|
Abstract:
|
Sequential filtering approaches are common in estimating parameters for dynamical systems. In this work, we develop an approach that deals with sequential calibration of computer model of a dynamical system that evolves over time and space. As data from the system is observed at regular time intervals, inference on input parameters need to be updated periodically. In particular, we consider stochastic computer models where repeated runs of the simulation at same input yield a realization from an unknown distribution. The proposed double sequential calibration strategy employs two sequential learning schemes that learn the optimal values of the input parameters sequentially within each time interval given a limited simulation budget and then across time horizon as more and more data becomes available. The full posterior distribution of the input parameters is obtained at the end and uncertainty measures on the predictions are directly available from the calibrated simulations. An epidemic model that simulates infectious disease through contact network (e.g., Covid-19) will be used for illustration.
|
