Efficient model-free Q-faetor approximation in value space via log-sum-exp neural networks

We propose an efficient technique for performing data-driven optimal control of discrete-time systems. In particular, we show that log-sum-exp (LSE) neural networks, which are smooth and convex universal approximators of convex functions, can be efficiently used to approximate Q-factors arising from finite-horizon optimal control problems with continuous state space. The key advantage of these networks over classical approximation techniques is that they are convex and hence readily amenable to efficient optimization.