Importance Sampling, Large Deviations, and Differential Games

From MaRDI portal





Let \(S_n=Y_0+\dots+Y_{n-1}\), where \(Y_i\) are i.i.d. random vectors in \(\mathbb{R}^d\). The authors consider approximation of the probability of a rare event \(p_n=\Pr\{S_n/n\in A\}\) (\(n\) large) using importance sampling scheme with dynamic adaptive changes of measures. I.e. the sums \(\overline S_n=\sum_{i=0}^n\overline Y_i^n\) are simulated in which the distribution of \(\overline Y_j^n\) depends on the historical mean \(\sum_{i=0}^{j-1}\overline Y_i^n\). Asymptotic optimality (as \(n\to\infty\)) of such sampling scheme is introduced and the optimal scheme is described. This construction is based on the large deviations (Cràmer's) theorem and a representation of the problem in terms of differentiable games. The solution is derived by the Isaacs equation. Numerical examples related to normally distributed random variables are presented. It is noted that the approach can be extended to estimates of path dependent events, stochastic networks and processes with state dependent dynamics.




Cited in
(74)








This page was built for publication: Importance Sampling, Large Deviations, and Differential Games

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q4659569)