problems: learning of a non-convex function as a convex mapping could result in sig- Neural networks nificant function approximation error, and we also note that the existing representations cannot capture Optimal control simple dynamic structures like linear time delay systems. We attempt to address the above problems by introduction of a new neural network architecture, which we call the CDiNN, which learns the function as a difference of polyhedral convex functions from data. We also discuss that, in some cases, the optimal input can be obtained from CDiNN through difference of convex optimization with convergence guaran- tees and that at each iteration, the problem is reduced to a linear programming problem. Ó 2022 Elsevier . All rights reserved. 1. Introduction tion. The most popular approach is the back-propagation approach, which is a gradient descent algorithm. Several vari- Neural networks have been used for function approximation ants of this algorithm have been evaluated for neural network and classification tasks in several applications. The fact that these training. Since the loss function is usually non-convex, issues networks are shown to be universal function approximators is a related to convergence and
