Parameter inference for stochastic biological models: the cross-entropy method
by Jeremy Revell
16:00 (40 min) in CT 7.01
Computational modelling is an essential component to achieving the aims of synthetic biology, which aims to develop a systems biology system level understanding of complex biological processes. In this project we aim at making fundamental advances in parameter inference and parameter synthesis, two of the most important computational modelling problems. In particular, we focus on stochastic biological models. In recent years, it has become well understood that stochasticity within finite populations can produce dynamics profoundly different from the predictions of corresponding deterministic models. Unfortunately, analytic solutions to stochastic time-evolution equations are often intractable, while numerical solutions are often computationally infeasible.
Over recent months, we have investigated a novel computational approach to parameter inference for stochastic biological models based on the cross-entropy method: a recent advance in the field of rare-event estimation. In this talk, we will present the Monte Carlo Expectation-Maximisation with Modified Cross-Entropy Method (MCEM2), a highly efficient Maximum Likelihood Estimate (MLE) algorithm developed for the stochastic simulation service by leaders in the field (StochSS, Petzold et al, 2012). We argue that the cross-entropy method can in fact `do all the work’, and is sufficient without the need for expectation-maximisation. We present a simpler Cross-Entropy Method for Parameter Inference: where early findings not only produce better or equivalent parameter estimates – but can deliver them in fractions of the time (with a relative speedup of 8500× in one model!).