Particle Reallocation Search for hyperparameter tuning of particle filters

Abstract. Particle-filter performance depends strongly on model- and algorithm-side hyperparameters. Exhaustive offline procedures, such as grid search, are computationally expensive because every candidate configuration is compared through a filtering run that uses the target number of particles N over a long time horizon. We propose Particle Reallocation Search (PRS), a hyperparameter search that operates under a fixed total computational cost and treats the number of particles as a controllable resource. Following the successive-halving idea used for hyperparameter optimization in machine learning, PRS first screens many candidates with small ensembles and discards poorly scoring candidates at intermediate screening times. The released particles are then reallocated to the surviving candidates by deterministic duplication such that their filters are continued instead of being restarted. Two devices stabilize the search. First, candidates are evaluated with coupled Monte Carlo draws through common random numbers, which reduces the variance of pairwise score differences. Second, the search region is repeatedly recentered and rescaled around the best configuration found so far. Numerical experiments on a linear random-walk model and the nonlinear Lorenz-63 and Lorenz-96 systems showed that the configuration selected by PRS is close to the seed-averaged exhaustive-grid optimum – on one task, its single-run evaluation exceeds the seed-averaged grid-best score – while using approximately 2 %–25 % of the cost of grid search. The largest absolute computational saving was achieved on the high-dimensional Lorenz-96 model.