Continuous Delayed-Memory Stochastic Gradient Descent for Continuous-Time Reinforcement Learning from Astrophysical Time Series

This paper reviews the application of stochastic differential equations (SDEs) with neural network parameterizations to model quasar brightness variations from ground-based survey light curve data. The approach addresses the statistical challenge of extracting information about supermassive black holes from stochastic astrophysical time series. The methodology employs continuous delayed-memory stochastic gradient descent for continuous-time reinforcement learning from historical astronomical observations.