Abstract
In many real-world scenarios, such as gas leak detection or environmentalpollutant tracking, solving the Inverse Source Localization andCharacterization problem involves navigating complex, dynamic fields withsparse and noisy observations. Traditional methods face significant challenges,including partial observability, temporal and spatial dynamics,out-of-distribution generalization, and reward sparsity. To address theseissues, we propose a hierarchical framework that integrates Bayesian inferenceand reinforcement learning. The framework leverages an attention-enhancedparticle filtering mechanism for efficient and accurate belief updates, andincorporates two complementary execution strategies: Attention ParticleFiltering Planning and Attention Particle Filtering Reinforcement Learning.These approaches optimize exploration and adaptation under uncertainty.Theoretical analysis proves the convergence of the attention-enhanced particlefilter, while extensive experiments across diverse scenarios validate theframework's superior accuracy, adaptability, and computational efficiency. Ourresults highlight the framework's potential for broad applications in dynamicfield estimation tasks.