CMStatistics 2018: Start Registration
View Submission - CMStatistics
B1682
Title: Stochastic decision-making using particle methods Authors:  Maciej Marowka - Imperial College London (United Kingdom) [presenting]
Nikolas Kantas - Imperial College (United Kingdom)
Abstract: A novel numerical, particle based method is proposed to estimate the optimal control inputs for a risk sensitive stochastic decision-making problem where a multiplicative reward is used. The problem is to identify a control sequence such that the resulting observations from the non-linear non Gaussian state-space model match a required deterministic reference sequence with respect to the particular choice of reward. The approach is based on earlier efforts for deterministic systems and is essentially a sequential Monte Carlo (SMC) algorithm for an appropriate dual filtering problem. Extensions using SMC$^{2}$ for nonlinear models based on hierarchical (deep) state space models are developed. We will illustrate the performance of the method on particular synthetic data examples and discuss possible applications to the optimal trading problem in the environment with time varying cointegration model, where the cointegration space bases are driven by a latent stochastic process.