Predictive Distance Weighting in Surrogate-Assisted ABC SMC : A Comparative Study of Predictive distance Weighting Strategies in ABC SMC
Rajapaksha Pathirannahalage, Shashikala Lankadari (2026)
2026
Master's Programme in Computing Sciences and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2026-06-24
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202606237893
https://urn.fi/URN:NBN:fi:tuni-202606237893
Tiivistelmä
Approximate Bayesian Computation (ABC) has become an important framework for likelihood-free Bayesian inference in situations where the likelihood function is analytically intractable or computationally expensive to evaluate. Among ABC methods, Sequential Monte Carlo Approximate Bayesian Computation (ABC SMC) improves sampling efficiency by evolving weighted particle populations through a sequence of progressively stricter tolerance thresholds. However, ABC SMC may still require a large number of simulator evaluations, particularly for nonlinear and high-dimensional inference problems.
This thesis investigates predictive distance weighting strategies within a Gaussian Process (GP)-assisted ABC SMC framework in order to improve computational efficiency while maintaining posterior approximation quality. Traditional weighting (standard ABC SMC importance weighting), mean-based weighting, annealing weighting, and a proposed probabilistic weighting strategy are the four weighting strategies evaluated in this study. A GP surrogate is used to approximate the discrepancy between simulated and observed data and to guide surrogate-assisted particle reweighting before resampling. The weighting strategies were evaluated using a one-dimensional Gaussian model, a two-dimensional banana-shaped model, and a Lotka–Volterra predator–prey model under increasing model complexity.
The experimental results demonstrate that GP-assisted weighting strategies can improve the computational efficiency of ABC SMC while maintaining posterior approximation quality. Among the evaluated methods, the mean-based weighting strategy provided the most consistent balance between posterior approximation quality, acceptance behavior, and computational efficiency, while the probabilistic weighting strategy demonstrated competitive computational performance by incorporating both the GP predictive mean and predictive uncertainty into the weighting process. The annealing weighting approach provided a balance between exploration and exploitation during sequential inference.
In conclusion, the results suggest that surrogate-assisted predictive distance weighting can enhance the efficiency of ABC SMC inference while preserving reliable posterior approximation performance. The proposed GP-assisted framework provides a flexible approach for parameter inference in nonlinear and computationally intensive simulator-based models, making it a promising direction for future likelihood-free Bayesian methodologies.
This thesis investigates predictive distance weighting strategies within a Gaussian Process (GP)-assisted ABC SMC framework in order to improve computational efficiency while maintaining posterior approximation quality. Traditional weighting (standard ABC SMC importance weighting), mean-based weighting, annealing weighting, and a proposed probabilistic weighting strategy are the four weighting strategies evaluated in this study. A GP surrogate is used to approximate the discrepancy between simulated and observed data and to guide surrogate-assisted particle reweighting before resampling. The weighting strategies were evaluated using a one-dimensional Gaussian model, a two-dimensional banana-shaped model, and a Lotka–Volterra predator–prey model under increasing model complexity.
The experimental results demonstrate that GP-assisted weighting strategies can improve the computational efficiency of ABC SMC while maintaining posterior approximation quality. Among the evaluated methods, the mean-based weighting strategy provided the most consistent balance between posterior approximation quality, acceptance behavior, and computational efficiency, while the probabilistic weighting strategy demonstrated competitive computational performance by incorporating both the GP predictive mean and predictive uncertainty into the weighting process. The annealing weighting approach provided a balance between exploration and exploitation during sequential inference.
In conclusion, the results suggest that surrogate-assisted predictive distance weighting can enhance the efficiency of ABC SMC inference while preserving reliable posterior approximation performance. The proposed GP-assisted framework provides a flexible approach for parameter inference in nonlinear and computationally intensive simulator-based models, making it a promising direction for future likelihood-free Bayesian methodologies.