Skip to Main Content
The particle filter (PF) is a state estimation algorithm that is inherently suitable for parallel computing. When the PF is implemented on a parallel computer, it is crucial to reduce the number of data transfers in the resampling procedure. One effective way to do this is to divide the particles into multiple groups. If the resampling is then performed only within each group, data transfers are reduced effectively. However, when the resampling is limited to within a small group, the imbalance of the weights of the particles cannot be resolved sufficiently, and this can depress the estimation accuracy. To evaluate this imbalance, we introduce a metric based on the entropy and observe that the accuracy actually does become worse as the imbalance of weights becomes more evident. We then propose a recipe in which the imbalance of weights is resolved when the metric of the imbalance is less than a predetermined threshold value. Finally, we demonstrate that this recipe notably improves the estimation accuracy without requiring substantial additional computational cost.