[论文解读] Robust, partially alive particle Metropolis-Hastings via the Frankenfilter

Chris Sherlock, Andrew Golightly+1 more|arXiv (Cornell University)|Jan 30, 2026

Target Tracking and Data Fusion in Sensor Networks被引用 0

一句话总结

The Frankenfilter provides an unbiased, robust particle filtering approach that targets a user-defined amount of success with bounds on simulations, improving PMMH performance in challenging hidden Markov models.

ABSTRACT

When a hidden Markov model permits the conditional likelihood of an observation given the hidden process to be zero, all particle simulations from one observation time to the next could produce zeros. If so, the filtering distribution cannot be estimated and the estimated parameter likelihood is zero. The alive particle filter addresses this by simulating a random number of particles for each inter-observation interval, stopping after a target number of non-zero conditional likelihoods. For outlying observations or poor parameter values, a non-zero result can be extremely unlikely, and computational costs prohibitive. We introduce the Frankenfilter, a principled, partially alive particle filter that targets a user-defined amount of success whilst fixing lower and upper bounds on the number of simulations. The Frankenfilter produces unbiased estimators of the likelihood, suitable for pseudo-marginal Metropolis--Hastings (PMMH). We demonstrate that PMMH with the Frankenfilter is more robust to outliers and mis-specified initial parameter values than PMMH using standard particle filters, and is typically at least 2-3 times more efficient. We also provide advice for choosing the amount of success. In the case of n exact observations, this is particularly simple: target n successes.

研究动机与目标

• Motivate robust parameter inference for hidden Markov models where conditional likelihoods can be zero for some simulations.
• Develop an unbiased particle filtering method (the Frankenfilter) that is partially alive and bounded in computational cost.
• Show that Frankenfilter-based PMMH is more robust to outliers and mis-specified parameters than standard filters, with improved efficiency.
• Provide practical tuning guidance to select the success target and simulation bounds for various observation regimes.

提出的方法

• Introduce the Frankenfilter, a generalisation of the alive particle filter with a minimum and maximum number of simulations (m- and m+).
• Allow a non-binary measure of success and any non-negative conditional likelihood, enabling informed proposals or bridges.
• Prove unbiasedness of the likelihood estimator produced by the Frankenfilter (and its variants for complete and partial observations).
• Present algorithms (Basic Frankenfilter, FrankenFilterOne, FrankenFilter, and Ancestor Sampling) that control stopping rules to ensure unbiased likelihood estimates.
• Derive tuning guidance to set the target number of successes s and the bounds m-, m+ to achieve a desired relative variance in PMMH.
• Demonstrate robustness and efficiency gains of PMMH with the Frankenfilter through simulation studies.

Figure 1: Death model. Ratio of the expectation of the estimator of the likelihood and exact likelihood at $\theta=0.01$ , based on the Frankenfilter (black lines) and alive particle filter (grey lines) using D50 (left panel) and D50mod (right panel).

实验结果

研究问题

• RQ1 Can the Frankenfilter provide unbiased likelihood estimates under bounded computational cost through controlled success thresholds?
• RQ2 How should one choose the target number of successes and the min/max number of simulations to balance robustness and efficiency in complete vs. partial observation regimes?
• RQ3 Does PMMH with the Frankenfilter outperform standard particle filters in the presence of outliers or mis-specified parameters?
• RQ4 What are practical tuning recommendations for achieving a target relative variance in the likelihood estimator across different observation scenarios?

主要发现

• The Frankenfilter yields an unbiased estimator of the data likelihood within PMMH.
• It offers robustness to outliers and mis-specified parameters compared to standard particle filters.
• Tuning guidance shows how to set the success target s and bounds m- and m+ to achieve desirable variance properties, often with s around 2+T/log(1+Vrel).
• Theoretical results establish unbiasedness for complete and partial observations (Algorithms 5 and related proofs).
• Simulation results indicate PMMH with the Frankenfilter is typically 2-3x more efficient than comparable approaches.
• The framework supports complete and partial observations and can accommodate non-binary weights and bridge proposals.

Figure 2: Protein dimerisation model. ESS/s versus required total success $\mathfrak{s}$ , from the output of $50K$ iterations of PMMH, using data sets with 10 observations (left panel), 30 observations (middle panel), 50 observations (right panel). Solid lines correspond to data sets P10a, P30a and