Abstract
Auxiliary particle filters (APFs) are a class of sequential Monte Carlo (SMC) methods for Bayesian inference in state-space models. In their original derivation, APFs operate in an extended state space to improve the quality of the proposal distribution. This paper introduces the Optimized Auxiliary Particle Filter (OAPF), a framework that treats the selection of proposal weights as a convex optimization problem. By matching the proposal distribution to the target posterior at key evaluation points, OAPF achieves lower variance and better accuracy than state-of-the-art filters while maintaining computational efficiency through weight sparsity.
Introduction
In fields ranging from robotics to finance, State-Space Models (SSMs) are essential for tracking hidden variables based on noisy observations. Particle Filters (PFs) are the standard tool for this task, representing the state belief through a set of weighted samples. However, standard PFs frequently suffer from weight degeneracy, where most particles end up with near-zero weights, leading to computational waste.
The Auxiliary Particle Filter (APF) was developed to mitigate this by "looking ahead" at future observations before sampling. Despite its success, the original APF often relies on heuristics or mathematically intractable "optimal" choices. The OAPF addresses this gap by providing a principled, computationally efficient method to optimize auxiliary weights for any given model.
Methodology
The OAPF framework introduces a three-step approach to proposal adaptation:
- Mixture Proposal Interpretation: The authors re-interpret auxiliary variables as weights ($\lambda$) in a mixture distribution composed of $K$ different kernels.
- Convex Optimization: To determine the optimal weights, the system matches the mixture proposal to the shape of the true filtering posterior at a set of $E$ evaluation points.
- Non-Negative Least Squares (NNLS): The weight optimization is formulated as an NNLS problem. Unlike non-convex "black-box" optimization, NNLS is a convex problem with a unique global optimum that can be solved reliably and quickly.
This framework decouples the number of kernels ($K$) and evaluation points ($E$) from the number of particles ($M$), ensuring the filter remains efficient as sample sizes increase.
Experimental Evaluation
The authors evaluated OAPF against standard benchmarks, including Linear Dynamical Models (LDM), the chaotic Stochastic Lorenz 63 model, and the high-dimensional Multivariate Stochastic Volatility Model (SVM).
Robustness to Multimodality
A key challenge in particle filtering is capturing multimodal distributions. In toy examples, OAPF demonstrated a unique ability to accurately match multimodal posteriors, whereas competing filters often collapsed onto a single mode.
Key Results
Accuracy and Efficiency
In linear models, OAPF achieved a lower Normalized Mean-Squared Error (NMSE) than all competitors, approaching the theoretical limit set by the Kalman Filter. Furthermore, in the Stochastic Volatility Model, OAPF significantly increased the Effective Sample Size (ESS), indicating a more efficient use of particles.
Computational Sparsity
The NNLS optimization naturally produces sparse weights, effectively pruning the proposal to focus on the most relevant regions of the state space. In tests with $M=1000$ particles, OAPF achieved an 88% reduction in the effective number of kernels, keeping runtime competitive despite the added optimization step.
Discussion and Conclusion
The OAPF represents a significant advancement in SMC methods by replacing heuristics with a robust convex optimization framework. The authors proved that OAPF estimators are unbiased and consistent, ensuring convergence to the true solution. While performance still faces the "curse of dimensionality" inherent to all particle filters, OAPF offers a more mathematically grounded and efficient alternative for complex dynamical systems. Future work may explore more sophisticated strategies for selecting evaluation points to further enhance performance.
Summary updated