Given some random variable
For situations where we have special interests in rare events (which can be described by a subset
The importance sampling can let one draw more frequently from
In our context, importance sampling has to be performed at the level of trajectories. Trajectories generated by the model are distributed according to some unknown pdf P0({X(t)}0 ≤ t ≤ T={x(t)}0 ≤ t ≤ T). Suppose given the model trajectory, we can evaluate some quantity of interest as given by
Suppose from Eq 2 and 3 we can estimate the probability of total DJF precipitation being smaller than b as denoted by
Suppose the simulation horizon is
1. Iterate each trajectory from time ti − 1 = (i−1)τ to time ti = iτ,
2. At time ti, stop the simulation and estimate the weight associated with each trajectory n as given by
where
3. Randomly sample N new trajectories (with replacement) according to the probability mass function
4. Add small random perturbations to the states of the
5. Increment
Then more weights we put on the trajectories distributed near the domain of interets (
where
I used a simple Markov chain to simulate daily precipitation over a 200-day interval. The probabilities of rain and no rain for today are only conditioned on the state of yeasterday and if we roll 'rain' for today, the intensity is drawn from a log normal distribution. The parameter values are arbitrarily assigned. Run control_run.m
to do a control run of 100,000 realizations (i.e., 100,000 years) and run alter_run.m
to repeat for 20 times the altered simulation using the abovementioned algorithm (and each altered run consists of 128 realizations/years). Note that step 4 was not used in this toy problem since I used a stochastic process already.
Ragone, F., Wouters, J., & Bouchet, F. (2018). Computation of extreme heat waves in climate models using a large deviation algorithm. Proceedings of the National Academy of Sciences, 115(1), 24-29.
Giardina, C., Kurchan, J., Lecomte, V., & Tailleur, J. (2011). Simulating rare events in dynamical processes. Journal of statistical physics, 145, 787-811.