Abstract
This paper deals with a class of neural SDEs and studies the limitingbehavior of the associated sampled optimal control problems as the sample sizegrows to infinity. The neural SDEs with $N$ samples can be linked to the$N$-particle systems with centralized control. We analyze theHamilton-Jacobi-Bellman equation corresponding to the $N$-particle system andestablish regularity results which are uniform in $N$. The uniform regularityestimates are obtained by the stochastic maximum principle and the analysis ofa backward stochastic Riccati equation. Using these uniform regularity results,we show the convergence of the minima of the objective functionals and optimalparameters of the neural SDEs as the sample size $N$ tends to infinity. Thelimiting objects can be identified with suitable functions defined on theWasserstein space of Borel probability measures. Furthermore, quantitativeconvergence rates are also obtained.