CDiNN – Convex difference neural networks

Neural networks with ReLU activation function have been shown to be universal approximators and learn mapping as non-smooth functions. Recently, there is considerable interest in the use of neural applications such optimal control. It well-known that optimization involving non-convex, functions are computationally intensive limited convergence guarantees. Moreover, choice hyper-parameters used gradient descent/ascent significantly affect quality obtained solutions. A new network architecture called Input Convex Networks (ICNNs) output a convex inputs thereby allowing efficient methods. Use ICNNs for determining input minimizing has two major problems: learning non-convex could result significant approximation error, we also note existing representations cannot capture simple dynamic structures like linear time delay systems. We attempt address above problems by introduction architecture, which call CDiNN, learns difference polyhedral from data. discuss that, some cases, can CDiNN through guarantees at each iteration, problem reduced programming problem.