Abstract

Partially observed nonlinear stochastic dynamic systems raise inference challenges. Sequential Monte Carlo (SMC) methods provide a route to accessing the likelihood function. However, despite the advantage of applicability to a wide class of nonlinear models, standard SMC methods have a limitation that they scale poorly to large systems. We present a profile likelihood approach, properly adjusted for Monte Carlo uncertainty, that enables likelihood-based inference in systems for which Monte Carlo error remains large despite stretching the limits of available computational resources. Together with state-of-the-art SMC algorithms, this technique permits effective inference on some scientific problems in panel time series analysis, spatiotemporal modeling, and inferring population dynamic models from genetic sequence data. The results presented are joint work with Carles Breto, Joonha Park, Alex Smith and Aaron King.