Title:Online Rolling Controlled Sequential Monte Carlo

Abstract:We introduce methodology for real-time inference in general-state-space hidden Markov models. Specifically, we extend recent advances in controlled sequential Monte Carlo (CSMC) methods-originally proposed for offline smoothing-to the online setting via a rolling window mechanism. Our novel online rolling controlled sequential Monte Carlo (ORCSMC) algorithm employs two particle systems to simultaneously estimate twisting functions and perform filtering, ensuring real-time adaptivity to new observations while maintaining bounded computational cost. Numerical results on linear-Gaussian, stochastic volatility, and neuroscience models demonstrate improved estimation accuracy and robustness in higher dimensions, compared to standard particle filtering approaches. The method offers a statistically efficient and practical solution for sequential and real-time inference in complex latent variable models.