BIMSA >
Probability and Dynamical Systems Seminar
Neural Adaptive Importance Sampling for High-Dimensional Autonomous Landing Systems
Neural Adaptive Importance Sampling for High-Dimensional Autonomous Landing Systems
Organizers
Speaker
Time
Tuesday, November 18, 2025 3:15 PM - 4:15 PM
Venue
A3-1-101
Abstract
This report presents the fundamental principles of importance sampling and examines the challenges encountered when applying Monte Carlo methods in high-dimensional settings. Our ongoing research aims to develop a non-uniform Monte Carlo framework specifically designed for high-dimensional control problems in autonomous landing systems, with a focus on variance reduction and improved sample efficiency.
We introduce a Neural Adaptive Importance Sampling (NAIS) methodology in which two neural networks jointly learn an optimized proposal distribution and a surrogate agent model. By tightly coupling sampling strategies with the underlying control dynamics, the proposed framework seeks to substantially reduce estimation variance, lower the required number of samples, and enhance the reliability of performance evaluation in rare-event scenarios.
Current efforts concentrate on establishing a unified learning scheme that includes the construction of loss functions balancing distribution alignment and predictive accuracy, coordination between sampling and model training, stabilization of the joint optimization process, and comprehensive error analysis. Ultimately, this work aims to improve the robustness, accuracy, and computational efficiency of autonomous landing systems through intelligent, data-driven sampling techniques.
We introduce a Neural Adaptive Importance Sampling (NAIS) methodology in which two neural networks jointly learn an optimized proposal distribution and a surrogate agent model. By tightly coupling sampling strategies with the underlying control dynamics, the proposed framework seeks to substantially reduce estimation variance, lower the required number of samples, and enhance the reliability of performance evaluation in rare-event scenarios.
Current efforts concentrate on establishing a unified learning scheme that includes the construction of loss functions balancing distribution alignment and predictive accuracy, coordination between sampling and model training, stabilization of the joint optimization process, and comprehensive error analysis. Ultimately, this work aims to improve the robustness, accuracy, and computational efficiency of autonomous landing systems through intelligent, data-driven sampling techniques.